[mpich-discuss] Hydra binding questions

[Jain, Rohit]

Hi Pavan,

I have filed ticket for it.

I have few follow up questions.

1. We used taskset with Hydra. We are seeing about 20% slowdown with (2-1.4 Hydra + taskset) Vs (2-1.0.6 mpd + taskset) on same machine with same taskset binding. We were expecting Hydra to perform same or faster than mpd. Is that not true?
 
2. We are seeing relatively high system activity seen in 'top' using new mpich (with or without hydra binding). Also, user time on the cpu cores hovers around 98-99% compared to old mpich (with mpd) where is always nearly 100%. Is that expected due to new Hydra?

3) Do you have any performance benchmarks with mpd vs hydra? Above issues are hindering our plan to upgrade to Hydra.

Regards,
Rohit

-----Original Message-----
From: mpich-discuss-bounces at mcs.anl.gov [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of Pavan Balaji
Sent: Tuesday, August 02, 2011 9:49 AM
To: mpich-discuss at mcs.anl.gov
Cc: Jain, Rohit
Subject: Re: [mpich-discuss] Hydra binding questions


On 08/01/2011 06:38 PM, Jain, Rohit wrote:
> 1)Is there any debug output by Hydra, which tells particular
> core/processor binding (core numbers) associated with a process? We can
> use this info to see variation of binding methods on different
> architectures.

Yes, good point. We don't print out the binding information in the 
verbose output currently. I'll add that in. Can you add a ticket for 
this? https://trac.mcs.anl.gov/projects/mpich2/newticket

> 2)We expected cpu:sockets and cache:l3 binding to behave same on an AMD
> box due to its layout (2 processors with 6 cores each). But, there is
> difference in performance of about 10%. Is that expected? Again, any
> information on binding will help, currently it is black box for us of
> what Hydra does.

Once I add print out the binding, you'll know what binding it's using. 
That might help debug this.

> 3)While running on AMD Opteron machine, I monitored 'top' as application
> was running. I found that processes are shuffling between different
> cores during the run. Is that expected, as processes are bind to a
> particular core? I also noticed it on Intel machine, but happening less
> frequently.

That shouldn't happen.

> 4)We currently use numactl/taskset to set affinity with mpd. Does hydra
> allow specifying affinity with numactl/taskset, instead of -binding?

Yes, of course. numactl/taskset are independent of the process manager. 
You can use them with any process manager.

> 5)What parameter does hydra read from machine to determines binding?
> /proc/cpuinfo or more?

It uses hwloc to get the process topology information.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji
_______________________________________________
mpich-discuss mailing list
mpich-discuss at mcs.anl.gov
https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss