Hi Louise,
I see thank you for the response, maybe there was some confusion, the feature I was trying to explain with LSF is that you *don't* need to tell it the required resources for a job and it will still be able to run all the submitted jobs on a node without crashing even if the jobs submitted need e.g. 10 more cores are available (that is 10 more cores than LSF thought it needed). LSF will just temporarily suspend jobs in mid-run on a node to keep the load down, but nothing will ever crash even if you are running jobs that require 20 CPUs and you only have 2. Thought maybe there was a way to do this with TORQUE. If LSF or TORQUE are explcitly passed the resources needed then they will never need to temporarily suspend because they will pick a node with those resources free. That being said, your method is more efficient for this reason as it will pick the right node with the cores available instead of picking a node with maybe just one core available and then running the multithreaded job slower because it has to periodically suspend it as it is running.
Also I wonder do you run any Java command-line tools via Galaxy? I can't seem to find with the JVM exactly how many cores it needs during execution or how to limit it to a certain max, it just jumps around in CPU usage from 50% to over 400%
regards,
Leandro
2011/5/19 Louise-Amélie Schmitt <louise-amelie.schmitt@embl.de>
Hi again Leandro
Well I might not have been really clear, perhaps I should have re-read the mail before posting it :)
The thing is, it was not an issue of Torque starting jobs when there were not enough resources available, but rather it believing the needed resources for each job being fewer that they were (e.g. always assuming the jobs were single-threaded even if the actual tools needed more tan one core). if Torque is properly notified of the needed resources, it will dispatch them or make them wait accordingly (since it knows the nodes' limits and load), like your LSF does.
This hack is not very sexy but it just notifies Torque of the cores needed by every multithreaded tool, so it doesn't run a multithreaded job when there's only one core available in the chosen node.
Hope that helps :)
Regards,
L-A
On 05/19/2011 03:05 PM, Leandro Hermida wrote:Hi Louise-Amelie,
Thank you for the post reference, this is exactly what I was looking for. For us for for example when I want to execute a tool that is a Java command the JVM typically will typically use multiple cores as it's running. You said with TORQUE it will crash when there aren't enough resources when the job is submitted. I wonder if you can do the same thing we have done here with LSF? With LSF you can configure a maximum server load for each node and if the submitted jobs push the node load above this threshold (e.g. more cores requested than available) LSF will temporarily suspend jobs (using some kind of heuristics) so that the load stays below the threshold and unsuspend as resources become available. So for us things just will run slower when we cannot pass the requested number of cores to LSF.
I would think maybe there is a way with TORQUE to have it achieve the same thing so jobs don't crash when resources requested are more than available?
regards,
Leandro
2011/5/19 Louise-Amélie Schmitt <louise-amelie.schmitt@embl.de>
Hi,
In a previous message, I explained how I did to multithreads certain jobs, perhaps you can modify the corresponding files for drmaa in a similar way:
On 04/26/2011 11:26 AM, Louise-Amélie Schmitt wrote:Just one little fix on line 261: 261 if ( len(l) > 1 and l[0] == job_wrapper.tool.id ): Otherwise it pathetically crashes when non-multithreaded jobs are submitted. Sorry about that. Regards, L-A Le mardi 19 avril 2011 à 14:33 +0200, Louise-Amélie Schmitt a écrit :Hello everyone, I'm using TORQUE with Galaxy, and we noticed that if a tool is multithreaded, the number of needed cores is not communicated to pbs, leading to job crashes if the required resources are not available when the job is submitted. Therefore I modified a little the code as follows in lib/galaxy/jobs/runners/pbs.py 256 # define PBS job options 257 attrs.append( dict( name = pbs.ATTR_N, value = str( "%s_%s_% s" % ( job_wrapper.job_id, job_wrapper.tool.id, job_wrapper.user ) ) ) ) 258 mt_file = open('tool-data/multithreading.csv', 'r') 259 for l in mt_file: 260 l = string.split(l) 261 if ( l[0] == job_wrapper.tool.id ): 262 attrs.append( dict( name = pbs.ATTR_l, resource = 'nodes', value = '1:ppn='+str(l[1]) ) ) 263 attrs.append( dict( name = pbs.ATTR_l, resource = 'mem', value = str(l[2]) ) ) 264 break 265 mt_file.close() 266 job_attrs = pbs.new_attropl( len( attrs ) + len( pbs_options ) ) (sorry it didn't come out very well due to line breaking) The csv file contains a list of the multithreaded tools, each line containing: <tool id>\t<number of threads>\t<memory needed>\n And it works fine, the jobs wait for their turn properly, but information is duplicated. Perhaps there would be a way to include something similar in galaxy's original code (if it is not already the case, I may not be up-to-date) without duplicating data. I hope that helps :) Best regards, L-A ___________________________________________________________ The Galaxy User list should be used for the discussion of Galaxy analysis and other features on the public server at usegalaxy.org. Please keep all replies on the list by using "reply all" in your mail client. For discussion of local Galaxy instances and the Galaxy source code, please use the Galaxy Development list: http://lists.bx.psu.edu/listinfo/galaxy-dev To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
On 05/19/2011 12:03 PM, Leandro Hermida wrote:___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/Hi,
When Galaxy is configured to use the DRMAA job runner is there a way for a tool to tell DRMAA the number of cores it would like to request? The equivalent of bsub -n X in LSF where X is min number of cores to have available on node.
best,
leandro