Andrew Gallatin wrote:
Nicolas Williams writes:
> On Mon, May 08, 2006 at 11:38:00AM -0400, Andrew Gallatin wrote:
> > In the context of "networking", I would like to make a request for an
> > behavior change which would be of great use to our (Myricom)
> > non-sockets, non-IP high-performance computing cluster interconnect
> > networking software (GM, MX): Please "fix" bug 6208547. It seems like
> > the redesign of the VM system would be a perfect time to do it :)
>
I will leave it up to the VM team to answer wether they plan on fixing
this bug or not. BTW Stan Studzinski lists a work around in the
evaluation does that work for you ?
> Ah, so RNICs exist. Good to know.
Its not really an "rnic" in the rdma buzzword sort of way. We have
our own API tailored to MPI, but which can be used by anybody.
You can think of us as the precursor to IB.
Where is the API listed ?
Check out the Pallas benchmarks at http://www.myri.com/opteron/
I followed the above page to the results where you list Solaris and
Linux results. Solaris does pretty well with TCP but not with UDP. Can
you try running and reporting the results using snv_27 or above as the
yosemite project went into snv_27 and this project has increased udp
performance considerably.
Rao.
Drew
_______________________________________________
networking-discuss mailing list
[email protected]
_______________________________________________
networking-discuss mailing list
[email protected]
