[mvapich-discuss] Announcing the release of MVAPICH2 0.9.5 with SRQ, intergrated multi-rail and TotalView support

[Pavel Shamis (Pasha)]

You measurements are absolutely correct the difference in IB send/recv 
latency is much bigger. But in most cases in mvapich small messages will 
be send via fast_path that doesn't effected by SRQ performance, is not it ?

Thanks,
Pasha

Sayantan Sur wrote:

> I used the standard `perftest' distributed by OpenIB. In particular, the 
> tests send_lat.c and send_bw.c seem to measure the latency and bandwidth 
> of send/recv operations. I trivially modified them to use SRQ instead of 
> posting receive to the Queue pair receive queue. The performance numbers 
> are noted below. Based on these numbers, it seems it is higher than the 
> 2-3% threshold. May I request you post your Gen2 level comparison 
> numbers? If there is something simple I'm missing, I'd like to correct it.
> 
> Thanks,
> Sayantan.
> 
> ===========
> Gen2 Latency (us)
> #Size    SRQ      Send/Recv
> 2        10.66    6.06
> 4        10.65    6.07
> 8        10.65    6.12
> 16       10.71    6.12
> 32       10.76    6.31
> 64       10.89    6.51
> 128      11.04    6.51
> 256      11.58    7.00
> 512      12.56    7.99
> 1024     13.87    9.40
> 
> Gen2 Bandwidth (MB/s)
> #Size    SRQ      Send/Recv
> 2        0.14     0.81
> 4        0.29     1.68
> 8        0.57     3.26
> 16       1.14     6.68
> 32       2.34     13.46
> 64       4.55     26.08
> 128      9.06     51.99
> 256      18.22    105.34
> 512      36.44    214.36
> 1024     75.05    423.72
> 
> 
> 


-- 
Pavel Shamis (Pasha)
Software Engineer
Mellanox Technologies LTD.
pasha at mellanox.co.il