[RFC Home] [TEXT|PDF|HTML] [Tracker] [IPR] [Info page]
INFORMATIONAL
Network Working Group J. Hadi Salim
Request for Comments: 2884 Nortel Networks
Category: Informational U. Ahmed
Carleton University
July 2000
Performance Evaluation of Explicit Congestion Notification (ECN)
in IP Networks
Status of this Memo
This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2000). All Rights Reserved.
Abstract
This memo presents a performance study of the Explicit Congestion
Notification (ECN) mechanism in the TCP/IP protocol using our
implementation on the Linux Operating System. ECN is an end-to-end
congestion avoidance mechanism proposed by [6] and incorporated into
RFC 2481[7]. We study the behavior of ECN for both bulk and
transactional transfers. Our experiments show that there is
improvement in throughput over NON ECN (TCP employing any of Reno,
SACK/FACK or NewReno congestion control) in the case of bulk
transfers and substantial improvement for transactional transfers.
A more complete pdf version of this document is available at:
http://www7.nortel.com:8080/CTL/ecnperf.pdf
This memo in its current revision is missing a lot of the visual
representations and experimental results found in the pdf version.
1. Introduction
In current IP networks, congestion management is left to the
protocols running on top of IP. An IP router when congested simply
drops packets. TCP is the dominant transport protocol today [26].
TCP infers that there is congestion in the network by detecting
packet drops (RFC 2581). Congestion control algorithms [11] [15] [21]
are then invoked to alleviate congestion. TCP initially sends at a
higher rate (slow start) until it detects a packet loss. A packet
loss is inferred by the receipt of 3 duplicate ACKs or detected by a
Salim & Ahmed Informational [Page 1]
RFC 2884 ECN in IP Networks July 2000
timeout. The sending TCP then moves into a congestion avoidance state
where it carefully probes the network by sending at a slower rate
(which goes up until another packet loss is detected). Traditionally
a router reacts to congestion by dropping a packet in the absence of
buffer space. This is referred to as Tail Drop. This method has a
number of drawbacks (outlined in Section 2). These drawbacks coupled
with the limitations of end-to-end congestion control have led to
interest in introducing smarter congestion control mechanisms in
routers. One such mechanism is Random Early Detection (RED) [9]
which detects incipient congestion and implicitly signals the
oversubscribing flow to slow down by dropping its packets. A RED-
enabled router detects congestion before the buffer overflows, based
on a running average queue size, and drops packets probabilistically
before the queue actually fills up. The probability of dropping a new
arriving packet increases as the average queue size increases above a
low water mark minth, towards higher water mark maxth. When the
average queue size exceeds maxth all arriving packets are dropped.
An extension to RED is to mark the IP header instead of dropping
packets (when the average queue size is between minth and maxth;
above maxth arriving packets are dropped as before). Cooperating end
systems would then use this as a signal that the network is congested
and slow down. This is known as Explicit Congestion Notification
(ECN). In this paper we study an ECN implementation on Linux for
both the router and the end systems in a live network. The memo is
organized as follows. In Section 2 we give an overview of queue
management in routers. Section 3 gives an overview of ECN and the
changes required at the router and the end hosts to support ECN.
Section 4 defines the experimental testbed and the terminologies used
throughout this memo. Section 5 introduces the experiments that are
carried out, outlines the results and presents an analysis of the
results obtained. Section 6 concludes the paper.
2. Queue Management in routers
TCP's congestion control and avoidance algorithms are necessary and
powerful but are not enough to provide good service in all
circumstances since they treat the network as a black box. Some sort
of control is required from the routers to complement the end system
congestion control mechanisms. More detailed analysis is contained in
[19]. Queue management algorithms traditionally manage the length of
packet queues in the router by dropping packets only when the buffer
overflows. A maximum length for each queue is configured. The router
will accept packets till this maximum size is exceeded, at which
point it will drop incoming packets. New packets are accepted when
buffer space allows. This technique is known as Tail Drop. This
method has served the Internet well for years, but has the several
drawbacks. Since all arriving packets (from all flows) are dropped
Salim & Ahmed Informational [Page 2]
RFC 2884 ECN in IP Networks July 2000
when the buffer overflows, this interacts badly with the congestion
control mechanism of TCP. A cycle is formed with a burst of drops
after the maximum queue size is exceeded, followed by a period of
underutilization at the router as end systems back off. End systems
then increase their windows simultaneously up to a point where a
burst of drops happens again. This phenomenon is called Global
Synchronization. It leads to poor link utilization and lower overall
throughput [19] Another problem with Tail Drop is that a single
connection or a few flows could monopolize the queue space, in some
circumstances. This results in a lock out phenomenon leading to
synchronization or other timing effects [19]. Lastly, one of the
major drawbacks of Tail Drop is that queues remain full for long
periods of time. One of the major goals of queue management is to
reduce the steady state queue size[19]. Other queue management
techniques include random drop on full and drop front on full [13].
2.1. Active Queue Management
Active queue management mechanisms detect congestion before the queue
overflows and provide an indication of this congestion to the end
nodes [7]. With this approach TCP does not have to rely only on
buffer overflow as the indication of congestion since notification
happens before serious congestion occurs. One such active management
technique is RED.
2.1.1. Random Early Detection
Random Early Detection (RED) [9] is a congestion avoidance mechanism
implemented in routers which works on the basis of active queue
management. RED addresses the shortcomings of Tail Drop. A RED
router signals incipient congestion to TCP by dropping packets
probabilistically before the queue runs out of buffer space. This
drop probability is dependent on a running average queue size to
avoid any bias against bursty traffic. A RED router randomly drops
arriving packets, with the result that the probability of dropping a
packet belonging to a particular flow is approximately proportional
to the flow's share of bandwidth. Thus, if the sender is using
relatively more bandwidth it gets penalized by having more of its
packets dropped. RED operates by maintaining two levels of
thresholds minimum (minth) and maximum (maxth). It drops a packet
probabilistically if and only if the average queue size lies between
the minth and maxth thresholds. If the average queue size is above
the maximum threshold, the arriving packet is always dropped. When
the average queue size is between the minimum and the maximum
threshold, each arriving packet is dropped with probability pa, where
pa is a function of the average queue size. As the average queue
length varies between minth and maxth, pa increases linearly towards
a configured maximum drop probability, maxp. Beyond maxth, the drop
Salim & Ahmed Informational [Page 3]
RFC 2884 ECN in IP Networks July 2000
probability is 100%. Dropping packets in this way ensures that when
some subset of the source TCP packets get dropped and they invoke
congestion avoidance algorithms that will ease the congestion at the
gateway. Since the dropping is distributed across flows, the problem
of global synchronization is avoided.
3. Explicit Congestion Notification
Explicit Congestion Notification is an extension proposed to RED
which marks a packet instead of dropping it when the average queue
size is between minth and maxth [7]. Since ECN marks packets before
congestion actually occurs, this is useful for protocols like TCP
that are sensitive to even a single packet loss. Upon receipt of a
congestion marked packet, the TCP receiver informs the sender (in the
subsequent ACK) about incipient congestion which will in turn trigger
the congestion avoidance algorithm at the sender. ECN requires
support from both the router as well as the end hosts, i.e. the end
hosts TCP stack needs to be modified. Packets from flows that are not
ECN capable will continue to be dropped by RED (as was the case
before ECN).
3.1. Changes at the router
Router side support for ECN can be added by modifying current RED
implementations. For packets from ECN capable hosts, the router marks
the packets rather than dropping them (if the average queue size is
between minth and maxth). It is necessary that the router identifies
that a packet is ECN capable, and should only mark packets that are
from ECN capable hosts. This uses two bits in the IP header. The ECN
Capable Transport (ECT) bit is set by the sender end system if both
the end systems are ECN capable (for a unicast transport, only if
both end systems are ECN-capable). In TCP this is confirmed in the
pre-negotiation during the connection setup phase (explained in
Section 3.2). Packets encountering congestion are marked by the
router using the Congestion Experienced (CE) (if the average queue
size is between minth and maxth) on their way to the receiver end
system (from the sender end system), with a probability proportional
to the average queue size following the procedure used in RED
(RFC2309) routers. Bits 10 and 11 in the IPV6 header are proposed
respectively for the ECT and CE bits. Bits 6 and 7 of the IPV4 header
DSCP field are also specified for experimental purposes for the ECT
and CE bits respectively.
3.2. Changes at the TCP Host side
The proposal to add ECN to TCP specifies two new flags in the
reserved field of the TCP header. Bit 9 in the reserved field of the
TCP header is designated as the ECN-Echo (ECE) flag and Bit 8 is
Salim & Ahmed Informational [Page 4]
RFC 2884 ECN in IP Networks July 2000
designated as the Congestion Window Reduced (CWR) flag. These two
bits are used both for the initializing phase in which the sender and
the receiver negotiate the capability and the desire to use ECN, as
well as for the subsequent actions to be taken in case there is
congestion experienced in the network during the established state.
There are two main changes that need to be made to add ECN to TCP to
an end system and one extension to a router running RED.
1. In the connection setup phase, the source and destination TCPs
have to exchange information about their desire and/or capability to
use ECN. This is done by setting both the ECN-Echo flag and the CWR
flag in the SYN packet of the initial connection phase by the sender;
on receipt of this SYN packet, the receiver will set the ECN-Echo
flag in the SYN-ACK response. Once this agreement has been reached,
the sender will thereon set the ECT bit in the IP header of data
packets for that flow, to indicate to the network that it is capable
and willing to participate in ECN. The ECT bit is set on all packets
other than pure ACK's.
2. When a router has decided from its active queue management
mechanism, to drop or mark a packet, it checks the IP-ECT bit in the
packet header. It sets the CE bit in the IP header if the IP-ECT bit
is set. When such a packet reaches the receiver, the receiver
responds by setting the ECN-Echo flag (in the TCP header) in the next
outgoing ACK for the flow. The receiver will continue to do this in
subsequent ACKs until it receives from the sender an indication that
it (the sender) has responded to the congestion notification.
3. Upon receipt of this ACK, the sender triggers its congestion
avoidance algorithm by halving its congestion window, cwnd, and
updating its congestion window threshold value ssthresh. Once it has
taken these appropriate steps, the sender sets the CWR bit on the
next data outgoing packet to tell the receiver that it has reacted to
the (receiver's) notification of congestion. The receiver reacts to
the CWR by halting the sending of the congestion notifications (ECE)
to the sender if there is no new congestion in the network.
Note that the sender reaction to the indication of congestion in the
network (when it receives an ACK packet that has the ECN-Echo flag
set) is equivalent to the Fast Retransmit/Recovery algorithm (when
there is a congestion loss) in NON-ECN-capable TCP i.e. the sender
halves the congestion window cwnd and reduces the slow start
threshold ssthresh. Fast Retransmit/Recovery is still available for
ECN capable stacks for responding to three duplicate acknowledgments.
Salim & Ahmed Informational [Page 5]
RFC 2884 ECN in IP Networks July 2000
4. Experimental setup
For testing purposes we have added ECN to the Linux TCP/IP stack,
kernels version 2.0.32. 2.2.5, 2.3.43 (there were also earlier
revisions of 2.3 which were tested). The 2.0.32 implementation
conforms to RFC 2481 [7] for the end systems only. We have also
modified the code in the 2.1,2.2 and 2.3 cases for the router portion
as well as end system to conform to the RFC. An outdated version of
the 2.0 code is available at [18]. Note Linux version 2.0.32
implements TCP Reno congestion control while kernels >= 2.2.0 default
to New Reno but will opt for a SACK/FACK combo when the remote end
understands SACK. Our initial tests were carried out with the 2.0
kernel at the end system and 2.1 (pre 2.2) for the router part. The
majority of the test results here apply to the 2.0 tests. We did
repeat these tests on a different testbed (move from Pentium to
Pentium-II class machines)with faster machines for the 2.2 and 2.3
kernels, so the comparisons on the 2.0 and 2.2/3 are not relative.
We have updated this memo release to reflect the tests against SACK
and New Reno.
4.1. Testbed setup
----- ----
| ECN | | ECN |
| ON | | OFF |
data direction ---->> ----- ----
| |
server | |
---- ------ ------ | |
| | | R1 | | R2 | | |
| | -----| | ---- | | ----------------------
---- ------ ^ ------ |
^ |
| -----
congestion point ___| | C |
| |
-----
The figure above shows our test setup.
All the physical links are 10Mbps ethernet. Using Class Based
Queuing (CBQ) [22], packets from the data server are constricted to a
1.5Mbps pipe at the router R1. Data is always retrieved from the
server towards the clients labelled , "ECN ON", "ECN OFF", and "C".
Since the pipe from the server is 10Mbps, this creates congestion at
the exit from the router towards the clients for competing flows. The
machines labeled "ECN ON" and "ECN OFF" are running the same version
Salim & Ahmed Informational [Page 6]
RFC 2884 ECN in IP Networks July 2000
of Linux and have exactly the same hardware configuration. The server
is always ECN capable (and can handle NON ECN flows as well using the
standard congestion algorithms). The machine labeled "C" is used to
create congestion in the network. Router R2 acts as a path-delay
controller. With it we adjust the RTT the clients see. Router R1
has RED implemented in it and has capability for supporting ECN
flows. The path-delay router is a PC running the Nistnet [16]
package on a Linux platform. The latency of the link for the
experiments was set to be 20 millisecs.
4.2. Validating the Implementation
We spent time validating that the implementation was conformant to
the specification in RFC 2481. To do this, the popular tcpdump
sniffer [24] was modified to show the packets being marked. We
visually inspected tcpdump traces to validate the conformance to the
RFC under a lot of different scenarios. We also modified tcptrace
[25] in order to plot the marked packets for visualization and
analysis.
Both tcpdump and tcptrace revealed that the implementation was
conformant to the RFC.
4.3. Terminology used
This section presents background terminology used in the next few
sections.
* Congesting flows: These are TCP flows that are started in the
background so as to create congestion from R1 towards R2. We use the
laptop labeled "C" to introduce congesting flows. Note that "C" as is
the case with the other clients retrieves data from the server.
* Low, Moderate and High congestion: For the case of low congestion
we start two congesting flows in the background, for moderate
congestion we start five congesting flows and for the case of high
congestion we start ten congesting flows in the background.
* Competing flows: These are the flows that we are interested in.
They are either ECN TCP flows from/to "ECN ON" or NON ECN TCP flows
from/to "ECN OFF".
* Maximum drop rate: This is the RED parameter that sets the maximum
probability of a packet being marked at the router. This corresponds
to maxp as explained in Section 2.1.
Salim & Ahmed Informational [Page 7]
RFC 2884 ECN in IP Networks July 2000
Our tests were repeated for varying levels of congestion with varying
maximum drop rates. The results are presented in the subsequent
sections.
* Low, Medium and High drop probability: We use the term low
probability to mean a drop probability maxp of 0.02, medium
probability for 0.2 and high probability for 0.5. We also
experimented with drop probabilities of 0.05, 0.1 and 0.3.
* Goodput: We define goodput as the effective data rate as observed
by the user, i.e., if we transmitted 4 data packets in which two of
them were retransmitted packets, the efficiency is 50% and the
resulting goodput is 2*packet size/time taken to transmit.
* RED Region: When the router's average queue size is between minth
and maxth we denote that we are operating in the RED region.
4.4. RED parameter selection
In our initial testing we noticed that as we increase the number of
congesting flows the RED queue degenerates into a simple Tail Drop
queue. i.e. the average queue exceeds the maximum threshold most of
the times. Note that this phenomena has also been observed by [5]
who proposes a dynamic solution to alleviate it by adjusting the
packet dropping probability "maxp" based on the past history of the
average queue size. Hence, it is necessary that in the course of our
experiments the router operate in the RED region, i.e., we have to
make sure that the average queue is maintained between minth and
maxth. If this is not maintained, then the queue acts like a Tail
Drop queue and the advantages of ECN diminish. Our goal is to
validate ECN's benefits when used with RED at the router. To ensure
that we were operating in the RED region we monitored the average
queue size and the actual queue size in times of low, moderate and
high congestion and fine-tuned the RED parameters such that the
average queue zones around the RED region before running the
experiment proper. Our results are, therefore, not influenced by
operating in the wrong RED region.
5. The Experiments
We start by making sure that the background flows do not bias our
results by computing the fairness index [12] in Section 5.1. We
proceed to carry out the experiments for bulk transfer presenting the
results and analysis in Section 5.2. In Section 5.3 the results for
transactional transfers along with analysis is presented. More
details on the experimental results can be found in [27].
Salim & Ahmed Informational [Page 8]
RFC 2884 ECN in IP Networks July 2000
5.1. Fairness
In the course of the experiments we wanted to make sure that our
choice of the type of background flows does not bias the results that
we collect. Hence we carried out some tests initially with both ECN
and NON ECN flows as the background flows. We repeated the
experiments for different drop probabilities and calculated the
fairness index [12]. We also noticed (when there were equal number
of ECN and NON ECN flows) that the number of packets dropped for the
NON ECN flows was equal to the number of packets marked for the ECN
flows, showing thereby that the RED algorithm was fair to both kind
of flows.
Fairness index: The fairness index is a performance metric described
in [12]. Jain [12] postulates that the network is a multi-user
system, and derives a metric to see how fairly each user is treated.
He defines fairness as a function of the variability of throughput
across users. For a given set of user throughputs (x1, x2...xn), the
fairness index to the set is defined as follows:
f(x1,x2,.....,xn) = square((sum[i=1..n]xi))/(n*sum[i=1..n]square(xi))
The fairness index always lies between 0 and 1. A value of 1
indicates that all flows got exactly the same throughput. Each of
the tests was carried out 10 times to gain confidence in our results.
To compute the fairness index we used FTP to generate traffic.
Experiment details: At time t = 0 we start 2 NON ECN FTP sessions in
the background to create congestion. At time t=20 seconds we start
two competing flows. We note the throughput of all the flows in the
network and calculate the fairness index. The experiment was carried
out for various maximum drop probabilities and for various congestion
levels. The same procedure is repeated with the background flows as
ECN. The fairness index was fairly constant in both the cases when
the background flows were ECN and NON ECN indicating that there was
no bias when the background flows were either ECN or NON ECN.
Max Fairness Fairness
Drop With BG With BG
Prob flows ECN flows NON ECN
0.02 0.996888 0.991946
0.05 0.995987 0.988286
0.1 0.985403 0.989726
0.2 0.979368 0.983342
Salim & Ahmed Informational [Page 9]
RFC 2884 ECN in IP Networks July 2000
With the observation that the nature of background flows does not
alter the results, we proceed by using the background flows as NON
ECN for the rest of the experiments.
5.2. Bulk transfers
The metric we chose for bulk transfer is end user throughput.
Experiment Details: All TCP flows used are RENO TCP. For the case of
low congestion we start 2 FTP flows in the background at time 0. Then
after about 20 seconds we start the competing flows, one data
transfer to the ECN machine and the second to the NON ECN machine.
The size of the file used is 20MB. For the case of moderate
congestion we start 5 FTP flows in the background and for the case of
high congestion we start 10 FTP flows in the background. We repeat
the experiments for various maximum drop rates each repeated for a
number of sets.
Observation and Analysis:
We make three key observations:
1) As the congestion level increases, the relative advantage for ECN
increases but the absolute advantage decreases (expected, since there
are more flows competing for the same link resource). ECN still does
better than NON ECN even under high congestion. Infering a sample
from the collected results: at maximum drop probability of 0.1, for
example, the relative advantage of ECN increases from 23% to 50% as
the congestion level increases from low to high.
2) Maintaining congestion levels and varying the maximum drop
probability (MDP) reveals that the relative advantage of ECN
increases with increasing MDP. As an example, for the case of high
congestion as we vary the drop probability from 0.02 to 0.5 the
relative advantage of ECN increases from 10% to 60%.
3) There were hardly any retransmissions for ECN flows (except the
occasional packet drop in a minority of the tests for the case of
high congestion and low maximum drop probability).
We analyzed tcpdump traces for NON ECN with the help of tcptrace and
observed that there were hardly any retransmits due to timeouts.
(Retransmit due to timeouts are inferred by counting the number of 3
DUPACKS retransmit and subtracting them from the total recorded
number of retransmits). This means that over a long period of time
(as is the case of long bulk transfers), the data-driven loss
recovery mechanism of the Fast Retransmit/Recovery algorithm is very
effective. The algorithm for ECN on congestion notification from ECE
Salim & Ahmed Informational [Page 10]
RFC 2884 ECN in IP Networks July 2000
is the same as that for a Fast Retransmit for NON ECN. Since both are
operating in the RED region, ECN barely gets any advantage over NON
ECN from the signaling (packet drop vs. marking).
It is clear, however, from the results that ECN flows benefit in bulk
transfers. We believe that the main advantage of ECN for bulk
transfers is that less time is spent recovering (whereas NON ECN
spends time retransmitting), and timeouts are avoided altogether.
[23] has shown that even with RED deployed, TCP RENO could suffer
from multiple packet drops within the same window of data, likely to
lead to multiple congestion reactions or timeouts (these problems are
alleviated by ECN). However, while TCP Reno has performance problems
with multiple packets dropped in a window of data, New Reno and SACK
have no such problems.
Thus, for scenarios with very high levels of congestion, the
advantages of ECN for TCP Reno flows could be more dramatic than the
advantages of ECN for NewReno or SACK flows. An important
observation to make from our results is that we do not notice
multiple drops within a single window of data. Thus, we would expect
that our results are not heavily influenced by Reno's performance
problems with multiple packets dropped from a window of data. We
repeated these tests with ECN patched newer Linux kernels. As
mentioned earlier these kernels would use a SACK/FACK combo with a
fallback to New Reno. SACK can be selectively turned off (defaulting
to New Reno). Our results indicate that ECN still improves
performance for the bulk transfers. More results are available in the
pdf version[27]. As in 1) above, maintaining a maximum drop
probability of 0.1 and increasing the congestion level, it is
observed that ECN-SACK improves performance from about 5% at low
congestion to about 15% at high congestion. In the scenario where
high congestion is maintained and the maximum drop probability is
moved from 0.02 to 0.5, the relative advantage of ECN-SACK improves
from 10% to 40%. Although this numbers are lower than the ones
exhibited by Reno, they do reflect the improvement that ECN offers
even in the presence of robust recovery mechanisms such as SACK.
5.3. Transactional transfers
We model transactional transfers by sending a small request and
getting a response from a server before sending the next request. To
generate transactional transfer traffic we use Netperf [17] with the
CRR (Connect Request Response) option. As an example let us assume
that we are retrieving a small file of say 5 - 20 KB, then in effect
we send a small request to the server and the server responds by
sending us the file. The transaction is complete when we receive the
complete file. To gain confidence in our results we carry the
simulation for about one hour. For each test there are a few thousand
Salim & Ahmed Informational [Page 11]
RFC 2884 ECN in IP Networks July 2000
of these requests and responses taking place. Although not exactly
modeling HTTP 1.0 traffic, where several concurrent sessions are
opened, Netperf-CRR is nevertheless a close approximation. Since
Netperf-CRR waits for one connection to complete before opening the
next one (0 think time), that single connection could be viewed as
the slowest response in the set of the opened concurrent sessions (in
HTTP). The transactional data sizes were selected based on [2] which
indicates that the average web transaction was around 8 - 10 KB; The
smaller (5KB) size was selected to guestimate the size of
transactional processing that may become prevalent with policy
management schemes in the diffserv [4] context. Using Netperf we are
able to initiate these kind of transactional transfers for a variable
length of time. The main metric of interest in this case is the
transaction rate, which is recorded by Netperf.
* Define Transaction rate as: The number of requests and complete
responses for a particular requested size that we are able to do per
second. For example if our request is of 1KB and the response is 5KB
then we define the transaction rate as the number of such complete
transactions that we can accomplish per second.
Experiment Details: Similar to the case of bulk transfers we start
the background FTP flows to introduce the congestion in the network
at time 0. About 20 seconds later we start the transactional
transfers and run each test for three minutes. We record the
transactions per second that are complete. We repeat the test for
about an hour and plot the various transactions per second, averaged
out over the runs. The experiment is repeated for various maximum
drop probabilities, file sizes and various levels of congestion.
Observation and Analysis
There are three key observations:
1) As congestion increases (with fixed drop probability) the relative
advantage for ECN increases (again the absolute advantage does not
increase since more flows are sharing the same bandwidth). For
example, from the results, if we consider the 5KB transactional flow,
as we increase the congestion from medium congestion (5 congesting
flows) to high congestion (10 congesting flows) for a maximum drop
probability of 0.1 the relative gain for ECN increases from 42% to
62%.
2) Maintaining the congestion level while adjusting the maximum drop
probability indicates that the relative advantage for ECN flows
increase. From the case of high congestion for the 5KB flow we
Salim & Ahmed Informational [Page 12]
RFC 2884 ECN in IP Networks July 2000
observe that the number of transactions per second increases from 0.8
to 2.2 which corresponds to an increase in relative gain for ECN of
20% to 140%.
3) As the transactional data size increases, ECN's advantage
diminishes because the probability of recovering from a Fast
Retransmit increases for NON ECN. ECN, therefore, has a huge
advantage as the transactional data size gets smaller as is observed
in the results. This can be explained by looking at TCP recovery
mechanisms. NON ECN in the short flows depends, for recovery, on
congestion signaling via receiving 3 duplicate ACKs, or worse by a
retransmit timer expiration, whereas ECN depends mostly on the TCP-
ECE flag. This is by design in our experimental setup. [3] shows
that most of the TCP loss recovery in fact happens in timeouts for
short flows. The effectiveness of the Fast Retransmit/Recovery
algorithm is limited by the fact that there might not be enough data
in the pipe to elicit 3 duplicate ACKs. TCP RENO needs at least 4
outstanding packets to recover from losses without going into a
timeout. For 5KB (4 packets for MTU of 1500Bytes) a NON ECN flow will
always have to wait for a retransmit timeout if any of its packets
are lost. ( This timeout could only have been avoided if the flow had
used an initial window of four packets, and the first of the four
packets was the packet dropped). We repeated these experiments with
the kernels implementing SACK/FACK and New Reno algorithms. Our
observation was that there was hardly any difference with what we saw
with Reno. For example in the case of SACK-ECN enabling: maintaining
the maximum drop probability to 0.1 and increasing the congestion
level for the 5KB transaction we noticed that the relative gain for
the ECN enabled flows increases from 47-80%. If we maintain the
congestion level for the 5KB transactions and increase the maximum
drop probabilities instead, we notice that SACKs performance
increases from 15%-120%. It is fair to comment that the difference
in the testbeds (different machines, same topology) might have
contributed to the results; however, it is worth noting that the
relative advantage of the SACK-ECN is obvious.
6. Conclusion
ECN enhancements improve on both bulk and transactional TCP traffic.
The improvement is more obvious in short transactional type of flows
(popularly referred to as mice).
* Because less retransmits happen with ECN, it means less traffic on
the network. Although the relative amount of data retransmitted in
our case is small, the effect could be higher when there are more
contributing end systems. The absence of retransmits also implies an
improvement in the goodput. This becomes very important for scenarios
Salim & Ahmed Informational [Page 13]
RFC 2884 ECN in IP Networks July 2000
where bandwidth is expensive such as in low bandwidth links. This
implies also that ECN lends itself well to applications that require
reliability but would prefer to avoid unnecessary retransmissions.
* The fact that ECN avoids timeouts by getting faster notification
(as opposed to traditional packet dropping inference from 3 duplicate
ACKs or, even worse, timeouts) implies less time is spent during
error recovery - this also improves goodput.
* ECN could be used to help in service differentiation where the end
user is able to "probe" for their target rate faster. Assured
forwarding [1] in the diffserv working group at the IETF proposes
using RED with varying drop probabilities as a service
differentiation mechanism. It is possible that multiple packets
within a single window in TCP RENO could be dropped even in the
presence of RED, likely leading into timeouts [23]. ECN end systems
ignore multiple notifications, which help in countering this scenario
resulting in improved goodput. The ECN end system also ends up
probing the network faster (to reach an optimal bandwidth). [23] also
notes that RENO is the most widely deployed TCP implementation today.
It is clear that the advent of policy management schemes introduces
new requirements for transactional type of applications, which
constitute a very short query and a response in the order of a few
packets. ECN provides advantages to transactional traffic as we have
shown in the experiments.
7. Acknowledgements
We would like to thank Alan Chapman, Ioannis Lambadaris, Thomas Kunz,
Biswajit Nandy, Nabil Seddigh, Sally Floyd, and Rupinder Makkar for
their helpful feedback and valuable suggestions.
8. Security Considerations
Security considerations are as discussed in section 9 of RFC 2481.
9. References
[1] Heinanen, J., Finland, T., Baker, F., Weiss, W. and J.
Wroclawski, "Assured Forwarding PHB Group", RFC 2597, June 1999.
[2] B.A. Mat. "An empirical model of HTTP network traffic." In
proceedings INFOCOMM'97.
Salim & Ahmed Informational [Page 14]
RFC 2884 ECN in IP Networks July 2000
[3] Balakrishnan H., Padmanabhan V., Seshan S., Stemn M. and Randy
H. Katz, "TCP Behavior of a busy Internet Server: Analysis and
Improvements", Proceedings of IEEE Infocom, San Francisco, CA,
USA, March '98
http://nms.lcs.mit.edu/~hari/papers/infocom98.ps.gz
[4] Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z. and W.
Weiss, "An Architecture for Differentiated Services", RFC 2475,
December 1998.
[5] W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for
Eliminating Packet Loss in Congested TCP/IP Networks", U.
Michigan CSE-TR-349-97, November 1997.
[6] S. Floyd. "TCP and Explicit Congestion Notification." ACM
Computer Communications Review, 24, October 1994.
[7] Ramakrishnan, K. and S. Floyd, "A Proposal to add Explicit
Congestion Notification (ECN) to IP", RFC 2481, January 1999.
[8] Kevin Fall, Sally Floyd, "Comparisons of Tahoe, RENO and Sack
TCP", Computer Communications Review, V. 26 N. 3, July 1996,
pp. 5-21
[9] S. Floyd and V. Jacobson. "Random Early Detection Gateways for
Congestion Avoidance". IEEE/ACM Transactions on Networking,
3(1), August 1993.
[10] E. Hashem. "Analysis of random drop for gateway congestion
control." Rep. Lcs tr-465, Lav. Fot Comput. Sci., M.I.T., 1989.
[11] V. Jacobson. "Congestion Avoidance and Control." In Proceedings
of SIGCOMM '88, Stanford, CA, August 1988.
[12] Raj Jain, "The art of computer systems performance analysis",
John Wiley and sons QA76.9.E94J32, 1991.
[13] T. V. Lakshman, Arnie Neidhardt, Teunis Ott, "The Drop From
Front Strategy in TCP Over ATM and Its Interworking with Other
Control Features", Infocom 96, MA28.1.
[14] P. Mishra and H. Kanakia. "A hop by hop rate based congestion
control scheme." Proc. SIGCOMM '92, pp. 112-123, August 1992.
[15] Floyd, S. and T. Henderson, "The NewReno Modification to TCP's
Fast Recovery Algorithm", RFC 2582, April 1999.
Salim & Ahmed Informational [Page 15]
RFC 2884 ECN in IP Networks July 2000
[16] The NIST Network Emulation Tool
http://www.antd.nist.gov/itg/nistnet/
[17] The network performance tool
http://www.netperf.org/netperf/NetperfPage.html
[18] ftp://ftp.ee.lbl.gov/ECN/ECN-package.tgz
[19] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, S.,
Estrin, D., Floyd, S., Jacobson, V., Minshall, G., Partridge,
C., Peterson, L., Ramakrishnan, K., Shenker, S., Wroclawski, J.
and L. Zhang, "Recommendations on Queue Management and
Congestion Avoidance in the Internet", RFC 2309, April 1998.
[20] K. K. Ramakrishnan and R. Jain. "A Binary feedback scheme for
congestion avoidance in computer networks." ACM Trans. Comput.
Syst.,8(2):158-181, 1990.
[21] Mathis, M., Mahdavi, J., Floyd, S. and A. Romanow, "TCP
Selective Acknowledgement Options", RFC 2018, October 1996.
[22] S. Floyd and V. Jacobson, "Link sharing and Resource Management
Models for packet Networks", IEEE/ACM Transactions on
Networking, Vol. 3 No.4, August 1995.
[23] Prasad Bagal, Shivkumar Kalyanaraman, Bob Packer, "Comparative
study of RED, ECN and TCP Rate Control".
http://www.packeteer.com/technology/Pdf/packeteer-final.pdf
[24] tcpdump, the protocol packet capture & dumper program.
ftp://ftp.ee.lbl.gov/tcpdump.tar.Z
[25] TCP dump file analysis tool:
http://jarok.cs.ohiou.edu/software/tcptrace/tcptrace.html
[26] Thompson K., Miller, G.J., Wilder R., "Wide-Area Internet
Traffic Patterns and Characteristics". IEEE Networks Magazine,
November/December 1997.
[27] http://www7.nortel.com:8080/CTL/ecnperf.pdf
Salim & Ahmed Informational [Page 16]
RFC 2884 ECN in IP Networks July 2000
10. Authors' Addresses
Jamal Hadi Salim
Nortel Networks
3500 Carling Ave
Ottawa, ON, K2H 8E9
Canada
EMail: hadi@nortelnetworks.com
Uvaiz Ahmed
Dept. of Systems and Computer Engineering
Carleton University
Ottawa
Canada
EMail: ahmed@sce.carleton.ca
Salim & Ahmed Informational [Page 17]
RFC 2884 ECN in IP Networks July 2000
11. Full Copyright Statement
Copyright (C) The Internet Society (2000). 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.
Salim & Ahmed Informational [Page 18]