Re: bifurcated results from netperf

Rick Jones (raj@cup.hp.com)
Mon, 12 Apr 1999 11:17:35 -0700

My favorite candidate for performance evil is lost packets. Check the
netstat stats when the test runs poorly and compare with when it runs
well.

rick jones

Greg Tillman wrote:
>
> Ok, I know i'm out of my field here, but i'd appreciate any help. I'm
> using netperf to test the performance between a pair of Intel's
> EtherExpress Pro/100 100Mbs cards, each on a Linux box (RedHat 5.2, kernel
> 2.0.36). I'm just running a basic
>
> netperf -l 30 -H boxb
>
> The results from a series of 30-second tests seem to be bifurcated,
> clustering around either 12 or 80 Mbs, which are obviously way different
> numbers:
>
> 11,82,12,12,83,13,11,14,12,83,83,60,83,12,83,44,83,83,83,14
>
> Ok, i thought, that will smooth out if we up it to 60 seconds. But, as
> they say, "NOT":
>
> 83,13,13,83,83,83,15,12,12,13,83,13,83,13,83,12,82,83,14,83
>
> Even at 30 minutes, it is by no means nice and smooth and repeatable,
> although it's not as bad as the shorter tests. But if anyone can give me a
> hint what's going on here, I'd appreciate it. One box has 2 network cards,
> because I wanted to test it as a router (once I had some baseline numbers
> to compare the router performance to :-( But anyway, I don't see why two
> cards in a box should matter. I've cut down on the number of other
> processes running on each box to the bare minimum. Each box has 32MB of
> memory, which should be plenty to prevent any kind of swapping. I've
> toggled the '-- -D' option, to no effect. Debug output doesn't give me any
> insight.
>
> Thanks for any help.
>
> - greg tillman
> university of maine at farmington

-- 
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to email, or post, but please do not do both...
my email address is raj in the cup.hp.com domain...