Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!uwm.edu!gem.mps.ohio-state.edu!pacific.mps.ohio-state.edu!tut.cis.ohio-state.edu!ucbvax!D1.DARTMOUTH.EDU!c11234
From: c11234@D1.DARTMOUTH.EDU (Stan Dunten)
Newsgroups: comp.protocols.appletalk
Subject: Re: ATP timeouts
Message-ID: <8910231712.AA16158@dartvax.dartmouth.edu>
Date: 23 Oct 89 20:12:00 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 41

Van Jacobson (van@LBL-CSAM.ARPA) did some great work with stream (tcp)
protocol timeouts. The copy I have came from Article 2361 of
comp.protocols.tcp-ip. Basically he suggests measuring the average reply delay
and it's variation then using a timeout of average+2*variation.

When an reply comes in, the elapsed reply time, request->reply, is used in
these equations:

diff = reply time - ave
ave = ave + diff/T1
dev = dev + (abs(diff)-dev)/T2

The timeout for later requests is ave+2*dev. He suggests a T1 of 8 and a
T2 of 4. Also if a packet is lost (timeout goes off) he doubles the
average. I have both modeled and used this. It works great.

But this is for a stream protocol. The lumpiness of atp and the requirement
to provide an 8 packet timeout are real pains. If atp ignored the callers
time out time (what does he know) and computed its own by:

Use the atp request -> first data packet time in the above equations to
compute a timeout.
Restart the timeout each time a good packet comes in.

Your equation 32*(8*20+50) = 7 sec then becomes 32*(1*20+50) = 2.2 sec.

My tracing of an AppleShare file transfer shows 16 short atp interactions
plus those for the actual data. If most file transfers are less than 16
atp interactions (16*8*512 = 65K files) then most users will be doing short
interactions. The Jacobson equations should react to this further reducing
the 2.2 sec. Users scratching their heads will help still more.

If you have software control of only one side of the atp interaction then:
When making requests the preceding can be used.
When responding to requests your software could:
  Only send 1 or 2 data packets even though more were requested. Setting
  the STS bit creates a 1 packet reply for use in the Jacobson equations.
  Early timeouts from the requesting end would be ignored.

Disclaimer: We haven't tried any of the atp ideas. Good luck.
all run spell*** "<*pass1"