Newsgroups: comp.parallel.pvm
From: raj@cup.hp.com (Rick Jones)
Subject: Re: performance hit across subnets
Organization: http://www.cup.hp.com/netperf/NetperfPage.html
Date: 7 Nov 1994 21:23:09 GMT
Message-ID: <39m5rt$apm@hpindda.cup.hp.com>

Going between subnets means going through an additional peice of
equipment - the brdige/router. This can have several implications.

If enough traffic hits a router, it is possible to overflow a queue or
two, leading to packet loss, which will lead to a considerable loss in
performance. (Actually, that should be "can" lead to, if one is using
TCP with "fast retransmit," then bulk throughput is failry robust in
the face of occasional packet loss.) This is especially true if you
work in a request/reply fashion.

The other issue is that some systems will send smaller TCP segments
(or IP datagram fragments) when going between different subnets. For
Ethernet, this can lead to a 3X increase in the number of segments
exchanged to transfer the same quantity of data. For FDDI, the
increase is even larger. Many vendors will offer a mechanism by which
one can say that subnets are "local" and hence should use the full
Interface MTU. The syntax will vary from system - on one system, it is
a command called subnetconfig. On others, it could be a flag to
ifconfig, or a command to manuipulate routing tables. On others, it
could be a kernel tunable.

The increase in the liklihood of a collision is proably not a very big
deal - collisions aren't all that "evil" and they do not mean that a
packet is lost (unless you are talking about "late collisions" or have
more than 15 of them in a row for the same frame)

rick jones


