Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!rutgers!sri-spam!mordor!lll-tis!ames!sdcsvax!ucbvax!ACC.ARPA!art
From: art@ACC.ARPA
Newsgroups: comp.protocols.tcp-ip
Subject: TCP SRTT
Message-ID: <8708211856.AA09676@ucbvax.Berkeley.EDU>
Date: Fri, 21-Aug-87 15:09:00 EDT
Article-I.D.: ucbvax.8708211856.AA09676
Posted: Fri Aug 21 15:09:00 1987
Date-Received: Sun, 23-Aug-87 02:43:25 EDT
Sender: daemon@ucbvax.BERKELEY.EDU
Reply-To: <art@acc.arpa>
Distribution: world
Organization: The ARPA Internet
Lines: 19


I implemented the SRTT algoritm posted by Van Jacobsen a while back.
This is the one that accumulates both a smoothed round trip time estimate
(SRTT) and also a smoothed variance estimate (which I call SRTV).  The
retransmit timer is set to: (SRTT + 2*SRTV)*rxmt_backoff_factor.

This has greatly improved the behavior of TCP over low speed links
(4800-9600 baud), while showing no noticable effect over Ethernet.
Using the standard SRTT algorithm, I was seeing as low as 10% link
utilization because of spurious retransmissions.  The new algorithm
is giving over 90% link utilization.

I would recommend others consider using this algoritm if there TCP
ever uses a path with large delay variance (such as low speed lines
or Arpa Internet).
					Art Berggreen
					art@acc.arpa

------