Newsgroups: comp.sys.sun.hardware,comp.dcom.lans.ethernet
From: vjs@calcite.rhyolite.com (Vernon Schryver)
Subject: Re: Ethernet chip bugs (slow network performance)
Keywords: Ethernet,bugs,slow
References: <2jc6iv$g99@redwood.cs.scarolina.edu>  <1994Feb17.155706.997@newstand.syr.edu>
Organization: Rhyolite Software
Date: Fri, 18 Feb 1994 15:23:22 GMT
Message-ID: 
Lines: 58
Xref: geraldo.cc.utexas.edu comp.sys.sun.hardware:10469 comp.dcom.lans.ethernet:5853

In article <1994Feb17.155706.997@newstand.syr.edu> ppmorris@rodan.syr.edu (Peter P Morrissey) writes:
>
>u have the decimal point wrong.
>>3% or 4% collisions means the Ethernet is idle.
>>10% is means the Ethernet is lightly loaded.
>>30% or 40% is something to worry about.
> 
>>Do the arithmetic yourself to see that a 4% collision rate will have
>>an insignificant effect on throughput either between any pair of
>>stations or for the entire network.
> 
>I believe you mean to say "3% - 4% *utilization* is idle...30-40% something
>to start worrying about..." 

No, I meant "3-4% collisions means the wire is idle" and
"30-40% collisions are something to worry about."

>I am computing collisions by looking at the total packets, and the total 
>collisions, total collisions/total packets.
>
>If there is a 10% collision rate, that means 1 in 10 of every packet
>transmitted successfully was involved in a collision. This seems like a lot
>to me, especially if you consider that when there is a collision, it is because
>at least two people were attempting to transmit at the same time, so the
>percentage might be even higher than my formula indicates. I have seen
>serious problems on networks with collision rates over 10%.

Let's do the arithmetic.  Assume your network has the usual bimodal traffic
distribution, about half small packets and half large packets.  (Eg. TCP
with 2 1518 byte data packets for every 64 byte ACK (58 increased to the
min.) or NFS with 1500 UDP/IP frags and small requests and getattrs).
Make the simplifying assumption the average packet size is 700 bytes.
Take a 10% collision rate to mean that 10% of all transmissions by any
single station suffer a single collision.  That means 10% of all packets
are delayed by one slot time or 64 bytes or less than 10%, which means
that the effective waste is about 1%.  Even if you double that to
incorrectly account for the fact that 2 stations are involved, you still
have an insignificant effect.  (It is incorrect to double it because the
total time for both stations to transmit their packets will be the time
for the collision plus the time needed for the first packet plus the time
needed for the second packet plus various small fudges.)  Even if you
assume your network traffic consists only of minimum length 64-byte
packets, a 10% collision rate yields only about a 10% performance effect.

Thus, any "serious problems" you have see have had nothing to do with a
collision rate as low as 10%.

>Also, I have never observed a functioning network with collision rates 
>above 20%. Has anyone else?

Many Ethernets in production run with collision rates higher than 20% for
extended periods of time.

Do not believe all of the nonsense of the switching hub and bridge sales
critters.


Vernon Schryver    vjs@rhyolite.com