[Linux-aus] Request for Qualified Opinion: netperf, iperf, others

Arjen Lentz arjen at lentz.com.au
Tue Jul 25 14:52:09 AEST 2017


Hi Steve

thanks for your reply and input!


> I can't comment on netperf, as it didn't tick the boxes I needed ticked
> at the time.
> 
> Iperf3 is a project of ESNet (Energy Science Network, think AARNet but
> for a specific subnet of the R&E Sector). They have a very different use
> case to most people, as they need to confirm their 100Gbps (and above)
> links and interconnects are actually capable of doing 100Gbps (and
> above). Lots of assumptions occur in that space, like all your links
> have jumbo frame support enabled, you're looking at performance stats
> for long-running (2-3 day) high bandwidth (90Gbps) transfers.
> 
> If you run tests with iperf3 with a card that doesn't have any TCP
> offload capabilities, you're going to get poor results. Iperf3 drops
> 128KB blocks into the send buffer and then walks away and lets the stack
> offload stuff do it's thing. Without any form of a stack offload, you'll
> likely get I/O bound at 4-5Gbps. With a good offload engine, you can
> push 9.5Gbps+ out of a 10Gbps card without breaking a sweat.

Very interesting, thanks.

> If you're looking to see if servers with TOE-support are doing their
> thing, iperf3 will probably tick that box. If you're looking to see if
> the 1Gbps link you've bough actually does 1Gbps, then iperf2 will tick
> the box happily.

Yea so many of us in the real world have to deal with such low bandwidth all the time.
I know you're spoilt at work ;-)

And this is indeed the issue.
Both tools and considerations will differ for different speeds of network.
Also tuning of course, what's good tuning advice a machine on a highspeed network can easily cause low performance on a lower speed network.


> It sounds like what you're wanting to end up with is a Ethernet OAM [0]
> style testing framework, would that be a safe assumption?
> [0]https://en.wikipedia.org/wiki/Operations,_administration_and_management

I think so, yes - mostly.
And I care about links that aren't managed by me, because I need to rely on them.

You neatly described the specific ESnet needs. Similarly, perfSONAR is a wonderful project that enables research organisations globally to keep an eye on their networks.
But neither deal with the reality that most of us poor sods operate in (even professionally), which is hosted machines and users who operate from mobile, ADSL or various types of NBN connections at widely different bandwidth rates. In addition, serious testing can require moving lots of data around, which again can be a problem with many connection plans which have data limits and/or costs.
It's an interesting problem space.

thanks

Regards,
Arjen.


More information about the linux-aus mailing list