[mvapich-discuss] job metrics

[RIVERA, Orlando]

Hi Walid

You could take a look to LIKIWD , you can get Performance  Sampling every Time delay  for a certain time spam.  You will  measure  not only the application , but also the counters  related to the cpus/cores,  since you don’t need to manually  to instrument any application

(From the Likiwd site  : stethoscope mode: Measure performance counters for a variable time duration independent of any code running)   I think this is what you are looking for.

You could also try the intel performance monitor (you need to have intel Hardware ) , the advantage is , that you can measure also uncore counters and energy  use directly


And if you want to profile MPI, I would recommend IPM2 (look for it  with google )  I  submitted  small patches twice and I think the development  stalled for a while. Anyways  here you don’t need to instrument re-link or re-compile your application . IPM2 is just a simple dynamic library , which you can put between your Application and the MPI library through  LD_PRELOAD to get the metrics . The only that  you require is that your MPI-LiB supports PMPI week symbols which is most probably your case

Regards

Orlando


Von: mvapich-discuss [mailto:mvapich-discuss-bounces at cse.ohio-state.edu] Im Auftrag von Walid
Gesendet: 30 April 2014 03:55
An: Hari Subramoni
Cc: mvapich-discuss at cse.ohio-state.edu
Betreff: Re: [mvapich-discuss] job metrics

Hari, Jennifer,

Thanks for your feedback, i am representing the system side in my organisation, and application support would like to have these metrics as a regular feature, part of the job output. what i understand from both of you and a colleague of mine that you usually do profiling as part of a performance exercise, or to understand more about the job characteristics, we do use Infiniband and ethernet as interconnects, and some of our jobs do extend large number of cores up in 100s, and sometime a couple of 1000 cores. would not that be collecting traces from every node and writing it causes additional overhead, does the collection and saving of metrics scale? apart from profilers do you do anything different to monitor job trends, and correlate behaviour, or capacity planing?

kind regards

Walid

On 30 April 2014 00:33, Hari Subramoni <subramoni.1 at osu.edu<mailto:subramoni.1 at osu.edu>> wrote:
Hello Walid,

There are several profilers like mpiP, IPM, TAU and HPCToolkit that allows folks to characterize the job and workload in terms of number of / time spent in MPI calls with relatively low overheads. The amount of overhead cause and the level and kind of detail provided differs from profiler to profiler.
Thx,
Hari.

On Tue, Apr 29, 2014 at 5:08 AM, Walid <walid.shaari at gmail.com<mailto:walid.shaari at gmail.com>> wrote:
Hi,

What would be the best way to characterize the job and workload it processed in terms of mpi calls and communications, how is easy to profile a job run and how much overhead does the profiling cause?

any one doing extensive parallel jobs profiling, capacity planing that would like to share their thoughts?

kind regards

Walid

_______________________________________________
mvapich-discuss mailing list
mvapich-discuss at cse.ohio-state.edu<mailto:mvapich-discuss at cse.ohio-state.edu>
http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss


--
MTU Aero Engines AG
Vorstand/Board of Management: Reiner Winkler, Vorsitzender/CEO; Dr. Rainer Martens, Michael Schreyögg, Dr. Stefan Weingartner
Vorsitzender des Aufsichtsrats/Chairman of the Supervisory Board: Klaus Eberhardt
Sitz der Gesellschaft/Registered Office: Muenchen
Handelsregister/Commercial Register: Muenchen HRB 157206

Diese E-Mail sowie ihre Anhaenge enthalten MTU-eigene vertrauliche oder rechtlich geschuetzte Informationen.
Wenn Sie nicht der beabsichtigte Empfaenger sind, informieren Sie bitte den Absender und loeschen Sie diese
E-Mail sowie die Anhaenge. Das unbefugte Speichern, Kopieren oder Weiterleiten ist nicht gestattet.

This e-mail and any attached documents are proprietary to MTU, confidential or protected by law.
If you are not the intended recipient, please advise the sender and delete this message and its attachments.
Any unauthorised storing, copying or distribution is prohibited.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20140430/a426beb8/attachment-0001.html>