[RFC Home] [TEXT|PDF|HTML] [Tracker] [IPR] [Errata] [Info page]
EXPERIMENTAL
Errata ExistNetwork Working Group G. Apostolopoulos
Request for Comments: 2676 D. Williams
Category: Experimental IBM
S. Kamat
Lucent
R. Guerin
UPenn
A. Orda
Technion
T. Przygienda
Siara Systems
August 1999
QoS Routing Mechanisms and OSPF Extensions
Status of this Memo
This memo defines an Experimental Protocol for the Internet
community. It does not specify an Internet standard of any kind.
Discussion and suggestions for improvement are requested.
Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (1999). All Rights Reserved.
Abstract
This memo describes extensions to the OSPF [Moy98] protocol to
support QoS routes. The focus of this document is on the algorithms
used to compute QoS routes and on the necessary modifications to OSPF
to support this function, e.g., the information needed, its format,
how it is distributed, and how it is used by the QoS path selection
process. Aspects related to how QoS routes are established and
managed are also briefly discussed. The goal of this document is to
identify a framework and possible approaches to allow deployment of
QoS routing capabilities with the minimum possible impact to the
existing routing infrastructure.
In addition, experience from an implementation of the proposed
extensions in the GateD environment [Con], along with performance
measurements is presented.
Apostolopoulos, et al. Experimental [Page 1]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Table of Contents
1. Introduction 3
1.1. Overall Framework . . . . . . . . . . . . . . . . . . . . 3
1.2. Simplifying Assumptions . . . . . . . . . . . . . . . . . 5
2. Path Selection Information and Algorithms 7
2.1. Metrics . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2. Advertisement of Link State Information . . . . . . . . . 8
2.3. Path Selection . . . . . . . . . . . . . . . . . . . . .10
2.3.1. Path Computation Algorithm . . . . . . . . . . .11
3. OSPF Protocol Extensions 16
3.1. QoS -- Optional Capabilities . . . . . . . . . . . . . .17
3.2. Encoding Resources as Extended TOS . . . . . . . . . . .17
3.2.1. Encoding bandwidth resource . . . . . . . . . . .19
3.2.2. Encoding Delay . . . . . . . . . . . . . . . . .21
3.3. Packet Formats . . . . . . . . . . . . . . . . . . . . .21
3.4. Calculating the Inter-area Routes . . . . . . . . . . . .22
3.5. Open Issues . . . . . . . . . . . . . . . . . . . . . . .22
4. A Reference Implementation based on GateD 22
4.1. The Gate Daemon (GateD) Program . . . . . . . . . . . . .22
4.2. Implementing the QoS Extensions of OSPF . . . . . . . . .23
4.2.1. Design Objectives and Scope . . . . . . . . . . .23
4.2.2. Architecture . . . . . . . . . . . . . . . . . .24
4.3. Major Implementation Issues . . . . . . . . . . . . . . .25
4.4. Bandwidth and Processing Overhead of QoS Routing . . . .29
5. Security Considerations 32
A. Pseudocode for the BF Based Pre-Computation Algorithm 33
B. On-Demand Dijkstra Algorithm for QoS Path Computation 36
C. Precomputation Using Dijkstra Algorithm 39
D. Explicit Routing Support 43
Endnotes 45
References 46
Authors' Addresses 48
Full Copyright Statement 50
Apostolopoulos, et al. Experimental [Page 2]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
1. Introduction
In this document, we describe a set of proposed additions to the OSPF
routing protocol (these additions have been implemented on top of the
GateD [Con] implementation of OSPF V2 [Moy98]) to support Quality-
of-Service (QoS) routing in IP networks. Support for QoS routing can
be viewed as consisting of three major components:
1. Obtain the information needed to compute QoS paths and select a
path capable of meeting the QoS requirements of a given request,
2. Establish the path selected to accommodate a new request,
3. Maintain the path assigned for use by a given request.
Although we touch upon aspects related to the last two components,
the focus of this document is on the first one. In particular, we
discuss the metrics required to support QoS, the extension to the
OSPF link state advertisement mechanism to propagate updates of QoS
metrics, and the modifications to the path selection to accommodate
QoS requests. The goal of the extensions described in this document
is to improve performance for QoS flows (likelihood to be routed on a
path capable of providing the requested QoS), with minimal impact on
the existing OSPF protocol and its current implementation. Given the
inherent complexity of QoS routing, achieving this goal obviously
implies trading-off "optimality" for "simplicity", but we believe
this to be required in order to facilitate deployment of QoS routing
capabilities.
In addition to describing the proposed extensions to the OSPF
protocol, this document also reports experimental data based on
performance measurements of an implementation done on the GateD
platform (see Section 4).
1.1. Overall Framework
We consider a network (1) that supports both best-effort packets and
packets with QoS guarantees. The way in which the network resources
are split between the two classes is irrelevant, except for the
assumption that each QoS capable router in the network is able to
dedicate some of its resources to satisfy the requirements of QoS
packets. QoS capable routers are also assumed capable of identifying
and advertising resources that remain available to new QoS flows. In
addition, we limit ourselves to the case where all the routers
involved support the QoS extensions described in this document, i.e.,
we do not consider the problem of establishing a route in a
heterogeneous environment where some routers are QoS-capable and
others are not. Furthermore, in this document, we focus on the case
Apostolopoulos, et al. Experimental [Page 3]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
of unicast flows, although many of the additions we define are
applicable to multicast flows as well.
We assume that a flow with QoS requirements specifies them in some
fashion that is accessible to the routing protocol. For example,
this could correspond to the arrival of an RSVP [RZB+97] PATH
message, whose TSpec is passed to routing together with the
destination address. After processing such a request, the routing
protocol returns the path that it deems the most suitable given the
flow's requirements. Depending on the scope of the path selection
process, this returned path could range from simply identifying the
best next hop, i.e., a hop-by-hop path selection model, to specifying
all intermediate nodes to the destination, i.e., an explicit route
model. The nature of the path being returned impacts the operation
of the path selection algorithm as it translates into different
requirements for constructing and returning the appropriate path
information. However, it does not affect the basic operation of the
path selection algorithm (2).
For simplicity and also because it is the model currently supported
in the implementation (see Section 4 for details), in the rest of
this document we focus on the hop-by-hop path selection model. The
additional modifications required to support an explicit routing
model are discussed in appendix D, but are peripheral to the main
focus of this document which concentrates on the specific extensions
to the OPSF protocol to support computation of QoS routes.
In addition to the problem of selecting a QoS path and possibly
reserving the corresponding resources, one should note that the
successful delivery of QoS guarantees requires that the packets of
the associated "QoS flow" be forwarded on the selected path. This
typically requires the installation of corresponding forwarding state
in the router. For example, with RSVP [RZB+97] flows a classifier
entry is created based on the filter specs contained in the RESV
message. In the case of a Differentiated Service [KNB98] setting,
the classifier entry may be based on the destination address (or
prefix) and the corresponding value of the DS byte. The mechanisms
described in this document are at the control path level and are,
therefore, independent of data path mechanisms such as the packet
classification method used. Nevertheless, it is important to notice
that consistent delivery of QoS guarantees implies stability of the
data path. In particular, while it is possible that after a path is
first selected, network conditions change and result in the
appearance of "better" paths, such changes should be prevented from
unnecessarily affecting existing paths. In particular, switching
over to a new (and better) path should be limited to specific
conditions, e.g., when the initial selection turns out to be
inadequate or extremely "expensive". This aspect is beyond the scope
Apostolopoulos, et al. Experimental [Page 4]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
of QoS routing and belongs to the realm of path management, which is
outside the main focus of this document. However, because of its
potentially significant impact on the usefulness of QoS routing, we
briefly outline a possible approach to path management.
Avoiding unnecessary changes to QoS paths requires that state
information be maintained for each QoS path after it has been
selected. This state information is used to track the validity of
the path, i.e., is the current path adequate or should QoS routing be
queried again to generate a new and potentially better path. We say
that a path is "pinned" when its state specifies that QoS routing
need not be queried anew, while a path is considered "un-pinned"
otherwise. The main issue is then to define how, when, and where
path pinning and un-pinning is to take place, and this will typically
depend on the mechanism used to request QoS routes. For example,
when the RSVP protocol is the mechanism being used, it is desirable
that path management be kept as synergetic as possible with the
existing RSVP state management. In other words, pinning and un-
pinning of paths should be coordinated with RSVP soft states, and
structured so as to require minimal changes to RSVP processing rules.
A broad RSVP-routing interface that enables this is described in
[GKR97]. Use of such an interface in the context of reserving
resources along an explicit path with RSVP is discussed in [GLG+97].
Details of path management and a means for avoiding loops in case of
hop-by-hop path setup can be found in [GKH97], and are not addressed
further in this document.
1.2. Simplifying Assumptions
In order to achieve our goal of minimizing impact to the existing
protocol and implementation, we impose certain restrictions on the
range of extensions we initially consider to support QoS. The first
restriction is on the type of additional (QoS) metrics that will be
added to Link State Advertisements (LSAs) for the purpose of
distributing metrics updates. Specifically, the extensions to LSAs
that we initially consider, include only available bandwidth and
delay. In addition, path selection is itself limited to considering
only bandwidth requirements. In particular, the path selection
algorithm selects paths capable of satisfying the bandwidth
requirement of flows, while at the same time trying to minimize the
amount of network resources that need to be allocated, i.e., minimize
the number of hops used.
This focus on bandwidth is adequate in most instances, and meant to
keep initial complexity at an acceptable level. However, it does not
fully capture the complete range of potential QoS requirements. For
example, a delay-sensitive flow of an interactive application could
be put on a path using a satellite link, if that link provided a
Apostolopoulos, et al. Experimental [Page 5]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
direct path and had plenty of unused bandwidth. This would clearly
be an undesirable choice. Our approach to preventing such poor
choices, is to assign delay-sensitive flows to a "policy" that would
eliminate from the network all links with high propagation delay,
e.g., satellite links, before invoking the path selection algorithm.
In general, multiple policies could be used to capture different
requirements, each presenting to the path selection algorithm a
correspondingly pruned network topology, on which the same algorithm
would be used to generate an appropriate path. Alternatively,
different algorithms could be used depending on the QoS requirements
expressed by an incoming request. Such extensions are beyond the
scope of this document, which limits itself to describing the case of
a single metric, bandwidth. However, it is worth pointing out that a
simple extension to the path selection algorithm proposed in this
document allows us to directly account for delay, under certain
conditions, when rate-based schedulers are employed, as in the
Guaranteed Service proposal [SPG97]; details can be found in [GOW97].
Another important aspect to ensure that introducing support for QoS
routing has the minimal possible impact, is to develop a solution
that has the smallest possible computing overhead. Additional
computations are unavoidable, but it is desirable to keep the
computational cost of QoS routing at a level comparable to that of
traditional routing algorithms. One possible approach to achieve
this goal, is to allow pre-computation of QoS routes. This is the
method that was chosen for the implementation of the QoS extensions
to OSPF and is, therefore, the one described in detail in this
document. Alternative approaches are briefly reviewed in appendices.
However, it should be noted that although several alternative path
selection algorithms are possible, the same algorithm should be used
consistently within a given routing domain. This requirement may be
relaxed when explicit routing is used, as the responsibility for
selecting a QoS path lies with a single entity, the origin of the
request, which then ensures consistency even if each router uses a
different path selection algorithm. Nevertheless, the use of a
common path selection algorithm within an AS is recommended, if not
necessary, for proper operation.
A last aspect of concern regarding the introduction of QoS routing,
is to control the overhead associated with the additional link state
updates caused by more frequent changes to link metrics. The goal is
to minimize the amount of additional update traffic without adversely
affecting the performance of path selection. In Section 2.2, we
present a brief discussion of various alternatives that trade
accuracy of link state information for protocol overhead. Potential
enhancements to the path selection algorithm, which seek to
(directly) account for the inaccuracies in link metrics, are
described in [GOW97], while a comprehensive treatment of the subject
Apostolopoulos, et al. Experimental [Page 6]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
can be found in [LO98, GO99]. In Section 4, we also describe the
design choices made in a reference implementation, to allow future
extensions and experimentation with different link state update
mechanisms.
The rest of this document is structured as follows. In Section 2, we
describe the general design choices and mechanisms we rely on to
support QoS request. This includes details on the path selection
metrics, link state update extensions, and the path selection
algorithm itself. Section 3 focuses on the specific extensions that
the OSPF protocol requires, while Section 4 describes their
implementation in the GateD platform and also presents some
experimental results. Section 5 briefly addresses security issues
that the proposed schemes may raise. Finally, several appendices
provide additional material of interest, e.g., alternative path
selection algorithms and support for explicit routes, but somewhat
outside the main focus of this document.
2. Path Selection Information and Algorithms
This section reviews the basic building blocks of QoS path selection,
namely the metrics on the which the routing algorithm operates, the
mechanisms used to propagate updates for these metrics, and finally
the path selection algorithm itself.
2.1. Metrics
The process of selecting a path that can satisfy the QoS requirements
of a new flow relies on both the knowledge of the flow's requirements
and characteristics, and information about the availability of
resources in the network. In addition, for purposes of efficiency,
it is also important for the algorithm to account for the amount of
resources the network has to allocate to support a new flow. In
general, the network prefers to select the "cheapest" path among all
paths suitable for a new flow, and it may even decide not to accept a
new flow for which a feasible path exists, if the cost of the path is
deemed too high. Accounting for these aspects involves several
metrics on which the path selection process is based. They include:
- Link available bandwidth: As mentioned earlier, we currently
assume that most QoS requirements are derivable from a rate-
related quantity, termed "bandwidth." We further assume that
associated with each link is a maximal bandwidth value, e.g., the
link physical bandwidth or some fraction thereof that has been set
aside for QoS flows. Since for a link to be capable of accepting
a new flow with given bandwidth requirements, at least that much
bandwidth must be still available on the link, the relevant link
metric is, therefore, the (current) amount of available (i.e.,
Apostolopoulos, et al. Experimental [Page 7]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
unallocated) bandwidth. Changes in this metric need to be
advertised as part of extended LSAs, so that accurate information
is available to the path selection algorithm.
- Link propagation delay: This quantity is meant to identify high
latency links, e.g., satellite links, which may be unsuitable for
real-time requests. This quantity also needs to be advertised as
part of extended LSAs, although timely dissemination of this
information is not critical as this parameter is unlikely to
change (significantly) over time. As mentioned earlier, link
propagation delay can be used to decide on the pruning of specific
links, when selecting a path for a delay sensitive request; also,
it can be used to support a related extension, as described in
[GOW97].
- Hop-count: This quantity is used as a measure of the path cost to
the network. A path with a smaller number of hops (that can
support a requested connection) is typically preferable, since it
consumes fewer network resources. As a result, the path selection
algorithm will attempt to find the minimum hop path capable of
satisfying the requirements of a given request. Note that
contrary to bandwidth and propagation delay, hop count is a metric
that does not affect LSAs, and it is only used implicitly as part
of the path selection algorithm.
2.2. Advertisement of Link State Information
The new link metrics identified in the previous section need to be
advertised across the network, so that each router can compute
accurate and consistent QoS routes. It is assumed that each router
maintains an updated database of the network topology, including the
current state (available bandwidth and propagation delay) of each
link. As mentioned before, the distribution of link state (metrics)
information is based on extending OSPF mechanisms. The detailed
format of those extensions is described in Section 3, but in addition
to how link state information is distributed, another important
aspect is when such distribution is to take place.
One option is to mandate periodic updates, where the period of
updates is determined based on a tolerable corresponding load on the
network and the routers. The main disadvantage of such an approach
is that major changes in the bandwidth available on a link could
remain unknown for a full period and, therefore, result in many
incorrect routing decisions. Ideally, routers should have the most
current view of the bandwidth available on all links in the network,
so that they can make the most accurate decision of which path to
select. Unfortunately, this then calls for very frequent updates,
e.g., each time the available bandwidth of a link changes, which is
Apostolopoulos, et al. Experimental [Page 8]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
neither scalable nor practical. In general, there is a trade-off
between the protocol overhead of frequent updates and the accuracy of
the network state information that the path selection algorithm
depends on. We outline next a few possible link state update
policies, which strike a practical compromise.
The basic idea is to trigger link state advertisements only when
there is a significant change in the value of metrics since the last
advertisement. The notion of significance of a change can be based
on an "absolute" scale or a "relative" one. An absolute scale means
partitioning the range of values that a metric can take into
equivalence classes and triggering an update whenever the metric
changes sufficiently to cross a class boundary (3). A relative
scale, on the other hand, triggers updates when the percentage change
in the metric value exceeds a predefined threshold. Independent of
whether a relative or an absolute change trigger mechanism is used, a
periodic trigger constraint can also be added. This constraint can
be in the form of a hold-down timer, which is used to force a minimum
spacing between consecutive updates. Alternatively, a transmit timer
can also be used to ensure the transmission of an update after a
certain time has expired. Such a feature can be useful if link state
updates advertising bandwidth changes are sent unreliably. The
current protocol extensions described in Section 3 as well as the
implementation of Section 4 do not consider such an option as metric
updates are sent using the standard, and reliable, OSPF flooding
mechanism. However, this is clearly an extension worth considering
as it can help lower substantially the protocol overhead associated
with metrics updates.
In both the relative and absolute change approaches, the metric value
advertised in an LSA can be either the actual or a quantized value.
Advertising the actual metric value is more accurate and, therefore,
preferable when metrics are frequently updated. On the other hand,
when updates are less frequent, e.g., because of a low sensitivity
trigger or the use of hold-down timers, advertising quantized values
can be of benefit. This is because it can help increase the number
of equal cost paths and, therefore, improve robustness to metrics
inaccuracies. In general, there is a broad space of possible trade-
offs between accuracy and overhead and selecting an appropriate
design point is difficult and depends on many parameters (see
[AGKT98] for a more detailed discussion of these issues). As a
result, in order to help acquire a better understanding of these
issues, the implementation described in Section 4 supports a range of
options that allow exploration of the available design space. In
addition, Section 4 also reports experimental data on the traffic
load and processing overhead generated by links state updates for
different configurations.
Apostolopoulos, et al. Experimental [Page 9]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
2.3. Path Selection
There are two major aspects to computing paths for QoS requests. The
first is the actual path selection algorithm itself, i.e., which
metrics and criteria it relies on. The second is when the algorithm
is actually invoked.
The topology on which the algorithm is run is, as with the standard
OSPF path selection, a directed graph where vertices (4) consist of
routers and networks (transit vertices) as well as stub networks
(non-transit vertices). When computing a path, stub networks are
added as a post-processing step, which is essentially similar to what
is done with the current OSPF routing protocol. The optimization
criteria used by the path selection are reflected in the costs
associated with each interface in the topology and how those costs
are accounted for in the algorithm itself. As mentioned before, the
cost of a path is a function of both its hop count and the amount of
available bandwidth. As a result, each interface has associated with
it a metric, which corresponds to the amount of bandwidth that
remains available on this interface. This metric is combined with
hop count information to provide a cost value, whose goal is to pick
a path with the minimum possible number of hops among those that can
support the requested bandwidth. When several such paths are
available, the preference is for the path whose available bandwidth
(i.e., the smallest value on any of the links in the path) is
maximal. The rationale for the above rule is the following: we
focus on feasible paths (as accounted by the available bandwidth
metric) that consume a minimal amount of network resources (as
accounted by the hop-count metric); and the rule for selecting among
these paths is meant to balance load as well as maximize the
likelihood that the required bandwidth is indeed available.
It should be noted that standard routing algorithms are typically
single objective optimizations, i.e., they may minimize the hop-
count, or maximize the path bandwidth, but not both. Double
objective path optimization is a more complex task, and, in general,
it is an intractable problem [GJ79]. Nevertheless, because of the
specific nature of the two objectives being optimized (bandwidth and
hop count), the complexity of the above algorithm is competitive with
even that of standard single-objective algorithms. For readers
interested in a thorough treatment of the topic, with insights into
the connection between the different algorithms, linear algebra and
modification of metrics, [Car79] is recommended.
Before proceeding with a more detailed description of the path
selection algorithm itself, we briefly review the available options
when it comes to deciding when to invoke the algorithm. The two main
options are: 1) to perform on-demand computations, that is, trigger
Apostolopoulos, et al. Experimental [Page 10]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
a computation for each new request, and 2) to use some form of pre-
computation. The on-demand case involves no additional issues in
terms of when computations should be triggered, but running the path
selection algorithm for each new request can be computationally
expensive (see [AT98] for a discussion on this issue). On the other
hand, pre-computing paths amortizes the computational cost over
multiple requests, but each computation instance is usually more
expensive than in the on-demand case (paths are computed to all
destinations and for all possible bandwidth requests rather than for
a single destination and a given bandwidth request). Furthermore,
depending on how often paths are recomputed, the accuracy of the
selected paths may be lower. In this document, we primarily focus on
the case of pre-computed paths, which is also the only method
currently supported in the reference implementation described in
Section 4. In this case, clearly, an important issue is when such
pre-computation should take place. The two main options we consider
are periodic pre-computations and pre-computations after a given (N)
number of updates have been received. The former has the benefit of
ensuring a strict bound on the computational load associated with
pre-computations, while the latter can provide for a more responsive
solution (5). Section 4 provides some experimental results comparing
the performance and cost of periodic pre-computations for different
period values.
2.3.1. Path Computation Algorithm
This section describes a path selection algorithm, which for a given
network topology and link metrics (available bandwidth), pre-computes
all possible QoS paths, while maintaining a reasonably low
computational complexity. Specifically, the algorithm pre-computes
for any destination a minimum hop count path with maximum bandwidth,
and has a computational complexity comparable to that of a standard
Bellman-Ford shortest path algorithm. The Bellman-Ford (BF) shortest
path algorithm is adapted to compute paths of maximum available
bandwidth for all hop counts. It is a property of the BF algorithm
that, at its h-th iteration, it identifies the optimal (in our
context: maximal bandwidth) path between the source and each
destination, among paths of at most h hops. In other words, the cost
of a path is a function of its available bandwidth, i.e., the
smallest available bandwidth on all links of the path, and finding a
minimum cost path amounts to finding a maximum bandwidth path.
However, because the BF algorithm progresses by increasing hop count,
it essentially provides for free the hop count of a path as a second
optimization criteria.
Specifically, at the kth (hop count) iteration of the algorithm, the
maximum bandwidth available to all destinations on a path of no more
than k hops is recorded (together with the corresponding routing
Apostolopoulos, et al. Experimental [Page 11]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
information). After the algorithm terminates, this information
provides for all destinations and bandwidth requirements, the path
with the smallest possible number of hops and sufficient bandwidth to
accommodate the new request. Furthermore, this path is also the one
with the maximal available bandwidth among all the feasible paths
with at most these many hops. This is because for any hop count, the
algorithm always selects the one with maximum available bandwidth.
We now proceed with a more detailed description of the algorithm and
the data structure used to record routing information, i.e., the QoS
routing table that gets built as the algorithm progresses (the
pseudo-code for the algorithm can be found in Appendix A). As
mentioned before, the algorithm operates on a directed graph
consisting only of transit vertices (routers and networks), with
stub-networks subsequently added to the path(s) generated by the
algorithm. The metric associated with each edge in the graph is the
bandwidth available on the corresponding interface. Let us denote by
b(n;m) the available bandwidth on the link from node n to m. The
vertex corresponding to the router where the algorithm is being run,
i.e., the computing router, is denoted as the "source node" for the
purpose of path selection. The algorithm proceeds to pre-compute
paths from this source node to all possible destination networks and
for all possible bandwidth values. At each (hop count) iteration,
intermediate results are recorded in a QoS routing table, which has
the following structure:
The QoS routing table:
- a KxH matrix, where K is the number of destinations (vertices in
the graph) and H is the maximal allowed (or possible) number of
hops for a path.
- The (n;h) entry is built during the hth iteration (hop count
value) of the algorithm, and consists of two fields:
* bw: the maximum available bandwidth, on a path of at most h
hops between the source node (router) and destination node
n;
* neighbor: this is the routing information associated with
the h (or less) hops path to destination node n, whose
available bandwidth is bw. In the context of hop-by-hop
path selection (6), the neighbor information is simply the
identity of the node adjacent to the source node on that
path. As a rule, the "neighbor" node must be a router and
not a network, the only exception being the case where the
network is the destination node (and the selected path is
the single edge interconnecting the source to it).
Apostolopoulos, et al. Experimental [Page 12]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Next, we provide additional details on the operation of the algorithm
and how the entries in the routing table are updated as the algorithm
proceeds. For simplicity, we first describe the simpler case where
all edges count as "hops," and later explain how zero-hop edges are
handled. Zero-hop edges arise in the case of transit networks
vertices, where only one of the two incoming and outgoing edges
should be counted in the hop count computation, as they both
correspond to the same physical hop. Accounting for this aspect
requires distinguishing between network and router nodes, and the
steps involved are detailed later in this section as well as in the
pseudo-code of Appendix A.
When the algorithm is invoked, the routing table is first initialized
with all bw fields set to 0 and neighbor fields cleared. Next, the
entries in the first column (which corresponds to one-hop paths) of
the neighbors of the computing router are modified in the following
way: the bw field is set to the value of the available bandwidth on
the direct edge from the source. The neighbor field is set to the
identity of the neighbor of the computing router, i.e., the next
router on the selected path.
Afterwards, the algorithm iterates for at most H iterations
(considering the above initial iteration as the first). The value of
H could be implicit, i.e., the diameter of the network or, in order
to better control the worst case complexity, it can be set explicitly
thereby limiting path lengths to at most H hops. In the latter case,
H must be assigned a value larger than the length of the minimum
hop-count path to any node in the graph.
At iteration h, we first copy column h-1 into column h. In
addition, the algorithm keeps a list of nodes that changed their bw
value in the previous iteration, i.e., during the (h-1)-th iteration.
The algorithm then looks at each link (n;m) where n is a node whose
bw value changed in the previous iteration, and checks the maximal
available bandwidth on an (at most) h-hop path to node m whose final
hop is that link. This amounts to taking the minimum between the bw
field in entry (n;h-1) and the link metric value b(n;m) kept in the
topology database. If this value is higher than the present value of
the bw field in entry (m;h), then a better (larger bw value) path has
been found for destination m and with at most h hops. The bw field
of entry (m;h) is then updated to reflect this new value. In the
case of hop-by-hop routing, the neighbor field of entry (m;h) is set
to the same value as in entry (n;h-1). This records the identity of
the first hop (next hop from the source) on the best path identified
thus far for destination m and with h (or less) hops.
Apostolopoulos, et al. Experimental [Page 13]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
As mentioned earlier, extending the above algorithm to handle zero-
hop edges is needed due to the possible use of multi-access networks,
e.g., T/R, E/N, etc., to interconnect routers. Such entities are
also represented by means of a vertex in the OSPF topology, but a
network connecting two routers should clearly be considered as a
single hop path rather than a two hop path. For example, consider
three routers A, B, and C connected over an Ethernet network N, which
the OSPF topology represents as in Figure 1.
A----N----B
|
|
C
Figure 1: Zero-Hop Edges
In the example of Figure 1, although there are directed edges in both
directions, an edge from the network to any of the three routers must
have zero "cost", so that it is not counted twice. It should be
noted that when considering such environments in the context of QoS
routing, it is assumed that some entity is responsible for
determining the "available bandwidth" on the network, e.g., a subnet
bandwidth manager. The specification and operation of such an entity
is beyond the scope of this document.
Accommodating zero-hop edges in the context of the path selection
algorithm described above is done as follows: At each iteration h
(starting with the first), whenever an entry (m;h) is modified, it is
checked whether there are zero-cost edges (m;k) emerging from node m.
This is the case when m is a transit network. In that case, we
attempt to further improve the entry of node k within the current
iteration, i.e., entry (k;h) (rather than entry (k;h+1)), since the
edge (m;k) should not count as an additional hop. As with the
regular operation of the algorithm, this amounts to taking the
minimum between the bw field in entry (m;h) and the link metric value
b(m;k) kept in the topology database (7). If this value is higher
than the present value of the bw field in entry (k;h), then the bw
field of entry (k;h) is updated to this new value. In the case of
hop-by-hop routing, the neighbor field of entry (k;h) is set, as
usual, to the same value as in entry (m;h) (which is also the value
in entry (n;h-1)).
Apostolopoulos, et al. Experimental [Page 14]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Note that while for simplicity of the exposition, the issue of equal
cost, i.e., same hop count and available bandwidth, is not detailed
in the above description, it can be easily supported. It only
requires that the neighbor field be expanded to record the list of
next (previous) hops, when multiple equal cost paths are present.
Addition of Stub Networks
As was mentioned earlier, the path selection algorithm is run on a
graph whose vertices consist only of routers and transit networks and
not stub networks. This is intended to keep the computational
complexity as low as possible as stub networks can be added
relatively easily through a post-processing step. This second
processing step is similar to the one used in the current OSPF
routing table calculation [Moy98], with some differences to account
for the QoS nature of routes.
Specifically, after the QoS routing table has been constructed, all
the router vertices are again considered. For each router, stub
networks whose links appear in the router's link advertisements will
be processed to determine QoS routes available to them. The QoS
routing information for a stub network is similar to that of routers
and transit networks and consists of an extension to the QoS routing
table in the form of an additional row. The columns in that new row
again correspond to paths of different hop counts, and contain both
bandwidth and next hop information. We also assume that an available
bandwidth value has been advertised for the stub network. As before,
how this value is determined is beyond the scope of this document.
The QoS routes for a stub network S are constructed as follows:
Each entry in the row corresponding to stub network S has its bw(s)
field initialized to zero and its neighbor set to null. When a stub
network S is found in the link advertisement of router V, the value
bw(S,h) in the hth column of the row corresponding to stub network S
is updated as follows:
bw(S,h) = max ( bw(S,h) ; min ( bw(V,h) , b(V,S) ) ),
where bw(V,h) is the bandwidth value of the corresponding column for
the QoS routing table row associated with router V, i.e., the
bandwidth available on an h hop path to V, and b(V,S) is the
advertised available bandwidth on the link from V to S. The above
expression essentially states that the bandwidth of a h hop path to
stub network S is updated using a path through router V, only if the
minimum of the bandwidth of the h hop path to V and the bandwidth on
the link between V and S is larger than the current value.
Apostolopoulos, et al. Experimental [Page 15]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Update of the neighbor field proceeds similarly whenever the
bandwidth of a path through V is found to be larger than or equal to
the current value. If it is larger, then the neighbor field of V in
the corresponding column replaces the current neighbor field of S.
If it is equal, then the neighbor field of V in the corresponding
column is concatenated with the existing field for S, i.e., the
current set of neighbors for V is added to the current set of
neighbors for S.
Extracting Forwarding Information from Routing Table
When the QoS paths are precomputed, the forwarding information for a
flow with given destination and bandwidth requirement needs to be
extracted from the routing table. The case of hop-by-hop routing is
simpler than that of explicit routing. This is because, only the
next hop needs to be returned instead of an explicit route.
Specifically, assume a new request to destination, say, d, and with
bandwidth requirements B. The index of the destination vertex
identifies the row in the QoS routing table that needs to be checked
to generate a path. Assuming that the QoS routing table was
constructed using the Bellman-Ford algorithm presented later in this
section, the search then proceeds by increasing index (hop) count
until an entry is found, say at hop count or column index of h, with
a value of the bw field which is equal to or larger than B. This
entry points to the initial information identifying the selected
path.
If the path computation algorithm stores multiple equal cost paths,
then some degree of load balancing can be achieved at the time of
path selection. A next hop from the list of equivalent next hops can
be chosen in a round robin manner, or randomly with a probability
that is weighted by the actual available bandwidth on the local
interface. The latter is the method used in the implementation
described in Section 4.
The case of explicit routing is discussed in Appendix D.
3. OSPF Protocol Extensions
As stated earlier, one of our goals is to limit the additions to the
existing OSPF V2 protocol, while still providing the required level
of support for QoS based routing. To this end, all of the existing
OSPF mechanisms, data structures, advertisements, and data formats
remain in place. The purpose of this section of the document is to
describe the extensions to the OSPF protocol needed to support QoS as
outlined in the previous sections.
Apostolopoulos, et al. Experimental [Page 16]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
3.1. QoS -- Optional Capabilities
The OSPF Options field is present in OSPF Hello packets, Database
Description packets and all LSAs. The Options field enables OSPF
routers to support (or not support) optional capabilities, and to
communicate their capability level to other OSPF routers. Through
this mechanism, routers of differing capabilities can be mixed within
an OSPF routing domain. Currently, the OSPF standard [Moy98]
specifies the following 5 bits in the options octet:
+-----------------------------------------------+
| * | * | DC | EA | N/P | MC | E | * |
+-----------------------------------------------+
Note that the least significant bit (`T' bit) that was used to
indicate TOS routing capability in the older OSPF specification
[Moy94] has been removed. However, for backward compatibility with
previous versions of the OSPF specification, TOS-specific information
can be included in router-LSAs, summary-LSAs and AS-external-LSAs.
We propose to reclaim the `T' bit as an indicator of router's QoS
routing capability and refer to it as the `Q' bit. In fact, QoS
capability can be viewed as an extension of the TOS-capabilities and
QoS routing as a form of TOS-based routing. A router sets this bit
in its hello packets to indicate that it is capable of supporting
such routing. When this bit is set in a router or summary links link
state advertisement, it means that there are QoS fields to process in
the packet. When this bit is set in a network link state
advertisement it means that the network described in the
advertisement is QoS capable.
We need to be careful in this approach so as to avoid confusing any
old style (i.e., RFC 1583 based) TOS routing implementations. The
TOS metric encoding rules of QoS fields introduced further in this
section will show how this is achieved. Additionally, unlike the RFC
1583 specification that unadvertised TOS metrics be treated to have
same cost as TOS 0, for the purpose of computing QOS routes,
unadvertised TOS metrics (on a hop) indicate lack of connectivity for
the specific TOS metrics (for that hop).
3.2. Encoding Resources as Extended TOS
Introduction of QoS should ideally not influence the compatibility
with existing OSPFv2 routers. To achieve this goal, necessary
extensions in packet formats must be defined in a way that either is
understood by OSPFv2 routers, ignored, or in the worst case
"gracefully" misinterpreted. Encoding of QoS metrics in the TOS
field which fortunately enough is longer in OSPF packets than
Apostolopoulos, et al. Experimental [Page 17]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
officially defined in [Alm92], allows us to mimic the new facility as
extended TOS capability. OSPFv2 routers will either disregard these
definitions or consider those unspecified. Specific precautions are
taken to prevent careless OSPF implementations from influencing
traditional TOS routers (if any) when misinterpreting the QoS
extensions.
For QoS resources, 32 combinations are available through the use of
the fifth bit in TOS fields contained in different LSAs. Since
[Alm92] defines TOS as being four bits long, this definition never
conflicts with existing values. Additionally, to prevent naive
implementations that do not take all bits of the TOS field in OSPF
packets into considerations, the definitions of the `QoS encodings'
is aligned in their semantics with the TOS encoding. Only bandwidth
and delay are specified as of today and their values map onto
`maximize throughput' and `minimize delay' if the most significant
bit is not taken into account. Accordingly, link reliability and
jitter could be defined later if necessary.
OSPF encoding RFC 1349 TOS values
___________________________________________
0 0000 normal service
2 0001 minimize monetary cost
4 0010 maximize reliability
6 0011
8 0100 maximize throughput
10 0101
12 0110
14 0111
16 1000 minimize delay
18 1001
20 1010
22 1011
24 1100
26 1101
28 1110
30 1111
Apostolopoulos, et al. Experimental [Page 18]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
OSPF encoding `QoS encoding values'
-------------------------------------------
32 10000
34 10001
36 10010
38 10011
40 10100 bandwidth
42 10101
44 10110
46 10111
48 11000 delay
50 11001
52 11010
54 11011
56 11100
58 11101
60 11110
62 11111
Representing TOS and QoS in OSPF.
3.2.1. Encoding bandwidth resource
Given the fact that the actual metric field in OSPF packets only
provides 16 bits to encode the value used and that links supporting
bandwidth ranging into Gbits/s are becoming reality, linear
representation of the available resource metric is not feasible. The
solution is exponential encoding using appropriately chosen implicit
base value and number bits for encoding mantissa and the exponent.
Detailed considerations leading to the solution described are not
presented here but can be found in [Prz95].
Given a base of 8, the 3 most significant bits should be reserved for
the exponent part and the remaining 13 for the mantissa. This allows
a simple comparison for two numbers encoded in this form, which is
often useful during implementation.
The following table shows bandwidth ranges covered when using
different exponents and the granularity of possible reservations.
Apostolopoulos, et al. Experimental [Page 19]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
exponent
value x range (2^13-1)*8^x step 8^x
-------------------------------------------------
0 8,191 1
1 65,528 8
2 524,224 64
3 4,193,792 512
4 33,550,336 4,096
5 268,402,688 32,768
6 2,147,221,504 262,144
7 17,177,772,032 2,097,152
Ranges of Exponent Values for 13 bits,
base 8 Encoding, in Bytes/s
The bandwidth encoding rule may be summarized as: "represent
available bandwidth in 16 bit field as a 3 bit exponent (with assumed
base of 8) followed by a 13 bit mantissa as shown below and advertise
2's complement of the above representation."
0 8 16
| | |
-----------------
|EXP| MANT |
-----------------
Thus, the above encoding advertises a numeric value that is
2^16 -1 -(exponential encoding of the available bandwidth):
This has the property of advertising a higher numeric value for lower
available bandwidth, a notion that is consistent with that of cost.
Although it may seem slightly pedantic to insist on the property that
less bandwidth is expressed higher values, it has, besides
consistency, a robustness aspect in it. A router with a poor OSPF
implementation could misuse or misunderstand bandwidth metric as
normal administrative cost provided to it and compute spanning trees
with a "normal" Dijkstra. The effect of a heavily congested link
advertising numerically very low cost could be disastrous in such a
scenario. It would raise the link's attractiveness for future
traffic instead of lowering it. Evidence that such considerations
are not speculative, but similar scenarios have been encountered, can
be found in [Tan89].
Apostolopoulos, et al. Experimental [Page 20]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Concluding with an example, assume a link with bandwidth of 8 Gbits/s
= 1024^3 Bytes/s, its encoding would consist of an exponent value of
6 since 1024^3= 4,096*8^6, which would then have a granularity of 8^6
or approx. 260 kBytes/s. The associated binary representation would
then be %(110) 0 1000 0000 0000% or 53,248 (8). The bandwidth cost
(advertised value) of this link when it is idle, is then the 2's
complement of the above binary representation, i.e., %(001) 1 0111
1111 1111% which corresponds to a decimal value of (2^16 - 1) -
53,248 = 12,287. Assuming now a current reservation level of 6;400
Mbits/s = 200 * 1024^2, there remains 1;600 Mbits/s of available
bandwidth on the link. The encoding of this available bandwidth of
1'600 Mbits/s is 6,400 * 8^5, which corresponds to a granularity of
8^5 or approx. 30 kBytes/s, and has a binary representation of %(101)
1 1001 0000 0000% or decimal value of 47,360. The advertised cost of
the link with this load level, is then %(010) 0 0110 1111 1111%, or
(2^16-1) -47,360 = 18,175.
Note that the cost function behaves as it should, i.e., the less
bandwidth is available on a link, the higher the cost and the less
attractive the link becomes. Furthermore, the targeted property of
better granularity for links with less bandwidth available is also
achieved. It should, however, be pointed out that the numbers given
in the above examples match exactly the resolution of the proposed
encoding, which is of course not always the case in practice. This
leaves open the question of how to encode available bandwidth values
when they do not exactly match the encoding. The standard practice
is to round it to the closest number. Because we are ultimately
interested in the cost value for which it may be better to be
pessimistic than optimistic, we choose to round costs up and,
therefore, bandwidth down.
3.2.2. Encoding Delay
Delay is encoded in microseconds using the same exponential method as
described for bandwidth except that the base is defined to be 4
instead of 8. Therefore, the maximum delay that can be expressed is
(2^13-1) *4^7 i.e., approx. 134 seconds.
3.3. Packet Formats
Given the extended TOS notation to account for QoS metrics, no
changes in packet formats are necessary except for the
(re)introduction of T-bit as the Q-bit in the options field. Routers
not understanding the Q-bit should either not consider the QoS
metrics distributed or consider those as `unknown' TOS.
Apostolopoulos, et al. Experimental [Page 21]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
To support QoS, there are additions to two Link State Advertisements,
the Router Links Advertisement and the Summary Links Advertisement.
As stated above, a router identifies itself as supporting QoS by
setting the Q-bit in the options field of the Link State Header.
When a router that supports QoS receives either the Router Links or
Summary Links Advertisement, it should parse the QoS metrics encoded
in the received Advertisement.
3.4. Calculating the Inter-area Routes
This document proposes a very limited use of OSPF areas, that is, it
is assumed that summary links advertisements exist for all networks
in the area. This document does not discuss the problem of providing
support for area address ranges and QoS metric aggregation. This is
left for further studies.
3.5. Open Issues
Support for AS External Links, Virtual Links, and incremental updates
for summary link advertisements are not addressed in this document
and are left for further study. For Virtual Links that do exist, it
is assumed for path selection that these links are non-QoS capable
even if the router advertises QoS capability. Also, as stated
earlier, this document does not address the issue of non-QoS routers
within a QoS domain.
4. A Reference Implementation based on GateD
In this section we report on the experience gained from implementing
the pre-computation based approach of Section 2.3.1 in the GateD
[Con] environment. First, we briefly introduce the GateD
environment, and then present some details on how the QoS extensions
were implemented in this environment. Finally, we discuss issues
that arose during the implementation effort and present some
measurement based results on the overhead that the QoS extensions
impose on a QoS capable router and a network of QoS routers. For
further details on the implementation study, the reader is referred
to [AGK99]. Additional performance evaluation based on simulations
can be found in [AGKT98].
4.1. The Gate Daemon (GateD) Program
GateD [Con] is a popular, public domain (9) program that provides a
platform for implementing routing protocols on hosts running the Unix
operating system. The distribution of the GateD software also
includes implementations of many popular routing protocols, including
the OSPF protocol. The GateD environment offers a variety of
services useful for implementing a routing protocol. These services
Apostolopoulos, et al. Experimental [Page 22]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
include a) support for creation and management of timers, b) memory
management, c) a simple scheduling mechanism, d) interfaces for
manipulating the host's routing table and accessing the network, and
e) route management (e.g., route prioritization and route exchange
between protocols).
All GateD processing is done within a single Unix process, and
routing protocols are implemented as one or several tasks. A GateD
task is a collection of code associated with a Unix socket. The
socket is used for the input and output requirements of the task.
The main loop of GateD contains, among other operations, a select()
call over all task sockets to determine if any read/write or error
conditions occurred in any of them. GateD implements the OSPF link
state database using a radix tree for fast access to individual link
state records. In addition, link state records for neighboring
network elements (such as adjacent routers) are linked together at
the database level with pointers. GateD maintains a single routing
table that contains routes discovered by all the active routing
protocols. Multiple routes to the same destination are prioritized
according to a set of rules and administrative preferences and only a
single route is active per destination. These routes are
periodically downloaded in the host's kernel forwarding table.
4.2. Implementing the QoS Extensions of OSPF
4.2.1. Design Objectives and Scope
One of our major design objectives was to gain substantial experience
with a functionally complete QoS routing implementation while
containing the overall implementation complexity. Thus, our
architecture was modular and aimed at reusing the existing OSPF code
with only minimal changes. QoS extensions were localized to specific
modules and their interaction with existing OSPF code was kept to a
minimum. Besides reducing the development and testing effort, this
approach also facilitated experimentation with different alternatives
for implementing the QoS specific features such as triggering
policies for link state updates and QoS route table computation.
Several of the design choices were also influenced by our assumptions
regarding the core functionalities that an early prototype
implementation of QoS routing must demonstrate. Some of the
important assumptions/requirements are:
- Support for only hop-by-hop routing. This affected the path
structure in the QoS routing table as it only needs to store next
hop information. As mentioned earlier, the structure can be
easily extended to allow construction of explicit routes.
Apostolopoulos, et al. Experimental [Page 23]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
- Support for path pre-computation. This required the creation of a
separate QoS routing table and its associated path structure, and
was motivated by the need to minimize processing overhead.
- Full integration of the QoS extensions into the GateD framework,
including configuration support, error logging, etc. This was
required to ensure a fully functional implementation that could be
used by others.
- Ability to allow experimentation with different approaches, e.g.,
use of different update and pre-computation triggering policies
with support for selection and parameterization of these policies
from the GateD configuration file.
- Decoupling from local traffic and resource management components,
i.e., packet classifiers and schedulers and local call admission.
This is supported by providing an API between QoS routing and the
local traffic management module, which hides all internal details
or mechanisms. Future implementations will be able to specify
their own mechanisms for this module.
- Interface to RSVP. The implementation assumes that RSVP [RZB+97]
is the mechanism used to request routes with specific QoS
requirements. Such requests are communicated through an interface
based on [GKR97], and used the RSVP code developed at ISI, version
4.2a2 [RZB+97].
In addition, our implementation also relies on several of the
simplifying assumptions made earlier in this document, namely:
- The scope of QoS route computation is currently limited to a
single area.
- All routers within the area are assumed to run a QoS enabled
version of OSPF, i.e., inter-operability with non-QoS aware
versions of the OSPF protocol is not considered.
- All interfaces on a router are assumed to be QoS capable.
4.2.2. Architecture
The above design decisions and assumptions resulted in the
architecture shown in Figure 2. It consists of three major
components: the signaling component (RSVP in our case); the QoS
routing component; and the traffic manager. In the rest of this
section we concentrate on the structure and operation of the QoS
routing component. As can be seen in Figure 2, the QoS routing
extensions are further divided into the following modules:
Apostolopoulos, et al. Experimental [Page 24]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
- Update trigger module determines when to advertise local link
state updates. This module implements a variety of triggering
policies: periodic, threshold based triggering, and class based
triggering. This module also implements a hold-down timer that
enforces minimum spacing between two consecutive update
triggerings from the same node.
- Pre-computation trigger module determines when to perform QoS path
pre-computation. So far, this module implements only periodic
pre-computation triggering.
- Path pre-computation module computes the QoS routing table based
on the QoS specific link state information as described in Section
2.3.1.
- Path selection and management module selects a path for a request
with particular QoS requirements, and manages it once selected,
i.e., reacts to link or reservation failures. Path selection is
performed as described in Section 2.3.1. Path management
functionality is not currently supported.
- QoS routing table module implements the QoS specific routing
table, which is maintained independently of the other GateD
routing tables.
- Tspec mapping module maps request requirements expressed in the
form of RSVP Tspecs and Rspecs into the bandwidth requirements
that QoS routing uses.
4.3. Major Implementation Issues
Mapping the above design to the framework of the GateD implementation
of OSPF led to a number of issues and design decisions. These issues
mainly fell under two categories: a) interoperation of the QoS
extensions with pre-existing similar OSPF mechanisms, and b)
structure, placement, and organization of the QoS routing table.
Next, we briefly discuss these issues and justify the resulting
design decisions.
Apostolopoulos, et al. Experimental [Page 25]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
+--------------------------------------------------+
| +-----------------------------+ |
| | QoS Route Table Computation | |
| +-----------------------------+ |
| | | |
| V | |
| +-----------------+ | |
+-------------->| QoS Route Table | | |
| | +-----------------+ | |
| | | |
| | +----------------------+ +---------------+ |
| | | Core OSPF Functions | | Precomputation| |
| | | + | | Trigger | |
| | | (Enhanced) Topology | +---------------+ |
| | | Data Base | | |
| | +----------------------+ | |
| | | | | |
| | | +----------------------------+ |
| | | | Receive and update QoS-LSA | |
| | | +----------------------------+ |
| | | | |
| | | +----------------+ |
| | | | Local Interface| |
| | | | Status Monitor | |
| | | +----------------+ |
+----------------+ | | | |
| Path Selection | | +--------------+ +----------------+ |
| & Management | | | Build and | | Link State | |
+----------------+ | | Send QoS-LSA |----------| Update Trigger | |
| | +--------------+ +----------------+ |
+----------------+ | | |
| QoS Parameter | | | |
| Mapping | | OSPF with QoS Routing Extensions | |
|----------------+ +-------------------------------------------|------+
| |
+----------------+ +----------+
| QoS Route | | Local |
| Request Client |<---------------------------------------->| Resource |
| (e.g. RSVP) | | Manager |
+----------------+ +----------+
Figure 2: The software architecture
Apostolopoulos, et al. Experimental [Page 26]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
The ability to trigger link state updates in response to changes in
bandwidth availability on interfaces is an essential component of the
QoS extensions. Mechanisms for triggering these updates and
controlling their rate have been mentioned in Section 2.2. In
addition, OSPF implements its own mechanism for triggering link state
updates as well as its own hold down timer, which may be incompatible
with what is used for the QoS link state updates. We handle such
potential conflicts as follows. First, since OSPF triggers updates
on a periodic basis with low frequency, we expect these updates to be
only a small part of the total volume of updates generated. As a
result, we chose to maintain the periodic update triggering of OSPF.
Resolving conflicts in the settings of the different hold down timer
settings requires more care. In particular, it is important to
ensure that the existing OSPF hold down timer does not interfere with
QoS updates. One option is to disable the existing OSPF timer, but
protection against transient overloads calls for some hold down
timer, albeit with a small value. As a result, the existing OSPF
hold down timer was kept, but reduced its value to 1 second. This
value is low enough (actually is the lowest possible, since GateD
timers have a maximum resolution of 1 second) so that it does not
interfere with the generation of the QoS link state updates, which
will actually often have hold down timers of their own with higher
values. An additional complexity is that the triggering of QoS link
state updates needs to be made aware of updates performed by OSPF
itself. This is necessary, as regular OSPF updates also carry
bandwidth information, and this needs to be considered by QoS updates
to properly determine when to trigger a new link state update.
Another existing OSPF mechanism that has the potential to interfere
with the extensions needed for QoS routing, is the support for
delayed acknowledgments that allows aggregation of acknowledgments
for multiple LSAs. Since link state updates are maintained in
retransmission queues until acknowledged, excessive delay in the
generation of the acknowledgement combined with the increased rates
of QoS updates may result in overflows of the retransmission queues.
To avoid these potential overflows, this mechanism was bypassed
altogether and LSAs received from neighboring routers were
immediately acknowledged. Another approach which was considered but
not implemented, was to make QoS LSAs unreliable, i.e., eliminate
their acknowledgments, so as to avoid any potential interference.
Making QoS LSAs unreliable would be a reasonable design choice
because of their higher frequency compared to the regular LSAs and
the reduced impact that the loss of a QoS LSA has on the protocol
operation. Note that the loss of a QoS LSA does not interfere with
the base operation of OSPF, and only transiently reduces the quality
of paths discovered by QoS routing.
Apostolopoulos, et al. Experimental [Page 27]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
The structure and placement of the QoS routing table also raises some
interesting implementation issues. Pre-computed paths are placed
into a QoS routing table. This table is implemented as a set of path
structures, one for each destination, which contain all the available
paths to this destination. In order to be able to efficiently locate
individual path structures, an access structure is needed. In order
to minimize the develpement effort, the radix tree structure used for
the regular GateD routing tables was reused. In addition, the QoS
routing table was kept independent of the GateD routing tables to
conform to the design goal of localizing changes and minimizing the
impact on the existing OSPF code. An additional reason for
maintaining the QoS routing separate and self-contained is that it is
re-computed under conditions that are different from those used for
the regular routing tables.
Furthermore, since the QoS routing table is re-built frequently, it
must be organized so that its computation is efficient. A common
operation during the computation of the QoS routing table is mapping
a link state database entry to the corresponding path structure. In
order to make this operation efficient, the link state database
entries were extended to contain a pointer to the corresponding path
structure. In addition, when a new QoS routing table is to be
computed, the previous one must be de-allocated. This is
accomplished by traversing the radix tree in-order, and de-allocating
each node in the tree. This full de-allocation of the QoS routing
table is potentially wasteful, especially since memory allocation and
de-allocation is an expensive operation. Furthermore, because path
pre-computations are typically not triggered by changes in topology,
the set of destinations will usually remain the same and correspond
to an unchanged radix tree. A natural optimization would then be to
de-allocate only the path structures and maintain the radix tree. A
further enhancement would be to maintain the path structures as well,
and attempt to incrementally update them only when required.
However, despite the potential gains, these optimizations have not
been included in the initial implementation. The main reason is that
they involve subtle and numerous checks to ensure the integrity of
the overall data structure at all times, e.g., correctly remove
failed destinations from the radix tree and update the tree
accordingly.
Apostolopoulos, et al. Experimental [Page 28]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
4.4. Bandwidth and Processing Overhead of QoS Routing
After completing the implementation outlined in the previous
sections, it was possible to perform an experimental study of the
cost and nature of the overhead of the QoS routing extensions
proposed in this document. In particular, using a simple setup
consisting of two interconnected routers, it is possible to measure
the cost of individual QoS routing related operations. These
operations are: a) computation of the QoS routing table, b)
selection of a path from the QoS routing table, c) generation of a
link state update, and d) reception of a link state update. Note
that the last two operations are not really specific to QoS routing
since regular OSPF also performs them. Nevertheless, we expect the
more sensitive update triggering mechanisms required for effective
QoS routing to result in increased number of updates, making the cost
of processing updates an important component of the QoS routing
overhead. An additional cost dimension is the memory required for
storing the QoS routing table. Scaling of the above costs with
increasing sizes of the topology database was investigated by
artificially populating the topology databases of the routers under
measurement.
Table 1 shows how the measured costs depend on the size of the
topology. The topology used in the measurements was built by
replicating a basic building block consisting of four routers
connected with transit networks in a rectangular arrangement. The
details of the topology and the measurements can be found in [AGK99].
The system running the GateD software was an IBM IntelliStation Z Pro
with a Pentium Pro processor at 200 MHz, 64 MBytes or real memory,
running FreeBSD 2.2.5-RELEASE and GateD 4. From the results of Table
1, one can observe that the cost of path pre-computation is not much
higher than that of the regular SPF computation. However, path pre-
computation may need to be performed much more often than the SPF
computation, and this can potentially lead to higher processing
costs. This issue was investigated in a set of subsequent
experiments, that are described later in this section. The other
cost components reported in Table 1 include memory, and it can be
seen that the QoS routing table requires roughly 80% more memory than
the regular routing table. Finally, the cost of selecting a path is
found to be very small compared to the path pre-computation times.
As expected, all the measured quantities increase as the size of the
topology increases. In particular, the storage requirements and the
processing costs for both SPF computation and QoS path pre-
computation scale almost linearly with the network size.
Apostolopoulos, et al. Experimental [Page 29]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
________________________________________________________________________
|Link_state_database_size_______|_25_|__49_|__81__|__121_|__169_|__225_|
|Regular_SPF_time_(microsec)____|215_|_440_|_747__|_1158_|_1621_|_2187_|
|Pre-computation_time_(microsec)|736_|_1622|_2883_|_4602_|_6617_|_9265_|
|SPF_routing_table_size_(bytes)_|2608|_4984|_8152_|_12112|_16864|_22408|
|QoS_routing_table_size_(bytes)_|3924|_7952|_13148|_19736|_27676|_36796|
|Path_Selection_time_(microsec)_|_.7_|_1.6_|__2.8_|__4.6_|__6.6_|__9.2_|
Table 1: Stand alone QoS routing costs
In addition to the stand alone costs reported in Table 1, it is
important to assess the actual operational load induced by QoS
routing in the context of a large network. Since it is not practical
to reproduce a large scale network in a lab setting, the approach
used was to combine simulation and measurements. Specifically, a
simulation was used to obtain a time stamped trace of QoS routing
related events that occur in a given router in a large scale network.
The trace was then used to artificially induce similar load
conditions on a real router and its adjacent links. In particular,
it was used to measure the processing load at the router and
bandwidth usage that could be attributed to QoS updates. A more
complete discussion of the measurement method and related
considerations can be found in [AGK99].
The use of a simulation further allows the use of different
configurations, where network topology is varied together with other
QoS parameters such as a) period of pre-computation, and b) threshold
for triggering link state updates. The results reported here were
derived using two types of topologies. One based on a regular but
artificial 8x8 mesh network, and another (isp) which has been used in
several previous studies [AGKT98, AT98] and that approximates the
network of a nation-wide ISP. As far as pre-computation periods are
concerned, three values of 1, 5 and 50 seconds were chosen, and for
the triggering of link state update thresholds of 10% and 80% were
used. These values were selected as they cover a wide range in terms
of precision of pre-computed paths and accuracy of the link state
information available at the routers. Also note that 1 second is the
smallest pre-computation period allowed by GateD.
Table 2 provides results on the processing load at the router driven
by the simulation trace, for the two topologies and different
combinations of QoS parameters, i.e., pre-computation period and
threshold for triggering link state updates. Table 3 gives the
bandwidth consumption of QoS updates on the links adjacent to the
router.
Apostolopoulos, et al. Experimental [Page 30]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
________________________________________________________________
|_____________________|_________Pre-computation_Period_________|
|Link_state_threshold_|___1_sec____|____5_sec____|____50_sec___|
|_________10%_________|.45%_(1.6%)_|__.29%_(2%)__|__.17%_(3%)__|
|_________80%_________|.16%_(2.4%)_|__.04%_(3%)__|_.02%_(3.8%)_|
isp
________________________________________________________________
|_________10%_________|3.37%_(2.1%)|_2.23%_(3.3%)|_1.78%_(7.7%)|
|_________80%_________|1.54%_(5.4%)|_.42%_(6.6%)_|_.14%_(10.4%)|
8x8 mesh
Table 2: Router processing load and (bandwidth blocking).
In Table 2, processing load is expressed as the percentage of the
total CPU resources that are consumed by GateD processing. The same
table also shows the routing performance that is achieved for each
combination of QoS parameters, so that comparison of the different
processing cost/routing performance trade-offs can be made. Routing
performance is measured using the bandwidth blocking ratio, defined
as the sum of requested bandwidth of the requests that were rejected
over the total offered bandwidth. As can be seen from Table 2,
processing load is low even when the QoS routing table is recomputed
every second, and LSAs are generated every time the available
bandwidth on a link changes by more than 10% of the last advertised
value. This seems to indicate that given today's processor
technology, QoS routing should not be viewed as a costly enhancement,
at least not in terms of its processing requirements. Another
general observation is that while network size has obviously an
impact, it does not seem to drastically affect the relative influence
of the different parameters. In particular, despite the differences
that exist between the isp and mesh topologies, changing the pre-
computation period or the update threshold translates into
essentially similar relative changes.
Similar conclusions can be drawn for the update traffic shown in
Table 3. In all cases, this traffic is only a small fraction of the
link's capacity. Clearly, both the router load and the link
bandwidth consumption depend on the router and link that was the
target of the measurements and will vary for different choices. The
results shown here are meant to be indicative, and a more complete
discussion can be found in [AGK99].
Apostolopoulos, et al. Experimental [Page 31]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
_______________________________________
|_Link_state_threshold_|_______________|
|_________10%__________|3112_bytes/sec_|
|_________80%__________|177_bytes/sec__|
isp
________________________________________
|_________10%__________|15438_bytes/sec_|
|_________80%__________|1053_bytes/sec__|
8x8 mesh
Table 3: Link state update traffic
Summarizing, by carrying out the implementation of the proposed QoS
routing extensions to OSPF we demonstrated that such extensions are
fairly straightforward to implement. Furthermore, by measuring the
performance of the real system we were able to demonstrate that the
overheads associated with QoS routing are not excessive, and are
definitely within the capabilities of modern processor and
workstation technology.
5. Security Considerations
The QoS extensions proposed in this document do not raise any
security considerations that are in addition to the ones associated
with regular OSPF. The security considerations of OSPF are presented
in [Moy98]. However, it should be noted that this document assumes
the availability of some entity responsible for assessing the
legitimacy of QoS requests. For example, when the protocol used for
initiating QoS requests is the RSVP protocol, this capability can be
provided through the use of RSVP Policy Decision Points and Policy
Enforcement Points as described in [YPG97]. Similarly, a policy
server enforcing the acceptability of QoS requests by implementing
decisions based on the rules and languages of [RMK+98], would also be
capable of providing the desired functionality.
Apostolopoulos, et al. Experimental [Page 32]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
APPENDICES
A. Pseudocode for the BF Based Pre-Computation Algorithm
Note: The pseudocode below assumes a hop-by-hop forwarding approach
in updating the neighbor field. The modifications needed to support
explicit route construction are straightforward. The pseudocode also
does not handle equal cost multi-paths for simplicity, but the
modification needed to add this support are straightforward.
Input:
V = set of vertices, labeled by integers 1 to N.
L = set of edges, labeled by ordered pairs (n,m) of vertex labels.
s = source vertex (at which the algorithm is executed).
For all edges (n,m) in L:
* b(n,m) = available bandwidth (according to last received update)
on interface associated with the edge between vertices n and m.
* If(n,m) outgoing interface corresponding to edge (n,m) when n is
a router.
H = maximum hop-count (at most the graph diameter).
Type:
tab_entry: record
bw = integer,
neighbor = integer 1..N.
Variables:
TT[1..N, 1..H]: topology table, whose (n,h) entry is a tab_entry
record, such that:
TT[n,h].bw is the maximum available bandwidth (as
known thus far) on a path of at most h hops
between vertices s and n,
TT[n,h].neighbor is the first hop on that path (a
neighbor of s). It is either a router or the
destination n.
S_prev: list of vertices that changed a bw value in the TT table
in the previous iteration.
S_new: list of vertices that changed a bw value (in the TT table
etc.) in the current iteration.
The Algorithm:
begin;
for n:=1 to N do /* initialization */
begin;
TT[n,0].bw := 0;
Apostolopoulos, et al. Experimental [Page 33]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
TT[n,0].neighbor := null
TT[n,1].bw := 0;
TT[n,1].neighbor := null
end;
TT[s,0].bw := infinity;
reset S_prev;
for all neighbors n of s do
begin;
TT[n,1].bw := max( TT[n,1].bw, b[s,n]);
if (TT[n,1].bw = b[s,n]) then TT[n,1].neighbor := If(s,n);
/* need to make sure we are picking the maximum */
/* bandwidth path for routers that can be reached */
/* through both networks and point-to-point links */
if (n is a router) then
S_prev := S_prev union {n}
/* only a router is added to S_prev, */
/* if it is not already included in S_prev */
else /* n is a network: */
/* proceed with network--router edges, without */
/* counting another hop */
for all (n,k) in L, k <> s, do
/* i.e., for all other neighboring routers of n */
begin;
TT[k,1].bw := max( min( TT[n,1].bw, b[n,k]), TT[k,1].bw );
/* In case k could be reached through another path */
/* (a point-to-point link or another network) with */
/* more bandwidth, we do not want to update TT[k,1].bw */
if (min( TT[n,1].bw, b[n,k]) = TT[k,1].bw )
/* If we have updated TT[k,1].bw by going through */
/* network n */
then TT[k,1].neighbor := If(s,n);
/* neighbor is interface to network n */
if ( {k} not in S_prev) then S_prev := S_prev union {k}
/* only routers are added to S_prev, but we again need */
/* to check they are not already included in S_prev */
end
end;
for h:=2 to H do /* consider all possible number of hops */
begin;
reset S_new;
for all vertices m in V do
begin;
TT[m,h].bw := TT[m,h-1].bw;
TT[m,h].neighbor := TT[m,h-1].neighbor
end;
Apostolopoulos, et al. Experimental [Page 34]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
for all vertices n in S_prev do
/* as shall become evident, S_prev contains only routers */
begin;
for all edges (n,m) in L do
if min( TT[n,h-1].bw, b[n,m]) > TT[m,h].bw then
begin;
TT[m,h].bw := min( TT[n,h-1].bw, b[n,m]);
TT[m,h].neighbor := TT[n,h-1].neighbor;
if m is a router then S_new := S_new union {m}
/* only routers are added to S_new */
else /* m is a network: */
/* proceed with network--router edges, without counting */
/* them as another hop */
for all (m,k) in L, k <> n,
/* i.e., for all other neighboring routers of m */
if min( TT[m,h].bw, b[m,k]) > TT[k,h].bw then
begin;
/* Note: still counting it as the h-th hop, as (m,k) is */
/* a network--router edge */
TT[k,h].bw := min( TT[m,h].bw, b[m,k]);
TT[k,h].neighbor := TT[m,h].neighbor;
S_new := S_new union {k}
/* only routers are added to S_new */
end
end
end;
S_prev := S_new;
/* the two lists can be handled by a toggle bit */
if S_prev=null then h=H+1 /* if no changes then exit */
end;
end.
Apostolopoulos, et al. Experimental [Page 35]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
B. On-Demand Dijkstra Algorithm for QoS Path Computation
In the main text, we described an algorithm that allows pre-
computation of QoS routes. However, it may be feasible in some
instances, e.g., limited number of requests for QoS routes, to
instead perform such computations on-demand, i.e., upon receipt of a
request for a QoS route. The benefit of such an approach is that
depending on how often recomputation of pre-computed routes is
triggered, on-demand route computation can yield better routes by
using the most recent link metrics available. Another benefit of
on-demand path computation is the associated storage saving, i.e.,
there is no need for a QoS routing table. This is essentially the
standard trade-off between memory and processing cycles.
In this section, we briefly describe how a standard Dijkstra
algorithm can, for a given destination and bandwidth requirement,
generate a minimum hop path that can accommodate the required
bandwidth and also has maximum bandwidth. Because the Dijkstra
algorithm is already used in the current OSPF route computation, only
differences from the standard algorithm are described. Also, while
for simplicity we do not consider here zero-hop edges, the
modification required for supporting them is straightforward.
The algorithm essentially performs a minimum hop path computation, on
a graph from which all edges, whose available bandwidth is less than
that requested by the flow triggering the computation, have been
removed. This can be performed either through a pre-processing step,
or while running the algorithm by checking the available bandwidth
value for any edge that is being considered (see the pseudocode that
follows). Another modification to a standard Dijkstra based minimum
hop count path computation, is that the list of equal cost next
(previous) hops which is maintained as the algorithm proceeds, needs
to be sorted according to available bandwidth. This is to allow
selection of the minimum hop path with maximum available bandwidth.
Alternatively, the algorithm could also be modified to, at each step,
only keep among equal hop count paths the one with maximum available
bandwidth. This would essentially amount to considering a cost that
is function of both hop count and available bandwidth.
Note: The pseudocode below assumes a hop-by-hop forwarding approach
in updating the neighbor field. Addition of routes to stub networks
is done in a second phase as usual. The modifications needed to
support explicit route construction are straightforward. The
pseudocode also does not handle equal cost multi-paths for
simplicity, but the modifications needed to add this support are also
easily done.
Apostolopoulos, et al. Experimental [Page 36]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Input:
V = set of vertices, labeled by integers 1 to N.
L = set of edges, labeled by ordered pairs (n,m) of vertex labels.
s = source vertex (at which the algorithm is executed).
For all edges (n,m) in L:
* b(n,m) = available bandwidth (according to last received update)
on interface associated with the edge between vertices n and m.
* If(n,m) = outgoing interface corresponding to edge (n,m) when n is
a router.
d = destination vertex.
B = requested bandwidth for the flow served.
Type:
tab_entry: record
hops = integer,
neighbor = integer 1..N,
ontree = boolean.
Variables:
TT[1..N]: topology table, whose (n) entry is a tab_entry
record, such that:
TT[n].bw is the available bandwidth (as known
thus far) on a shortest-path between
vertices s and n,
TT[n].neighbor is the first hop on that path (a
neighbor of s). It is either a router or the
destination n.
S: list of candidate vertices;
v: vertex under consideration;
The Algorithm:
begin;
for n:=1 to N do /* initialization */
begin;
TT[n].hops := infinity;
TT[n].neighbor := null;
TT[n].ontree := FALSE;
end;
TT[s].hops := 0;
reset S;
v:= s;
while v <> d do
begin;
TT[v].ontree := TRUE;
for all edges (v,m) in L and b(v,m) >= B do
begin;
Apostolopoulos, et al. Experimental [Page 37]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
if m is a router
begin;
if not TT[m].ontree then
begin;
/* bandwidth must be fulfilled since all links >= B */
if TT[m].hops > TT[v].hops + 1 then
begin
S := S union { m };
TT[m].hops := TT[v].hops + 1;
TT[m].neighbor := v;
end;
end;
end;
else /* must be a network, iterate over all attached routers */
begin; /* each network -- router edge treated as zero hop edge */
for all (m,k) in L, k <> v,
/* i.e., for all other neighboring routers of m */
if not TT[k].ontree and b(m,k) >= B then
begin;
if TT[k].hops > TT[v].hops then
begin;
S := S union { k };
TT[k].hops := TT[v].hops;
TT[k].neighbor := v;
end;
end;
end;
end; /* of all edges from the vertex under consideration */
if S is empty then
begin;
v=d; /* which will end the algorithm */
end;
else
begin;
v := first element of S;
S := S - {v}; /* remove and store the candidate to consider */
end;
end; /* from processing of the candidate list */
end.
Apostolopoulos, et al. Experimental [Page 38]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
C. Precomputation Using Dijkstra Algorithm
This appendix outlines a Dijkstra-based algorithm that allows pre-
computation of QoS routes for all destinations and bandwidth values.
The benefit of using a Dijkstra-based algorithm is a greater synergy
with existing OSPF implementations. The solution to compute all
"best" paths is to consecutively compute shortest path spanning trees
starting from a complete graph and removing links with less bandwidth
than the threshold used in the previous computation. This yields
paths with possibly better bandwidth but of course more hops.
Despite the large number of Dijkstra computations involved, several
optimizations such as incremental spanning tree computation can be
used and allow for efficient implementations in terms of complexity
as well as storage since the structure of computed paths leans itself
towards path compression [ST83]. Details including measurements and
applicability studies can be found in [Prz95] and [BP95].
A variation of this theme is to trade the "accuracy" of the pre-
computed paths, (i.e., the paths being generated may be of a larger
hop count than needed) for the benefit of using a modified version of
Dijkstra shortest path algorithm and also saving on some
computations. This loss in accuracy comes from the need to rely on
quantized bandwidth values, which are used when computing a minimum
hop count path. In other words, the range of possible bandwidth
values that can be requested by a new flow is mapped into a fixed
number of quantized values, and minimum hop count paths are generated
for each quantized value. For example, one could assume that
bandwidth values are quantized as low, medium, and high, and minimum
hop count paths are computed for each of these three values. A new
flow is then assigned to the minimum hop path that can carry the
smallest quantized value, i.e., low, medium, or high, larger than or
equal to what it requested. We restrict our discussion here to this
"quantized" version of the algorithm.
Here too, we discuss the elementary case where all edges count as
"hops", and note that the modification required for supporting zero-
hop edges is straightforward.
As with the BF algorithm, the algorithm relies on a routing table
that gets built as the algorithm progresses. The structure of the
routing table is as follows:
The QoS routing table:
- a K x Q matrix, where K is the number of vertices and Q is the
number of quantized bandwidth values.
Apostolopoulos, et al. Experimental [Page 39]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
- The (n;q) entry contains information that identifies the minimum
hop count path to destination n, which is capable of accommodating
a bandwidth request of at least bw[q] (the qth quantized bandwidth
value). It consists of two fields:
* hops: the minimal number of hops on a path between the source
node and destination n, which can accommodate a request of at
least bw[q] units of bandwidth.
* neighbor: this is the routing information associated with the
minimum hop count path to destination node n, whose available
bandwidth is at least bw[q]. With a hop-by-hop routing
approach, the neighbor information is simply the identity of
the node adjacent to the source node on that path.
The algorithm operates again on a directed graph where vertices
correspond to routers and transit networks. The metric associated
with each edge in the graph is as before the bandwidth available on
the corresponding interface, where b(n;m) is the available bandwidth
on the edge between vertices n and m. The vertex corresponding to
the router where the algorithm is being run is selected as the source
node for the purpose of path selection, and the algorithm proceeds to
compute paths to all other nodes (destinations).
Starting with the highest quantization index, Q, the algorithm
considers the indices consecutively, in decreasing order. For each
index q, the algorithm deletes from the original network topology all
links (n;m) for which b(n;m) < bw[q], and then runs on the remaining
topology a Dijkstra-based minimum hop count algorithm (10) between
the source node and all other nodes (vertices) in the graph. Note
that as with the Dijkstra used for on-demand path computation, the
elimination of links such that b(n;m) < bw[q] could also be performed
while running the algorithm.
After the algorithm terminates, the q-th column in the routing table
is updated. This amounts to recording in the hops field the hop
count value of the path that was generated by the algorithm, and by
updating the neighbor field. As before, the update of the neighbor
field depends on the scope of the path computation. In the case of a
hop-by-hop routing decision, the neighbor field is set to the
identity of the node adjacent to the source node (next hop) on the
path returned by the algorithm. However, note that in order to
ensure that the path with the maximal available bandwidth is always
chosen among all minimum hop paths that can accommodate a given
quantized bandwidth, a slightly different update mechanism of the
neighbor field needs to be used in some instances. Specifically,
when for a given row, i.e., destination node n, the value of the hops
field in column q is found equal to the value in column q+1 (here we
Apostolopoulos, et al. Experimental [Page 40]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
assume q<Q), i.e., paths that can accommodate bw[q] and bw[q+ 1] have
the same hop count, then the algorithm copies the value of the
neighbor field from entry (n;q+1) into that of entry (n;q).
Note: The pseudocode below assumes a hop-by-hop forwarding approach
in updating the neighbor field. The modifications needed to support
explicit route construction are straightforward. The pseudocode also
does not handle equal cost multi-paths for simplicity, but the
modification needed to add this support have been described above.
Details of the post-processing step of adding stub networks are
omitted.
Input:
V = set of vertices, labeled by integers 1 to N.
L = set of edges, labeled by ordered pairs (n,m) of vertex labels.
s = source vertex (at which the algorithm is executed).
bw[1..Q] = array of bandwidth values to "quantize" flow requests to.
For all edges (n,m) in L:
* b(n,m) = available bandwidth (according to last received update)
on interface associated with the edge between vertices n and m.
* If(n,m) = outgoing interface corresponding to edge (n,m) when n is
a router.
Type:
tab_entry: record
hops = integer,
neighbor = integer 1..N,
ontree = boolean.
Variables:
TT[1..N, 1..Q]: topology table, whose (n,q) entry is a tab_entry
record, such that:
TT[n,q].bw is the available bandwidth (as known
thus far) on a shortest-path between
vertices s and n accommodating bandwidth b(q),
TT[n,q].neighbor is the first hop on that path (a
neighbor of s). It is either a router or the
destination n.
S: list of candidate vertices;
v: vertex under consideration;
q: "quantize" step
The Algorithm:
begin;
for r:=1 to Q do
begin;
for n:=1 to N do /* initialization */
Apostolopoulos, et al. Experimental [Page 41]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
begin;
TT[n,r].hops := infinity;
TT[n,r].neighbor := null;
TT[n,r].ontree := FALSE;
end;
TT[s,r].hops := 0;
end;
for r:=1 to Q do
begin;
S = {s};
while S not empty do
begin;
v := first element of S;
S := S - {v}; /* remove and store the candidate to consider */
TT[v,r].ontree := TRUE;
for all edges (v,m) in L and b(v,m) >= bw[r] do
begin;
if m is a router
begin;
if not TT[m,r].ontree then
begin;
/* bandwidth must be fulfilled since all links >= bw[r] */
if TT[m,r].hops > TT[v,r].hops + 1 then
begin
S := S union { m };
TT[m,r].hops := TT[v,r].hops + 1;
TT[m,r].neighbor := v;
end;
end;
end;
else /* must be a network, iterate over all attached
routers */
begin;
for all (m,k) in L, k <> v,
/* i.e., for all other neighboring routers of m */
if not TT[k,r].ontree and b(m,k) >= bw[r] then
begin;
if TT[k,r].hops > TT[v,r].hops + 2 then
begin;
S := S union { k };
TT[k,r].hops := TT[v,r].hops + 2;
TT[k,r].neighbor := v;
end;
end;
end;
end; /* of all edges from the vertex under consideration */
end; /* from processing of the candidate list */
end; /* of "quantize" steps */
Apostolopoulos, et al. Experimental [Page 42]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
end.
D. Explicit Routing Support
As mentioned before, the scope of the path selection process can
range from simply returning the next hop on the QoS path selected for
the flow, to specifying the complete path that was computed, i.e., an
explicit route. Obviously, the information being returned by the
path selection algorithm differs in these two cases, and constructing
it imposes different requirements on the path computation algorithm
and the data structures it relies on. While the presentation of the
path computation algorithms focused on the hop-by-hop routing
approach, the same algorithms can be applied to generate explicit
routes with minor modifications. These modifications and how they
facilitate constructing explicit routes are discussed next.
The general approach to facilitate construction of explicit routes is
to update the neighbor field differently from the way it is done for
hop-by-hop routing as described in Section 2.3. Recall that in the
path computation algorithms the neighbor field is updated to reflect
the identity of the router adjacent to the source node on the partial
path computed. This facilitates returning the next hop at the source
for the specific path. In the context of explicit routing, the
neighbor information is updated to reflect the identity of the
previous router on the path.
In general, there can be multiple equivalent paths for a given hop
count. Thus, the neighbor information is stored as a list rather
than single value. Associated with each neighbor, additional
information is stored to facilitate load balancing among these
multiple paths at the time of path selection. Specifically, we store
the advertised available bandwidth on the link from the neighbor to
the destination in the entry.
With this change, the basic approach used to extract the complete
list of vertices on a path from the neighbor information in the QoS
routing table is to proceed `recursively' from the destination to the
origin vertex. The path is extracted by stepping through the
precomputed QoS routing table from vertex to vertex, and identifying
at each step the corresponding neighbor (precursor) information. The
process is described as recursive since the neighbor node identified
in one step becomes the destination node for table look up in the
next step. Once the source router is reached, the concatenation of
all the neighbor fields that have been extracted forms the desired
explicit route. This applies to algorithms of Section 2.3.1 and
Appendix C. If at a particular stage there are multiple neighbor
choices (due to equal cost multi-paths), one of them can be chosen at
random with a probability that is weighted, for example, by the
Apostolopoulos, et al. Experimental [Page 43]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
associated bandwidth on the link from the neighbor to the (current)
destination.
Specifically, assume a new request to destination, say, d, and with
bandwidth requirements B. The index of the destination vertex
identifies the row in the QoS routing table that needs to be checked
to generate a path. The row is then searched to identify a suitable
path. If the Bellman-Ford algorithm of Section 2.3.1 was used, the
search proceeds by increasing index (hop) count until an entry is
found, say at hop count or column index of h, with a value of the bw
field that is equal to or greater than B. This entry points to the
initial information identifying the selected path. If the Dijkstra
algorithm of Appendix C is used, the first quantized value qB such
that qB >= B is first identified, and the associated column then
determines the first entry in the QoS routing table that identifies
the selected path.
Once this first entry has been identified, reconstruction of the
complete list of vertices on the path proceeds similarly, whether the
table was built using the algorithm of Section 2.3.1 or Appendix C.
Specifically, in both cases, the neighbor field in each entry points
to the previous node on the path from the source node and with the
same bandwidth capabilities as those associated with the current
entry. The complete path is, therefore, reconstructed by following
the pointers provided by the neighbor field of successive entries.
In the case of the Bellman-Ford algorithm of Section 2.3.1, this
means moving backwards in the table from column to column, using at
each step the row index pointed to by the neighbor field of the entry
in the previous column. Each time, the corresponding vertex index
specified in the neighbor field is pre-pended to the list of vertices
constructed so far. Since we start at column h, the process ends
when the first column is reached, i.e., after h steps, at which point
the list of vertices making up the path has been reconstructed.
In the case of the Dijkstra algorithm of Appendix C, the backtracking
process is similar although slightly different because of the
different relation between paths and columns in the routing table,
i.e., a column now corresponds to a quantized bandwidth value instead
of a hop count. The backtracking now proceeds along the column
corresponding to the quantized bandwidth value needed to satisfy the
bandwidth requirements of the flow. At each step, the vertex index
specified in the neighbor field is pre-pended to the list of vertices
constructed so far, and is used to identify the next row index to
move to. The process ends when an entry is reached whose neighbor
field specifies the origin vertex of the flow. Note that since there
are as many rows in the table as there are vertices in the graph,
i.e., N, it could take up to N steps before the process terminates.
Apostolopoulos, et al. Experimental [Page 44]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Note that the identification of the first entry in the routing table
is identical to what was described for the hop-by-hop routing case.
However, as described in this section, the update of the neighbor
fields while constructing the QoS routing tables, is being performed
differently in the explicit and hop-by-hop routing cases. Clearly,
two different neighbor fields can be kept in each entry and updates
to both could certainly be performed jointly, if support for both
xplicit routing and hop-by-hop routing is needed.
Endnotes
1. In this document we commit the abuse of notation of calling a
"network" the interconnection of routers and networks through
which we attempt to compute a QoS path.
2. This is true for uni-cast flows, but in the case of multi-cast
flows, hop-by-hop and an explicit routing clearly have different
implications.
3. Some hysteresis mechanism should be added to suppress updates when
the metric value oscillates around a class boundary.
4. In this document, we use the terms node and vertex
interchangeably.
5. Various hybrid methods can also be envisioned, e.g., periodic
computations except if more than a given number of updates are
received within a shorter interval, or periodic updates except if
the change in metrics corresponding to a given update exceeds a
certain threshold. Such variations are, however, not considered
in this document.
6. Modifications to support explicit routing are discussed in
Appendix D.
7. Note, that this does not say anything on whether to differentiate
between outgoing and incoming bandwidth on a shared media network.
As a matter of fact, a reasonable option is to set the incoming
bandwidth (from network to router) to infinity, and only use the
outgoing bandwidth value to characterize bandwidth availability on
the shared network.
8. exponent in parenthesis
9. Access to some of the more recent versions of the GateD software
is restricted to the GateD consortium members.
Apostolopoulos, et al. Experimental [Page 45]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
10. Note that a Breadth-First-Search (BFS) algorithm [CLR90] could
also be used. It has a lower complexity, but would not allow
reuse of existing code in an OSPF implementation.
References
[AGK99] G. Apostolopoulos, R. Guerin, and S. Kamat. Implementation
and performance meassurements of QoS routing extensions to
OSPF. In Proceedings of INFOCOM'99, pages 680--688, New
York, NY, March 1999.
[AGKT98] G. Apostolopoulos, R. Guerin, S. Kamat, and S. K. Tripathi.
QoS routing: A performance perspective. In Proceedings of
ACM SIGCOMM'98, pages 17--28, Vancouver, Canada, October
[Alm92] Almquist, P., "Type of Service in the Internet Protocol
Suite", RFC 1349, July 1992.
[AT98] G. Apostolopoulos and S. K. Tripathi. On reducing the
processing cost of on-demand QoS path computation. In
Proceedings of ICNP'98, pages 80--89, Austin, TX, October
1998.
[BP95] J.-Y. Le Boudec and T. Przygienda. A Route Pre-Computation
Algorithm for Integrated Services Networks. Journal of
Network and Systems Management, 3(4), 1995.
[Car79] B. Carre. Graphs and Networks. Oxford University Press,
ISBN 0-19-859622-7, Oxford, UK, 1979.
[CLR90] T. H. Cormen, C. E. Leiserson, and R. L. Rivest.
Introduction to Algorithms. MIT Press, Cambridge, MA, 1990.
[Con] Merit GateD Consortium. The Gate Daemon (GateD) project.
[GJ79] M.R. Garey and D.S. Johnson. Computers and Intractability.
Freeman, San Francisco, 1979.
[GKH97] R. Guerin, S. Kamat, and S. Herzog. QoS Path Management
with RSVP. In Proceedings of the 2nd IEEE Global Internet
Mini-Conference, pages 1914-1918, Phoenix, AZ, November
[GKR97] Guerin, R., Kamat, S. and E. Rosen, "An Extended RSVP
Routing Interface, Work in Progress.
[GLG+97] Der-Hwa G., Li, T., Guerin, R., Rosen, E. and S. Kamat,
"Setting Up Reservations on Explicit Paths using RSVP", Work
in Progress.
Apostolopoulos, et al. Experimental [Page 46]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
[GO99] R. Guerin and A. Orda. QoS-Based Routing in Networks with
Inaccurate Information: Theory and Algorithms. IEEE/ACM
Transactions on Networking, 7(3):350--364, June 1999.
[GOW97] R. Guerin, A. Orda, and D. Williams. QoS Routing Mechanisms
and OSPF Extensions. In Proceedings of the 2nd IEEE Global
Internet Mini-Conference, pages 1903-1908, Phoenix, AZ,
November 1997.
[KNB98] Nichols, K., Blake, S., Baker F. and D. Black, "Definition
of the Differentiated Services Field (DS Field) in the IPv4
and IPv6 Headers", RFC 2474, December 1998.
[LO98] D. H. Lorenz and A. Orda. QoS Routing in Networks with
Uncertain Parameters. IEEE/ACM Transactions on Networking,
6(6):768--778, December 1998.
[Moy94] Moy, J., "OSPF Version 2", RFC 1583, March 1994.
[Moy98] Moy, J., "OSPF Version 2", STD 54, RFC 2328, April 1998.
[Prz95] A. Przygienda. Link State Routing with QoS in ATM LANs.
Ph.D. Thesis Nr. 11051, Swiss Federal Institute of
Technology, April 1995.
[RMK+98] R. Rajan, J. C. Martin, S. Kamat, M. See, R. Chaudhury, D.
Verma, G. Powers, and R. Yavatkar. Schema for
differentiated services and integrated services in networks.
INTERNET-DRAFT, October 1998. work in progress.
[RZB+97] Braden, R., Editor, Zhang, L., Berson, S., Herzog, S. and S.
Jamin, "Resource reSerVation Protocol (RSVP) Version 1,
Functional Specification", RFC 2205, September 1997.
[SPG97] Shenker, S., Partridge, C. and R. Guerin, "Specification of
Guaranteed Quality of Service", RFC 2212, November 1997.
[ST83] D.D. Sleator and R.E. Tarjan. A Data Structure for Dynamic
Trees. Journal of Computer Systems, 26, 1983.
[Tan89] A. Tannenbaum. Computer Networks. Addisson Wesley, 1989.
[YPG97] Yavatkar, R., Pendarakis, D. and R. Guerin, "A Framework for
Policy-based Admission Control", INTERNET-DRAFT, April 1999.
Work in Progress.
Apostolopoulos, et al. Experimental [Page 47]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Authors' Addresses
George Apostolopoulos
IBM T.J. Watson Research Center
P.O. Box 704
Yorktown Heights, NY 10598
Phone: +1 914 784-6204
Fax: +1 914 784-6205
EMail: georgeap@watson.ibm.com
Roch Guerin
University Of Pennsylvania
Department of Electrical Engineering, Rm 367 GRW
200 South 33rd Street
Philadelphia, PA 19104--6390
Phone: +1 215-898-9351
EMail: guerin@ee.upenn.edu
Sanjay Kamat
Bell Laboratories
Lucent Technologies
Room 4C-510
101 Crawfords Corner Road
Holmdel, NJ 07733
Phone: (732) 949-5936
email: sanjayk@dnrc.bell-labs.com
Ariel Orda
Dept. Electrical Engineering
Technion - I.I.T
Haifa, 32000 - ISRAEL
Phone: +011 972-4-8294646
Fax: +011 972-4-8323041
EMail: ariel@ee.technion.ac.il
Apostolopoulos, et al. Experimental [Page 48]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Tony Przygienda
Siara Systems
300 Ferguson Drive
Moutain View
California 94043
Phone: +1 732 949-5936
Email: prz@siara.com
Doug Williams
IBM T.J. Watson Research Center
P.O. Box 704
Yorktown Heights, NY 10598
Phone: +1 914 784-5047
Fax: +1 914 784-6318
EMail: dougw@watson.ibm.com
Apostolopoulos, et al. Experimental [Page 49]
RFC 2676 QoS Routing Mechanisms and OSPF Extensions August 1999
Full Copyright Statement
Copyright (C) The Internet Society (1999). All Rights Reserved.
This document and translations of it may be copied and furnished to
others, and derivative works that comment on or otherwise explain it
or assist in its implementation may be prepared, copied, published
and distributed, in whole or in part, without restriction of any
kind, provided that the above copyright notice and this paragraph are
included on all such copies and derivative works. However, this
document itself may not be modified in any way, such as by removing
the copyright notice or references to the Internet Society or other
Internet organizations, except as needed for the purpose of
developing Internet standards in which case the procedures for
copyrights defined in the Internet Standards process must be
followed, or as required to translate it into languages other than
English.
The limited permissions granted above are perpetual and will not be
revoked by the Internet Society or its successors or assigns.
This document and the information contained herein is provided on an
"AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
Acknowledgement
Funding for the RFC Editor function is currently provided by the
Internet Society.
Apostolopoulos, et al. Experimental [Page 50]