Message boards : CMS Application : Multi core CMS
Message board moderation

To post messages, you must log in.

AuthorMessage
Harri Liljeroos
Avatar

Send message
Joined: 28 Sep 04
Posts: 674
Credit: 43,152,472
RAC: 15,698
Message 29135 - Posted: 9 Mar 2017, 16:11:35 UTC
Last modified: 9 Mar 2017, 16:29:40 UTC

I am running some CMS tasks with 2 CPU cores using app_config. But so far the results seem to point that this is not useful. The tasks seem to take as long as when running single core. The CPU time is copied to the run time on task pages showing twice as long time as it actually took. Here's one task: https://lhcathome.cern.ch/lhcathome/result.php?resultid=123844671
So it ties 2 CPU cores to the task and I get double credits for it. So I don't see the point doing this?

Edit: Actually this might be bad for project as only half of the work gets done.
ID: 29135 · Report as offensive     Reply Quote
Toby Broom
Volunteer moderator

Send message
Joined: 27 Sep 08
Posts: 798
Credit: 644,764,636
RAC: 232,155
Message 29137 - Posted: 9 Mar 2017, 16:49:52 UTC

I thought no I have seen some on Dev project, so coming in the future
ID: 29137 · Report as offensive     Reply Quote
Crystal Pellet
Volunteer moderator
Volunteer tester

Send message
Joined: 14 Jan 10
Posts: 1268
Credit: 8,421,616
RAC: 2,139
Message 29140 - Posted: 9 Mar 2017, 17:04:16 UTC

It's be a bit different:

Multi-core on ATLAS is processing 1 job on all allocated cores and the BOINC-task is ready when the job within the VM has finished.
Multi-core VM's for Theory and CMS are running a job on each core allocated to the VM. So a quad-core VM is doing 4 jobs at the same time.
When the elapsed time since the VM booted is still below 12 hours, the VM will request a new job after 1 of the busy ones has finished.

See a part of your linked result:

2017-03-09 17:01:23 (1268): Guest Log: [INFO] Job finished in slot1 with unknown exit code.
2017-03-09 17:38:56 (1268): Guest Log: [INFO] Job finished in slot2 with unknown exit code.
2017-03-09 17:49:07 (1268): Guest Log: [INFO] Condor exited with return value N/A.
2017-03-09 17:49:07 (1268): Guest Log: [INFO] Shutting Down.


Disadvantage is that towards the end of a task (normally 12 hours) not all jobs are finishing at the same time, so there will be idle time on 1 or more cores until the last job has finished.

Advantage is that you need to reserve less RAM for a multicore VM.
ID: 29140 · Report as offensive     Reply Quote
Harri Liljeroos
Avatar

Send message
Joined: 28 Sep 04
Posts: 674
Credit: 43,152,472
RAC: 15,698
Message 29146 - Posted: 9 Mar 2017, 18:47:02 UTC - in response to Message 29140.  

OK, I see. The memory consumption was the reason I started this experiment. I reserved 4400 MB memory for the two core task but Boinc shows the same amount memory consumed as a single core task (2384 MB) which isn't correct. Do you know the formula for counting the required memory of multi core CMS tasks?
ID: 29146 · Report as offensive     Reply Quote
Crystal Pellet
Volunteer moderator
Volunteer tester

Send message
Joined: 14 Jan 10
Posts: 1268
Credit: 8,421,616
RAC: 2,139
Message 29147 - Posted: 9 Mar 2017, 20:30:28 UTC - in response to Message 29146.  

Do you know the formula for counting the required memory of multi core CMS tasks?

What BOINC shows need not to be the same as what is used.
BOINC only reserves that amount of memory to calculate, whether there is enough RAM free for other tasks.

I tried last October 2 CMS-VM's simultaneously running with 4 cores with app_config.xml on the LHC-dev project and that was running fine.

The app_version part of app_config.xml should look like:
 <app_version>
  <app_name>CMS</app_name>
  <plan_class>vbox64</plan_class>
  <avg_ncpus>4.000000</avg_ncpus>
  <cmdline>--memory_size_mb 6144</cmdline>
 </app_version>

You could even try a lower RAM value and watch the console (ALT-F3), whether there is (after e.g. 1 hour running) still free RAM and/or how many memory is swapped.
ID: 29147 · Report as offensive     Reply Quote
Harri Liljeroos
Avatar

Send message
Joined: 28 Sep 04
Posts: 674
Credit: 43,152,472
RAC: 15,698
Message 29170 - Posted: 10 Mar 2017, 15:30:45 UTC - in response to Message 29147.  

Thanks for the info. I checked on one task that has been running for 8 hours and it showed about 3820 MB used, 530 MB free and 0 MB swapped. Mem setting was 4400 MB with 2 cores. So maybe I can drop the mem setting with a couple of hudred MB?

I am running also some Seti and Einstein tasks on the GPUs (2 x GTX970) and 1 CPDN task on CPU with 2 CMS tasks (2 core). This is a i7-3770 with 16 GB memory (now 91% of memory is being used) and it is also used as my daily driver with email, web browsing and DLNA server (usually just streaming music).
ID: 29170 · Report as offensive     Reply Quote
Erich56

Send message
Joined: 18 Dec 15
Posts: 1686
Credit: 100,391,895
RAC: 102,144
Message 29426 - Posted: 20 Mar 2017, 8:29:49 UTC

I intended to run 2 CMS tasks with 2 CPU cores each.
So, in the settings (Homepage) I put:
max # Jobs: 2
max # CPUs: 2

then, 2 tasks were downloaded, however, each one running on 1 core only. Why so?

Before, with the same settings for ATLAS, each task ran on 2 cores.
ID: 29426 · Report as offensive     Reply Quote
maeax

Send message
Joined: 2 May 07
Posts: 2071
Credit: 156,137,814
RAC: 105,274
Message 29427 - Posted: 20 Mar 2017, 8:34:42 UTC

Multicore for CMS and LHC are only in -dev possible.
ID: 29427 · Report as offensive     Reply Quote
Crystal Pellet
Volunteer moderator
Volunteer tester

Send message
Joined: 14 Jan 10
Posts: 1268
Credit: 8,421,616
RAC: 2,139
Message 29434 - Posted: 20 Mar 2017, 11:19:14 UTC - in response to Message 29427.  

Multicore for CMS and LHC are only in -dev possible.

It's possible with an app_config.xml, but will not speed up the task, cause on each core a job will run.
Due to several jobs running in one VM and the fact they will not all finish at the same time, towards the end of a task one or more cores will run idle until the last job has finished too.

So multi-core for CMS, Theory and LHCb is performance-wise only useful for machines with low RAM not able to run several single cores.
ID: 29434 · Report as offensive     Reply Quote

Message boards : CMS Application : Multi core CMS


©2024 CERN