[Beowulf] new release of GAMMA: x86-64 + PRO/1000

Douglas Eadline deadline at clustermonkey.net
Fri Feb 3 08:47:33 PST 2006


I think it still makes sense to develop GAMMA for GigE. Here are some
preliminary numbers I was able to achieve on a very low cost cluster

The Gamma ping-pong test gives latency through the switch as (useconds):

 (lowest < average < highest)
 9.13639 < 9.54739 < 15.5897

it also reports 72 MB/sec throughput (which is pretty good for a 32 bit/33
MHz NIC!) The netpipe TCP was 29 usec latency and 66 MB/sec.

Here are some *preliminary* MPI netpipe numbers as compared to LAM and MPICH.



Before I do more testing I needed to add more memory to the nodes and
update some software (working on this today as a matter of fact) I did run
the NAS suite without no problems.

Thanks for your efforts.


> Hello,
> I'm glad to announce that GAMMA has been ported to x86-64 CPU
> :-)
> This release of GAMMA has been tested on a GigEth cluster of dual-Xeon
> as well as a cluster of single-EM64T (courtesy of OpenCFD, Ltd).  So, we
> (and need!) tests on GigEth cluster of dual-processor x86-64 machines
> have no such hardware on site).
> A new, 64-bit capable, version of MPI/GAMMA has been released as well.
> GAMMA only supports the Intel PRO/1000 Gigabit Ethernet NIC (e1000 driver).
> Latency at MPI level is below 12 usec, switch included (6.5 usec
> No significant difference between 32-bit and 64-bit.
> More details at the GAMMA web page, http://www.disi.unige.it
> Some tests on computational fluid dynamics jobs (courtesy of OpenCFD,
> show a very interesting comparison between MPI/GAMMA and LAM.  I will
> more details on the GAMMA web page soon.
> A question:
> Does it still make sense to have a low-latency communication library for
Gigabit Ethernet?
> Bye,
> Giuseppe Ciaccio               http://www.disi.unige.it/person/CiaccioG/
DISI - Universita' di Genova   via Dodecaneso 35   16146 Genova,   Italy
phone +39 10 353 6637          fax +39 010 3536699 ciaccio at disi.unige.it
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf


More information about the Beowulf mailing list