Re: Mail Bridge Performance

Edward A. Cain (cain@EDN-UNIX.ARPA)
3 Mar 1986 17:03:11 EST


Thanks for the summary of mailbridge traffic. I think it does partially
explain why performance is so awful at times thru the mailbridges. The
correlation with school schedules is interesting, too, and probably a better
guess than any I've heard recently.

There is one other important consideration. Performance on the ARPANET alone
has been terrible at times. For example, ICMP ECHO and ECHO REPLY round-trip
measurements between east and west coast hosts were averaging 18 seconds on
Feb 3-4, with tails of the delay distribution out to 37 seconds, as measured
from DCEC (via arpanet) and at BRL (via milnet). Delays were very high
again during the Feb 12-14 time period. Even worse, on Feb 20th, one hour
in the afternoon the roundtrip delay from DCEC to the arpanet interface of
the ISI mailbridge was 30-40 seconds, and from DCEC to the arpanet
interface of the SRI mailbridge the delays were 45-47 seconds during the
same hour, with 90% packet loss!!!

Usually, this kind of behavior on the arpanet is coincident with the outage
of key lines or nodes in the arpanet. On Feb 20th for example, line 76
(utah to lbl2) and line 76 (sri2 to collins) were both down most of the
day because of flooded cableheads.!!! The loss of a key component in the
arpanet seems to create serious congestion when the traffic goes up. And
congestion is noticed quickly by the mailbridges, which are among the
busiest arpanet hosts in terms of both packets sent and connection blocks
used (in the IMP).

Some of the overhead traffic you mentioned, although still alarmingly
high, has decreased noticeably since a year ago. The decrease in Packets
Originating at a Gateway could be due mainly to hosts learning how to
handle ICMP Redirects.

I don't suspect replacing the mailbridges with Butterflies is going to
make any noticeable difference. The new congestion control scheme for the IMP
might help, if anyone does anything with Source Quench, because it paves
the way for gateways to learn about congestion in the networks (currently,
RFNM blocking is the only trick). Unless some action is taken to provide
a congestion control strategy at both the network and internet levels, or
atlernatively, enough spare capacity is provided in the arpanet to avoid
most of the congested situations, I don't think there will be any improvement
in mailbridge performance.

Ed Cain

This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:36:04 GMT