[mpich-discuss] ch3:nemesis performance and SMP cluster

[Rajeev Thakur]

Not sure what you are saying here. What behavior are you seeing with
1.1.1p1?

Rajeev 

> -----Original Message-----
> From: mpich-discuss-bounces at mcs.anl.gov 
> [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of 
> Vamvakopoulos Manolis
> Sent: Saturday, September 05, 2009 5:09 PM
> To: mpich-discuss at mcs.anl.gov
> Subject: [mpich-discuss] ch3:nemesis performance and SMP cluster
> 
> Dear Developers and Users
> 
> In our cluster we have already install  mpich2-1.0.8 as a 
> default production
> version configure with ch3:ssm device.
> 
> With the above lib ,an application  with alltoall 
> communication (let's say
> amber9) "runs" efficiently  on 16 procs with 1 procs per node  (16x1)
> and 16 procs with two procs per node (8x2) (run's time are comparable)
> 
> --------------------------------------------------------------
> ---------------
> I make run test with same application and  mpich2-1.1.1p1 
> with nemesis device
> and the same tests was faster (up to 30%) on first case (one 
> proc per node)
> 
> same behavior i found with  mpich2-1.1.1p and ch3:ssm device
> 
> 
> In all case with mpich2-1.1.1p the runs on 16 procs on 16 
> different node were
> faster
> 
> -----
> Our cluster consist from DL145 HP with 2 single CPU: opteron 
> 248 2.2 MHZ with
> 4GB per node with linux 2.6.9-78.0.1.ELsmp SL 4.7 and for the 
> interconnect we
> use a full crossbar dedicate Gigabit switch (with high 
> backplane 400Gb)
> 
> can somebody  shed a light to this issue
> 
> thank you very much in advance
> 
> E.V.
> 
> -- 
> University OF Ioannina
> Department of Computer Science
> P.O. BOX 1186 Ioannina, Greece
> Tel: (+30)-26510-98864
> Fax: (+30)-26510-98890
> 
> 
> 
>