[openib-general] Re: diff-perftest-07 replace pp_get_local_lid()

[Michael S. Tsirkin]

Quoting r. Grant Grundler <iod00d at hp.com>:
> Subject: Re: diff-perftest-07 replace pp_get_local_lid()
> 
> On Wed, May 18, 2005 at 04:51:30PM +0300, Michael S. Tsirkin wrote:
> > Quoting r. Grant Grundler <iod00d at hp.com>:
> > > Following patch to rdma_lat.c:
> > > o replaces pp_get_local_lid with code from ibv_pingpong.
> > >   This calls into libibverbs instead of fishing around in /sys FS.
> > > 
> > > o makes two minor white space fix-ups.
> > > 
> > > Signed-off-by: Grant Grundler <iod00d at hp.com>
> > 
> > No patch was attached :(
> 
> Doh! sorry...appended now
> 
> > > I'd like to slowly restructure main() into multiple distinct parts:
> > > 	1) parameter parsing/setting
> > > 	2) global data init (e.g. srand())
> > > 	3) setup connection
> > > 	3) negotiate test+parameters with server/client
> > > 	4) run test (maybe several iterations with different params)
> > > 	5) exit/cleanup
> > 
> > Good, but lets pass data around in function parameters
> > instead of in global data section.
> 
> Certainly. There should be very little/no global data
> when I'm done. srand/errno type stuff should be about it.
> 
> > > I'm thinking about how to keep the server running and iterating.
> > > The goal is to be able to run a sequence of tests just
> > > from the client side.
> > 
> > I just do ssh server $PWD/rdma_lat & ( sleep 1 && $PWD/rdma_lat)
> 
> hrm...that would work too though it's not as intuitive.
> 
> > > Or is this a waste of time?
> > > Should I rather be looking at fixing up netperf to support IB?
> > 
> > That may be kind of hard, given that uverbs API is completely different
> > from socket API.
> 
> Agreed - but that's not the part I'm trying to leverage.
> I want:
> o statistical infrastructure.
> o automatically run multiple iterations until confidence level is reached
> o CPU utilization (aka "Service Demand")
> o bind task to a particular CPU
> 
> I can add those things to rmda_lat but it seems like it's a wash
> to add IB support to netperf.
>
>
> And since netperf.org is pretty well
> known and no-one has a clue about rdma_lat..it seems like a more
> obvious place to do the work. I guess that's a political
> argument for doing the work in netperf. I'll think about it more
> and talk to the netperf maintainer later this week.
> 
> thanks,
> grant

Fair enough - the more the merrier.

I'd like to re-iterate what I am trying to do currently:
- show how to utilize uverbs efficiently
- get a benchmark useful e.g. for hardware tuning and/or testing

For both these goals its very important for the test to be simple and readable-
netperf sure is not it.

I started with latency and am looking at bandwidth now, it seems these
may be part of the same test - but CPU utilization may be harder.

-- 
MST - Michael S. Tsirkin