[openfabrics-ewg] Cisco SQA results so far for OFED 1.0 rc5

[Pavel Shamis (Pasha)]

Hi,
The default mtu size in mvapich-RC5 was changed to MTU1024 (it solves bw 
issue on PCI_X). You can try  VIADEV_DEFAULT_MTU=MTU2048 parameter on 
your platforms (PCI_EX) and I believe that you will get the rc4 latency 
numbers.
Also in RC6 we added mvapich.conf file that include all configuration 
options and description for best tunings on different platforms.

Regards
Pasha

Scott Weitzenkamp (sweitzen) wrote:
> I forgot to mention that Open MPI and Intel MPI small message latency 
> are slightly worse as well in rc5 compared to rc4.
>  
> Scott Weitzenkamp
> SQA and Release Manager
> Server Virtualization Business Unit
> Cisco Systems
>  
> 
>     ------------------------------------------------------------------------
>     *From:* openfabrics-ewg-bounces at openib.org
>     [mailto:openfabrics-ewg-bounces at openib.org] *On Behalf Of *Scott
>     Weitzenkamp (sweitzen)
>     *Sent:* Monday, June 05, 2006 12:51 AM
>     *To:* openfabrics-ewg at openib.org
>     *Subject:* [openfabrics-ewg] Cisco SQA results so far for OFED 1.0 rc5
> 
>     We've focused on SDP and MPI performance so far on RHEL4 U3 x86_64. 
>     Some bad news:
> 
>         * MVAPICH MPI small message latency is slightly worse across the
>           board, as measured by OSU latency benchmarks.  For example:
>               o Lion Mini SDR: small message latency up from 3.43 usec
>                 to 3.53 usec
>               o Cheetah SDR: small message latency up from 3.57 usec to
>                 3.73 usec
>         * SDP throughput worse and small message latency worse than
>           rc4.   For example
>               o SDP small message latency */worse/* than IPoIB now
>               o Lion Mini SDR: throughput dropped from 7.24 Gb/sec to
>                 6.45 Gb/sec
>               o Lion Mini SDR: latency grew from 17.5 usec to 30.0 usec
> 
>     Scott Weitzenkamp
>     SQA and Release Manager
>     Server Virtualization Business Unit
>     Cisco Systems
>      
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> openfabrics-ewg mailing list
> openfabrics-ewg at openib.org
> http://openib.org/mailman/listinfo/openfabrics-ewg