[openib-general] Userspace testing results (2.6.15-rc7-git2 with modules)

[Troy Benjegerdes]

> Currently, I am running netpipe, iperf and netperf (these three tests
> are giving horrible results but we are pretty sure that it is a local
> issue, as both eth1 and ib0 based tests lead to poor performance) and
> also netpipe with a patch from Shirley Ma to run over native IB [1].
> Additionally, I am running the 4 pingpong tests (rc, srq, uc, ud) and
> the two perftest tests: rdma_lat and rdma_bw. There are some issues with
> some size combinations; or, at least, that is how it seems to me.
> 

I assume this is the same patch I have here..
http://source.scl.ameslab.gov/hg/netpipe3-dev

Brad Benton also found that the way NetPIPE does polling to see if a
message arrives does not work well with the relaxed memory consistency
model the Power5 systems with ehca seem to use. His workaround was to
add a 'dcbf' instruction in NetPIPE, but that's obviously not portable.

Shouldn't we have a function in openIB to force immediate cache
consitency checks?