[openib-general] performance tests uploaded to contrib

[Sayantan Sur]

* On Apr,5 Michael S. Tsirkin<mst at mellanox.co.il> wrote :
> Quoting r. Sayantan Sur <surs at cse.ohio-state.edu>:
> > Subject: Re: [openib-general] performance tests uploaded to contrib
> > 
> > Michael S. Tsirkin wrote:
> > >Quoting r. Sayantan Sur <surs at cse.ohio-state.edu>:
> > >
> > >>Subject: Re: [openib-general] performance tests uploaded to contrib
> > >>
> > >>Michael,
> > >>
> > >>
> > >>>Current results:
> > >>>
> > >>>I currently observe latency below 3.5 usec.
> > >>>
> > >>>Drop me a note if you find this useful.
> > >>
> > >>Thanks for putting this up on the contrib tree. I have run this
> > >>rdma_latency, and am getting around 3.35 us (without switch).
> > >>
> > >>Our platform is Dual Xeon EM64T with RH AS 4. Kernel version 2.6.11.7
> > >>
> > >>Do you have any idea when a port of the popular `perf_main' will be
> > >>available? As more people try to use the Gen2 verbs, `perf_main' (or
> > >>something similar) can help people evaluate different IB operations and
> > >>also to have example code to use different features of IB.
> > >>
> > >>Thanks,
> > >>Sayantan.
> > >>
> > >
> > >
> > >I dont plan to port the monolithic perf_main to gen2.
> > >
> > >Instead, I plan to upload a set of microbenchmarks each testing
> > >a specific feature: rdma latency/rc send latency/rdma bandwidth/rc send
> > >bandwidth etc.
> > >
> > >I hope that this will help achieve better code clarity than what we 
> > >have in perf_main.
> > 
> > That is fine. As long as there is some ibverbs level benchmark suite, it 
> > will ease the transition.
> > 
> > >
> > >What are the features you are most interested in?
> > 
> > By RDMA latency/bandwidth, do you mean both RDMA write & read?
> 
> RDMA write test is out there, I hope to upload the read test RSN.
> 
> > Will 
> > there be any Atomic latency tests also?
> 
> Sure, why not. Is it a priority for you?

Nope. Not a priority RDMA Write/Read will do for some time to come.

Thanks,
Sayantan.

> 
> > >
> > >By the way, you can already see some example code in libibverbs/examples,
> > >although that is not necessarily benchmark-oriented.
> > >I used the code in pingpong.c as a starting point for rdma latency
> > >test, and it can be used with very little changes as rc send latency
> > >test.
> > 
> > Yes, this was helpful for me too. Apart from providing code examples, I 
> > was suggesting that a much more comprehensive benchmark suite will help 
> > people moving from other vendor stacks to Gen2 to have a quick 
> > comparison of performance offered by both stacks.
> > 
> > Thanks,
> > Sayantan.
> > 
> 
> I agree.
> 
> -- 
> MST - Michael S. Tsirkin

-- 
---------------------------------------------------------
Sayantan Sur            Graduate Research Assistant

395 Dreese Labs,        Computer Science and Engineering
Ohio State University,  Office : 774, Dreese Labs
Columbus,               email  : surs at cse.ohio-state.edu
Ohio - 43210.           phone(res) : 614.688.9792
USA.                    phone(off) : 614.292.8501
---------------------------------------------------------