Network Working Group                                         T. Shepard
Request for Comments: 2416                                  C. Partridge
Category: Informational                                 BBN Technologies
                                                          September 1998


      When TCP Starts Up With Four Packets Into Only Three Buffers

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 (1998).  All Rights Reserved.

Abstract

   This memo is to document a simple experiment.  The experiment showed
   that in the case of a TCP receiver behind a 9600 bps modem link at
   the edge of a fast Internet where there are only 3 buffers before the
   modem (and the fourth packet of a four-packet start will surely be
   dropped), no significant degradation in performance is experienced by
   a TCP sending with a four-packet start when compared with a normal
   slow start (which starts with just one packet).

Background

   Sally Floyd has proposed that TCPs start their initial slow start by
   sending as many as four packets (instead of the usual one packet) as
   a means of getting TCP up-to-speed faster.  (Slow starts instigated
   due to timeouts would still start with just one packet.)  Starting
   with more than one packet might reduce the start-up latency over
   long-fat pipes by two round-trip times.  This proposal is documented
   further in [1], [2], and in [3] and we assume the reader is familiar
   with the details of this proposal.

   On the end2end-interest mailing list, concern was raised that in the
   (allegedly common) case where a slow modem is served by a router
   which only allocates three buffers per modem (one buffer being
   transmitted while two packets are waiting), that starting with four
   packets would not be good because the fourth packet is sure to be
   dropped.

   Vern Paxson replied with the comment (among other things) that the
   four-packet start is no worse than what happens after two round trip
   times in normal slow start, hence no new problem is introduced by
   starting with as many as four packets.  If there is a problem with a
   four-packet start, then the problem already exists in a normal slow-
   start startup after two round trip times when the slow-start
   algorithm will release into the net four closely spaced packets.

   The experiment reported here confirmed Vern Paxson's reasoning.

Scenario and experimental setup


+--------+  100 Mbps  +---+  1.5 Mbps   +---+  9600 bps    +----------+
| source +------------+ R +-------------+ R +--------------+ receiver |
+--------+  no delay  +---+ 25 ms delay +---+ 150 ms delay +----------+

              |                             |
              |                             |
          (we spy here)              (this router has only 3 buffers
                                      to hold packets going into the
                                      9600 bps link)

   The scenario studied and simulated consists of three links between
   the source and sink.  The first link is a 100 Mbps link with no
   delay.  It connects the sender to a router.  (It was included to have
   a means of logging the returning ACKs at the time they would be seen
   by the sender.)  The second link is a 1.5 Mbps link with a 25 ms
   one-way delay.  (This link was included to roughly model traversing
   an un-congested, intra-continental piece of the terrestrial
   Internet.) The third link is a 9600 bps link with a 150 ms one-way
   delay.  It connects the edge of the net to a receiver which is behind
   the 9600 bps link.

   The queue limits for the queues at each end of the first two links
   were set to 100 (a value sufficiently large that this limit was never
   a factor).  The queue limits at each end of the 9600 bps link were
   set to 3 packets (which can hold at most two packets while one is
   being sent).

   Version 1.2a2 of the the NS simulator (available from LBL) was used
   to simulate both one-packet and four-packet starts for each of the
   available TCP algorithms (tahoe, reno, sack, fack) and the conclusion
   reported here is independent of which TCP algorithm is used (in
   general, we believe).  In this memo, the "tahoe" module will be used
   to illustrate what happens.  In the 4-packet start cases, the
   "window-init" variable was set to 4, and the TCP implementations were
   modified to use the value of the window-init variable only on

   connection start, but to set cwnd to 1 on other instances of a slow-
   start. (The tcp.cc module as shipped with ns-1.2a2 would use the
   window-init value in all cases.)

   The packets in simulation are 1024 bytes long for purposes of
   determining the time it takes to transmit them through the links.
   (The TCP modules included with the LBL NS simulator do not simulate
   the TCP sequence number mechanisms.  They use just packet numbers.)

   Observations are made of all packets and acknowledgements crossing
   the 100 Mbps no-delay link, near the sender.  (All descriptions below
   are from this point of view.)

What happens with normal slow start

   At time 0.0 packet number 1 is sent.

   At time 1.222 an ack is received covering packet number 1, and
   packets 2 and 3 are sent.

   At time 2.444 an ack is received covering packet number 2, and
   packets 4 and 5 are sent.

   At time 3.278 an ack is received covering packet number 3, and
   packets 6 and 7 are sent.

   At time 4.111 an ack is received covering packet number 4, and
   packets 8 and 9 are sent.

   At time 4.944 an ack is received covering packet number 5, and
   packets 10 and 11 are sent.

   At time 5.778 an ack is received covering packet number 6, and
   packets 12 and 13 are sent.

   At time 6.111 a duplicate ack is recieved (covering packet number 6).

   At time 7.444 another duplicate ack is received (covering packet
   number 6).

   At time 8.278 a third duplicate ack is received (covering packet
   number 6) and packet number 7 is retransmitted.

   (And the trace continues...)

What happens with a four-packet start

   At time 0.0, packets 1, 2, 3, and 4 are sent.

   At time 1.222 an ack is received covering packet number 1, and
   packets 5 and 6 are sent.

   At time 2.055 an ack is received covering packet number 2, and
   packets 7 and 8 are sent.

   At time 2.889 an ack is received covering packet number 3, and
   packets 9 and 10 are sent.

   At time 3.722 a duplicate ack is received (covering packet number 3).

   At time 4.555 another duplicate ack is received (covering packet
   number 3).

   At time 5.389 a third duplicate ack is received (covering packet
   number 3) and packet number 4 is retransmitted.

   (And the trace continues...)

Discussion

   At the point left off in the two traces above, the two different
   systems are in almost identical states.  The two traces from that
   point on are almost the same, modulo a shift in time of (8.278 -
   5.389) = 2.889 seconds and a shift of three packets.  If the normal
   TCP (with the one-packet start) will deliver packet N at time T, then
   the TCP with the four-packet start will deliver packet N - 3 at time
   T - 2.889 (seconds).

   Note that the time to send three 1024-byte TCP segments through a
   9600 bps modem is 2.66 seconds.  So at what time does the four-
   packet-start TCP deliver packet N?  At time T - 2.889 + 2.66 = T -
   0.229 in most cases, and in some cases earlier, in some cases later,
   because different packets (by number) experience loss in the two
   traces.

   Thus the four-packet-start TCP is in some sense 0.229 seconds (or
   about one fifth of a packet) ahead of where the one-packet-start TCP
   would be.  (This is due to the extra time the modem sits idle while
   waiting for the dally timer to go off in the receiver in the case of
   the one-packet-start TCP.)

   The states of the two systems are not exactly identical.  They differ
   slightly in the round-trip-time estimators because the behavior at
   the start is not identical. (The observed round trip times may differ
   by a small amount due to dally timers and due to that the one-packet
   start experiences more round trip times before the first loss.)  In
   the cases where a retransmit timer did later go off, the additional

   difference in timing was much smaller than the 0.229 second
   difference discribed above.

Conclusion

   In this particular case, the four-packet start is not harmful.

Non-conclusions, opinions, and future work

   A four-packet start would be very helpful in situations where a
   long-delay link is involved (as it would reduce transfer times for
   moderately-sized transfers by as much as two round-trip times).  But
   it remains (in the authors' opinions at this time) an open question
   whether or not the four-packet start would be safe for the network.

   It would be nice to see if this result could be duplicated with real
   TCPs, real modems, and real three-buffer limits.

Security Considerations

   This document discusses a simulation study of the effects of a
   proposed change to TCP.  Consequently, there are no security
   considerations directly related to the document.  There are also no
   known security considerations associated with the proposed change.

References

   1.   S. Floyd, Increasing TCP's Initial Window (January 29, 1997).
        URL ftp://ftp.ee.lbl.gov/papers/draft.jan29.

   2.   S. Floyd and M. Allman, Increasing TCP's Initial Window (July,
        1997). URL http://gigahertz.lerc.nasa.gov/~mallman/share/draft-
        ss.txt

   3.   Allman, M., Floyd, S., and C. Partridge, "Increasing TCP's
        Initial Window", RFC 2414, September 1998.

Authors' Addresses

   Tim Shepard
   BBN Technologies
   10 Moulton Street
   Cambridge, MA 02138

   EMail: shep@alum.mit.edu


   Craig Partridge
   BBN Technologies
   10 Moulton Street
   Cambridge, MA 02138

   EMail: craig@bbn.com

Full Copyright Statement

   Copyright (C) The Internet Society (1998).  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.