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

[Grant Grundler]

On Wed, May 18, 2005 at 10:07:52PM +0300, Michael S. Tsirkin wrote:
...
> > 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.
> 
> Fair enough - the more the merrier.

Rick Jones thinks it should be straight forward to add netperf_ib.c.
I'll work with him to update his silly HPUX box to a more silly
parisc-linux box so we can run subversion. Then I can start "borrowing"
code from rmda_lat to do the same things using netperf infrastructure.

> 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.

Yes, I agree with both the goals and rdma_lat covers those pretty
well already.

I think netperf is more interesting for more complex stuff since
it has slightly different goals and serves a different purpose.
(e.g. binding tasks to CPUs and reporting statistics)
I'll finish up cleaning up rmda_lat and then use it as an example
(like you intended) to add IB functionality into netperf.

> I started with latency and am looking at bandwidth now, it seems these
> may be part of the same test

If the two tests can share code, I'd rather see two example tests
and parts of rdma_lat.c split off into a seperate file.
I expect all the socket handling and context setup stuff could be
shared.

The more bells and whistles you add to rdma_lat, the further
it gets away from your primary goal (simple, efficient example).

>- but CPU utilization may be harder.

Isn't the use of get_cycles() a sufficient form of CPU utilization
for the two goals above?
Or is there more kernel overhead we aren't catching somehow?

grant