[openib-general] Re: [PATCH] rdma_lat-09 and results

[Grant Grundler]

On Thu, Jun 16, 2005 at 01:11:39AM +0300, Michael S. Tsirkin wrote:
> Quoting r. Grant Grundler <iod00d at hp.com>:
> > 2) I working towards making the server side a daemon.
> >    ie move stuff out of main:
> 
> I have another idea: lets move all the socket stuff out of perftests
> altogether. Each test will just read/write standard input/output.
> Reports will go to standard error.

That sounds good to me.

> On top of this, we'll be able to have a script/program, or
> multiple different programs, that will perform the
> server/client communication in whatever way we want it to.

I've been heading down this path already (splitting out the test)
and have my latest version posted on 
	ftp://gsyprf10.external.hp.com/pub/openib/perftest/rdma_lat.c-10

Sorry - I haven't been able to prepare a patch based on
the current SVN content and probably won't tomorrow either.
The above is the best I can offer right now.

> One simple way would be to dup the socket descriptor
> to standard input/output.

I'm not a good network applications programmer...ie I'd be more
comfortable taking netperf/netserver and hacking it to handle
the "high level" stuff. I'm just not quite sure how to wedge
the test into netperf at this point.

> It seems this would also let the same test use cm for connection setup,
> but I didnt look at ucm, yet.

sorry - neither "cm" nor "ucm" mean anything to me (yet).
Infiniband has more acronyms than a stray dog has fleas. :^)

thanks,
grant