Limited Slow-Start for TCP with Large Congestion Windows
Easing into speed for high-capacity TCP connections
Floyd, S., "Limited Slow-Start for TCP with Large Congestion Windows", RFC 3742, DOI 10.17487/RFC3742, March 2004, <https://www.rfc-editor.org/info/rfc3742>.
@misc{rfc3742,
series = {Request for Comments},
number = 3742,
howpublished = {RFC 3742},
publisher = {RFC Editor},
doi = {10.17487/RFC3742},
url = {https://www.rfc-editor.org/info/rfc3742},
author = {S. Floyd},
title = {{Limited Slow-Start for TCP with Large Congestion Windows}},
pagetotal = 7,
year = 2004,
month = mar,
} RFC text
Contents
Network Working Group S. Floyd
Request for Comments: 3742 ICSI
Category: Experimental March 2004
Limited Slow-Start for TCP with Large Congestion Windows
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 (2004). All Rights Reserved.
Abstract
This document describes an optional modification for TCP's slow-start
for use with TCP connections with large congestion windows. For TCP
connections that are able to use congestion windows of thousands (or
tens of thousands) of MSS-sized segments (for MSS the sender's
MAXIMUM SEGMENT SIZE), the current slow-start procedure can result in
increasing the congestion window by thousands of segments in a single
round-trip time. Such an increase can easily result in thousands of
packets being dropped in one round-trip time. This is often
counter-productive for the TCP flow itself, and is also hard on the
rest of the traffic sharing the congested link. This note describes
Limited Slow-Start as an optional mechanism for limiting the number
of segments by which the congestion window is increased for one
window of data during slow-start, in order to improve performance for
TCP connections with large congestion windows.
1. Introduction
This note describes an optional modification for TCP's slow-start for
use with TCP connections with large congestion windows. For TCP
connections that are able to use congestion windows of thousands (or
tens of thousands) of MSS-sized segments (for MSS the sender's
MAXIMUM SEGMENT SIZE), the current slow-start procedure can result in
increasing the congestion window by thousands of segments in a single
round-trip time. Such an increase can easily result in thousands of
packets being dropped in one round-trip time. This is often
counter-productive for the TCP flow itself, and is also hard on the
rest of the traffic sharing the congested link. This note describes
Limited Slow-Start, limiting the number of segments by which the
congestion window is increased for one window of data during slow-
start, in order to improve performance for TCP connections with large
congestion windows.
When slow-start results in a large increase in the congestion window
in one round-trip time, a large number of packets might be dropped in
the network (even with carefully-tuned active queue management
mechanisms in the routers). This drop of a large number of packets
in the network can result in unnecessary retransmit timeouts for the
TCP connection. The TCP connection could end up in the congestion
avoidance phase with a very small congestion window, and could take a
large number of round-trip times to recover its old congestion
window. This poor performance is illustrated in [F02].
2. The Proposal for Limited Slow-Start
Limited Slow-Start introduces a parameter, "max_ssthresh", and
modifies the slow-start mechanism for values of the congestion window
where "cwnd" is greater than "max_ssthresh". That is, during Slow-
Start, when
cwnd <= max_ssthresh,
cwnd is increased by one MSS (MAXIMUM SEGMENT SIZE) for every
arriving ACK (acknowledgement) during slow-start, as is always the
case. During Limited Slow-Start, when
max_ssthresh < cwnd <= ssthresh,
the invariant is maintained so that the congestion window is
increased during slow-start by at most max_ssthresh/2 MSS per round-
trip time. This is done as follows:
For each arriving ACK in slow-start:
If (cwnd <= max_ssthresh)
cwnd += MSS;
else
K = int(cwnd/(0.5 max_ssthresh));
cwnd += int(MSS/K);
Thus, during Limited Slow-Start the window is increased by 1/K MSS
for each arriving ACK, for K = int(cwnd/(0.5 max_ssthresh)), instead
of by 1 MSS as in standard slow-start [RFC2581].
When
ssthresh < cwnd,
slow-start is exited, and the sender is in the Congestion Avoidance
phase.
Our recommendation would be for max_ssthresh to be set to 100 MSS.
(This is illustrated in the NS [NS] simulator, for snapshots after
May 1, 2002, in the tests "./test-all-tcpHighspeed tcp1A" and
"./test-all-tcpHighspeed tcpHighspeed1" in the subdirectory
"tcl/lib". Setting max_ssthresh to Infinity causes the TCP
connection in NS not to use Limited Slow-Start.)
With Limited Slow-Start, when the congestion window is greater than
max_ssthresh, the window is increased by at most 1/2 MSS for each
arriving ACK; when the congestion window is greater than 1.5
max_ssthresh, the window is increased by at most 1/3 MSS for each
arriving ACK, and so on.
With Limited Slow-Start it takes:
log(max_ssthresh)
round-trip times to reach a congestion window of max_ssthresh, and it
takes:
log(max_ssthresh) + (cwnd - max_ssthresh)/(max_ssthresh/2)
round-trip times to reach a congestion window of cwnd, for a
congestion window greater than max_ssthresh.
Thus, with Limited Slow-Start with max_ssthresh set to 100 MSS, it
would take 836 round-trip times to reach a congestion window of
83,000 packets, compared to 16 round-trip times without Limited
Slow-Start (assuming no packet drops). With Limited Slow-Start, the
largest transient queue during slow-start would be 100 packets;
without Limited Slow-Start, the transient queue during Slow-Start
would reach more than 32,000 packets.
By limiting the maximum increase in the congestion window in a
round-trip time, Limited Slow-Start can reduce the number of drops
during slow-start, and improve the performance of TCP connections
with large congestion windows.
3. Experimental Results
Tom Dunigan has added Limited Slow-Start to the Linux 2.4.16 Web100
kernel, and performed experiments comparing TCP with and without
Limited Slow-Start [D02]. Results so far show improved performance
for TCPs using Limited Slow-Start. There are also several
experiments comparing different values for max_ssthresh.
4. Related Proposals
There has been considerable research on mechanisms for the TCP sender
to learn about the limitations of the available bandwidth, and to
exit slow-start before receiving a congestion indication from the
network [VEGAS,H96]. Other proposals set TCP's slow-start parameter
ssthresh based on information from previous TCP connections to the
same destination [WS95,G00]. This document proposes a simple
limitation on slow-start that can be effective in some cases even in
the absence of such mechanisms. The max_ssthresh parameter does not
replace ssthresh, but is an additional parameter. Thus, Limited
Slow-Start could be used in addition to mechanisms for setting
ssthresh.
Rate-based pacing has also been proposed to improve the performance
of TCP during slow-start [VH97,AD98,KCRP99,ASA00]. We believe that
rate-based pacing could be of significant benefit, and could be used
in addition to the Limited Slow-Start in this proposal.
Appropriate Byte Counting [RFC3465] proposes that TCP increase its
congestion window as a function of the number of bytes acknowledged,
rather than as a function of the number of ACKs received.
Appropriate Byte Counting is largely orthogonal to this proposal for
Limited Slow-Start.
Limited Slow-Start is also orthogonal to other proposals to change
mechanisms for exiting slow-start. For example, FACK TCP includes an
overdamping mechanism to decrease the congestion window somewhat more
aggressively when a loss occurs during slow-start [MM96]. It is also
true that larger values for the MSS would reduce the size of the
congestion window in units of MSS needed to fill a given pipe, and
therefore would reduce the size of the transient queue in units of
MSS.
5. Acknowledgements
This proposal is part of a larger proposal for HighSpeed TCP for TCP
connections with large congestion windows, and resulted from
simulations done by Evandro de Souza, in joint work with Deb Agarwal.
This proposal for Limited Slow-Start draws in part from discussions
with Tom Kelly, who has used a similar modified slow-start in his own
research with congestion control for high-bandwidth connections. We
also thank Tom Dunigan for his experiments with Limited Slow-Start.
We thank Andrei Gurtov, Reiner Ludwig, members of the End-to-End
Research Group, and members of the Transport Area Working Group, for
feedback on this document.
6. Security Considerations
This proposal makes no changes to the underlying security of TCP.
7. References
7.1. Normative References
[RFC2581] Allman, M., Paxson, V. and W. Stevens, "TCP Congestion
Control", RFC 2581, April 1999.
[RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte
Counting (ABC)", RFC 3465, February 2003.
7.2. Informative References
[AD98] Mohit Aron and Peter Druschel, "TCP: Improving Start-up
Dynamics by Adaptive Timers and Congestion Control"",
TR98-318, Rice University, 1998. URL "http://cs-
tr.cs.rice.edu/Dienst/UI/2.0/Describe/ncstrl.rice_cs/TR98-
318/".
[ASA00] A. Aggarwal, S. Savage, and T. Anderson, "Understanding the
Performance of TCP Pacing", Proceedings of the 2000 IEEE
Infocom Conference, Tel-Aviv, Israel, March, 2000. URL
"http://www.cs.ucsd.edu/~savage/".
[D02] T. Dunigan, "Floyd's TCP slow-start and AIMD mods", 2002.
URL "http://www.csm.ornl.gov/~dunigan/net100/floyd.html".
[F02] S. Floyd, "Performance Problems with TCP's Slow-Start",
2002. URL "http://www.icir.org/floyd/hstcp/slowstart/".
[G00] A. Gurtov, "TCP Performance in the Presence of Congestion
and Corruption Losses", Master's Thesis, University of
Helsinki, Department of Computer Science, Helsinki,
December 2000. URL
"http://www.cs.helsinki.fi/u/gurtov/papers/ms_thesis.html".
[H96] J. C. Hoe, "Improving the Start-up Behavior of a Congestion
Control Scheme for TCP", SIGCOMM 96, 1996. URL
"http://www.acm.org/sigcomm/sigcomm96/program.html".
[KCRP99] J. Kulik, R. Coulter, D. Rockwell, and C. Partridge, "A
Simulation Study of Paced TCP", BBN Technical Memorandum
No. 1218, 1999. URL
"http://www.ir.bbn.com/documents/techmemos/index.html".
[MM96] M. Mathis and J. Mahdavi, "Forward Acknowledgment: Refining
TCP Congestion Control", SIGCOMM, August 1996.
[NS] The Network Simulator (NS). URL
"http://www.isi.edu/nsnam/ns/".
[VEGAS] Vegas Web Page, University of Arizona. URL
"http://www.cs.arizona.edu/protocols/".
[VH97] Vikram Visweswaraiah and John Heidemann, "Rate Based Pacing
for TCP", 1997. URL
"http://www.isi.edu/lsam/publications/rate_based_pacing/".
[WS95] G. Wright and W. Stevens, "TCP/IP Illustrated", Volume 2,
Addison-Wesley Publishing Company, 1995.
Authors' Address
Sally Floyd
ICIR (ICSI Center for Internet Research)
Phone: +1 (510) 666-2989
EMail: floyd@icir.org
URL: http://www.icir.org/floyd/
Full Copyright Statement
Copyright (C) The Internet Society (2004). This document is subject
to the rights, licenses and restrictions contained in BCP 78 and
except as set forth therein, the authors retain all their rights.
This document and the information contained herein are provided on an
"AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE
REPRESENTS OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE
INTERNET ENGINEERING TASK FORCE DISCLAIM 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.
Intellectual Property
The IETF takes no position regarding the validity or scope of any
Intellectual Property Rights or other rights that might be claimed
to pertain to the implementation or use of the technology
described in this document or the extent to which any license
under such rights might or might not be available; nor does it
represent that it has made any independent effort to identify any
such rights. Information on the procedures with respect to
rights in RFC documents can be found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use
of such proprietary rights by implementers or users of this
specification can be obtained from the IETF on-line IPR repository
at http://www.ietf.org/ipr.
The IETF invites any interested party to bring to its attention
any copyrights, patents or patent applications, or other
proprietary rights that may cover technology that may be required
to implement this standard. Please address the information to the
IETF at ietf-ipr@ietf.org.
Acknowledgement
Funding for the RFC Editor function is currently provided by the
Internet Society.
AI ELI5
On this page
§1The Slow-Start Problem
TCP slow-start is designed to probe the network's available capacity without overwhelming it. In the standard algorithm, the sender increases the congestion window by one Maximum Segment Size (MSS) for every acknowledgement received, effectively doubling the window each round-trip time. This aggressive growth works well for typical connections, but for high-speed long-distance links or modern data centers, the congestion window can grow to thousands of segments.
When the window reaches such sizes, doubling it in one round-trip time means injecting thousands of new packets into the network. Even well-tuned router queues cannot absorb such a burst without dropping many packets. The result is often counterproductive: instead of reaching a high sending rate, the sender experiences massive loss, triggering retransmission timeouts, and ends up at a very small congestion window in congestion avoidance mode. Recovery takes many round-trip times.
If you double your speed every second, you go from walking to crashing into a wall very quickly. That is not smart for getting there faster.
§2The Limited Slow-Start Mechanism
Limited Slow-Start introduces a parameter max_ssthresh, recommended to be set to 100 MSS. The algorithm proceeds in two phases: when the congestion window is at or below max_ssthresh, the sender uses standard slow-start, increasing the window by one MSS per ACK. When the window exceeds max_ssthresh, the sender enters Limited Slow-Start mode.
In Limited Slow-Start mode, the window growth is capped so that the maximum increase per round-trip time is max_ssthresh / 2 segments. For each arriving ACK, the sender increments the window by int(MSS / K), where K = int(cwnd / (0.5 * max_ssthresh)). This formula ensures that as the window grows larger, the increments per ACK shrink inversely, keeping the total growth per RTT bounded.
Instead of doubling speed every time, shift to adding a tiny bit each time. The bigger you get, the tinier the increments.
§3Growth Rates and Performance
With Limited Slow-Start set to max_ssthresh = 100 MSS, it takes approximately log(100) ≈ 7 round-trip times to reach 100 MSS, just like standard slow-start. Beyond that, reaching a window of 83,000 packets requires about 836 round-trip times instead of just 16 without Limited Slow-Start (assuming no losses).
The key advantage is queue management: the transient queue during standard slow-start reaching 83,000 packets would exceed 32,000 packets. With Limited Slow-Start, the largest transient queue stays around 100 packets. This massive reduction in buffering prevents the large packet drops that trigger timeouts. The result is fewer retransmissions and overall better performance even though it takes longer to reach full speed.
Getting to high speed takes longer, but you do not crash into a wall of dropped packets halfway there.
§4Deployment and Results
Implementation in Linux 2.4.16 Web100 kernel demonstrated improved performance for TCP connections using Limited Slow-Start compared to standard slow-start. Experiments showed that the algorithm works well with different values of max_ssthresh, though the recommended value of 100 MSS provides a good balance.
Limited Slow-Start is orthogonal to other TCP improvements and can be deployed alongside rate-based pacing, Appropriate Byte Counting, and other congestion control enhancements. The mechanism introduces no new options or packet overhead; it is a purely local sender-side change to the slow-start algorithm.
Tests showed the gentler approach actually works better in practice, even though it is not the fastest possible way to get to high speed.