[Beowulf] Measuring port to port latency

Douglas Eadline deadline at eadline.org
Thu Sep 20 08:31:57 PDT 2007


The benchmarks mentioned in the other responses are a
good place to start. However, these single port numbers
don't tell the whole story with multi-core where you may
have eight cores on a node all talking to other nodes at
the same time.

Multi-core requires thinking differently about performance.
Talking about single application performance in a multi-core
environment is like believing that all you need is a good
quarterback to have a winning football team (apologies to
all the non-US football fans)

 --
 Doug


> Hi,
>
> since the rush of emails seems to be over right now, I would like to
> start another thread. How are you measuring port to port latency in your
> network easily?
>
> So far, I have found a small MPI-pingpong code at
>
> http://cluster.fisica.unam.mx/curso/mpi_intro/para_pingpong.html
>
> as well as netpipe at
>
> http://www.scl.ameslab.gov/Projects/Netpipe/
>
> They look fine on their own, but since I'm just starting, any opinions?
>
> Cheers
>
> Carsten
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>
> !DSPAM:46f256ae60241691722815!
>


--
Doug



More information about the Beowulf mailing list