SUMMARY: collision rate (what's yours?)

From: Curt Welch - RDR (curt@to.mobil.com)
Date: Wed Aug 03 1994 - 08:28:13 CDT


Some time ago, I asked the net:

> What type of numbers do all of you see if you divide Collis by Opkts in
> the output of netstat -i for your Sun Sparc systems (le0 line only)? I
> see numbers ranging from around 2% (i.e. .02) to as high as 50% if we
> really load the network. It normally hangs around 5%. Is this normal
> for Sun systems???

> My network is a 10Base-T network with about 25 SPARC 10s doing "normal" NFS
> and X types of things. The avg load on the net is around 5%.

Here are the 6 replies I received:

---
| I'm running a 10base-T network with 17 Sparc 10's, my ratio for Opkts to 
| collisions is running a little over 3% (.03).  We're also doing the NFS and
| X type stuff.  Sounds like your network is pretty much "normal" to me,
| although "normal" can be a dangerous word!!
---
| We run about 40 clients in a production trading floor with market feeds and
| lots of big sybase queries. It hovers around 12-15%.
---
| I generally see values of 5% to 15%
---
| Here's a sampling from my network:
|
|                   Opkts  Oerrs Collis
| butthead:  2.3% (  628772   79  14493)  Sparc 10-512, Solaris 2.3
| beavis:   11.1% (  628012 1112  69756)  Sparc 10-512, Solaris 2.3
| hobbes:    1.9% ( 1328255    2  24791)  Sparc ELC, Solaris 2.3
| dagger:         (   51648    0    609)  Sparc ELC, Solaris 2.3
             1.2%
| calvin:    1.4% (18116248   26 250818)  Sparc IPC, SunOS 4.1.1
| opsnms:      0% (  234972    0      0)  IBM RS/6000, AIX 3.2
| beagle:      0% ( 3433290    0      0)  IBM RS/6000, AIX 3.2
| vdbsan:     .8% (  673614    1   5818)  IBM 486 clone, ISC 3.2.2
| komondor:   .3% ( 1380415    1   4232)  HP 9000/700, HP-UX A.09.03
---
| On our heavily loaded NEXTSTEP server we have these stats:
| (coll/okpts)*100=5.6%, so we are about the same as you....
---
| We have an unsegmented thernet with 10 SLC, 3 IPX, 15 PC, 15 X-terminal
| and a 670 SPARCServer.
| The collision rates (Collis/Opkts) vary from 0.01 to 0.25.
| The average (calulated as sum Collis/ sum Opkts) is usually
| around 0.07.
---

So the answer to my question seems to be yes. Collision rates from 5% to 10% are not unusual. Though it was interesting to note that the one person that responded with a mixed set of systems had higher collision rates on his two Sparc 10s than on any of his other systems.

No one reported rates as high as I have seen (50%), but I only saw numbers that high when I was testing the Sun Show-me video converencing system. It was pumping multiple streams of real-time digital video onto the net and using up something like 80% of the bandwith of the ethernet. I wasn't surprised to see the collision rate go through the roof in that case.

Thanks again to everyone that responded. Sorry for taking so long to post the summary...

Curt Welch curt@to.mobil.com



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:09:06 CDT