Peter Tröger peter at troeger.eu
Thu Jan 27 04:47:34 CST 2011


sitting here with Thijs from Platform, we figured out that LSF does not have to large JobArray operations support as suggested below. In fact, you can only kill submitted job arrays as a whole.

I would add an accordingly restricted version of the interface to the spec.


Am 14.01.2011 um 22:36 schrieb Peter Tröger:

> Hi Nadav,
> first let me thank you again for the in-depth analysis, which gives us confidence that the current spec design is the right one to go.
> I personally like your job array argumentation, since the proposal extends the existing bulk job facility in a natural way. Another pro-argument is that the according bulk operations would be also implementable by the DRMAA library itself, if the DRMS does not support them. As usual, every feature extension has the danger of forgetting nasty side effects in the job control flow, as Mariusz mentioned. 
> I am willing to open up a discussion on that, so here is a first proposal. If we just take the current status and introduce a "set of jobs" representation, we end up with something like this:
> ==== snip ===
> interface JobSession {
> 		...
> 		Job runJob(in DRMAA::JobTemplate jobTemplate)
> 		JobArray runBulkJobs(in DRMAA::JobTemplate jobTemplate, in  long beginIndex, in long endIndex, in  long step)
>                 ...
> }
> interface JobArray {
> 		readonly attribute string jobArrayId;
> 		sequence<Job> jobs;
> 		readonly attribute JobSession session;
> 		readonly attribute JobTemplate jobTemplate;
> 		readonly attribute Reservation reservation;
> 		void suspend()     // suspend all jobs of the array, partial failures in changing the state are ok
> 		void resume()      // resume all jobs of the array, partial failures in changing the state are ok
> 		void hold()        // put a queued bulk job on hold
> 		void release()     // release an array job on hold
> 		void terminate()   // terminate a running job
> 		Job waitAnyStarted(in TimeAmount timeout)      // similar to JobSession function
> 		Job waitAnyTerminated(in TimeAmount timeout)   // similar to JobSession function
> };
> ==== snip ===
> Fetching status information makes only sense on job level, so the according getInfo() call is not part of the JobArray interface.  
> I would also resist the temptation to add a JobArray counterpart of getJobs(JobInfo filter), since the filter semantics would become horrible to specify. 
> All functions should be implementable with the 'loop' fallback in the library, when we allow partial success in the bulk control functions.
> DRMAA folks, your comments please. Is this a feasible interface for the denoted DRM systems with direct job array control support ?
> Best,
> Peter.
> Am 13.01.2011 um 09:23 schrieb Nadav Brandes:
>> The newer API specification does look a great deal better, and obviously I came up with some irrelevant questions.
>> I'll let you decide what you think about those issues I mentioned that are still relevant, but first I want to elaborate a little bit about the job-arrays feature, which is the most crucial feature for us.
>> When dealing with job arrays, each task actually has two IDs (The ID of the whole job-array, and the index of the task within the job-array).
>> Therefore, in job-arrays, all of the queries and actions that are performed on jobs according to the current DRMAA specification, are actually performed upon tasks, which are identified by two IDs instead of one, and except of that are perfectly similar to single jobs. 
>> All I said so far doesn't make any significant difference, and is only a matter of terminology. But the important thing about job-arrays is the ability to perform inclusive queries and operations on them.
>> For example, one can terminate all of the tasks in a job-array using a single command (supplying only the ID of the whole job-array, without needing to give the ID of each task, which might be very exhausting for users).
>> An example for a more advanced logic that one might want to perform on job-arrays is to rerun all the failed tasks in a given job-array.
>> Another advanced logic might be to limit the number of tasks that may run simultaneously in a job-array (for example, submitting a job-array containing 1000 tasks, where only 10 tasks are allowed to run simultaneously at a given time).
>> The greatest advantage of job-arrays, is the ability of users to "remember" many tasks with a single ID, what is impossible to do when submitting many single jobs.
>> Many schedulers (like LSF) support all these features, and you can see it implemented in a growing number of scheduler.
>> We believe that DRMAA should support these features as well, by being more "job-arrays oriented". I truly believe that DRMAA will be better if it supports job-arrays.
>> 2011/1/12 Mariusz Mamoński <mamonski at man.poznan.pl>
>> Hi Nadav,
>> On 12 January 2011 17:03, Nadav Brandes <nadavbrandes at gmail.com> wrote:
>> > Hello everyone,
>> >
>> > I went over your API description with my team (as described in
>> > http://www.drmaa.org/drmaav2_draft5.pdf).
>> please us the wiki as it is the most up to date version of the DRMAA spec:
>> http://wikis.sun.com/display/DRMAAv2/Home
>> >
>> >
>> >
>> > If it's not too late, we have few questions/suggestions:
>> >
>> > ·         Can one get a 'Job' object representing a job already submitted
>> > once, given only the job index (as an integer)?
>> It is supported: The JobSession has a method:
>>                sequence<Job> getJobs(JobInfo filter);
>> which as i remember is not constrained to jobs submitted via DRMAA.
>> >
>> > ·         It seems like the 'JobInfo' interface misses few parameters given
>> > in the 'JobTemplate' interface. For example, can one get the 'remoteCommand'
>> > of a job that was already submitted, if he only has a 'Job' object in hand,
>> > and not the 'JobTemplate'?
>> >
>> > ·         Does DRMAA support job-arrays feature (meaning submitting a group
>> > of tasks in one job, that has a single ID)? Most schedulers support this
>> > feature (include LSF, Moab and SGE). You do have a feature of 'runBulkJobs'
>> > that sends a sequence of jobs altogether, but it also returns a sequence of
>> > 'Job' objects, and not a single job with a single ID.
>> IMHO most of the batch systems returns many job ids for job arrays but
>> they offer to do perform some of the operations on the whole array
>> (bulk) by giving common suffix of those job ids. Having one job id,
>> thus one Job complicates state model (what if half of the array
>> sub-jobs are running and the rest queued? What should be the state of
>> the whole array job?)
>> >
>> > ·         Does DRMAA support the notion of queues (a feature that is
>> > supported by all of the schedulers I know)? We believe that it could be very
>> > useful if one could determine a queue in 'JobTemplate', change the queue of
>> > an existing job, and also get a list of all the queues in the cluster.
>> this was already addressed (wiki!), except alteration of target queue
>> of already submitted job.
>> >
>> > ·         Many batch systems have a feature that allows giving a 'project
>> > name' to submitted jobs. We believe that it could also be very useful if
>> > 'JobTemplate' had such field.
>> has: it is called accountingId
>> >
>> > ·         Sometimes, especially when dealing with large clusters containing
>> > a large number of compute nodes (which some of them might be out of order),
>> > jobs might fail randomly, without any justified reason. We think it could be
>> > useful if DRMAA supported a feature that allows rerunning failed jobs (as
>> > many schedulers allow, like LSF).  Such 'rerun()' method could be added to
>> > the 'Job' interface.
>> We have: rerunnable attribute of the JobTemplate. So one can configure
>> batch system to rerun jobs that failed due to resources failure
>> >
>> > ·         Modern schedulers (like Moab and LSF) support advanced features of
>> > memory management, cores management, and also general resources management
>> > (like GPUs). In general, it means giving a list of required resources to
>> > each submitted job (for example, submitting a job that requires 5 cores,
>> > 12GB RAM, and 2 GPUs). Then the scheduler knows how to schedule the jobs so
>> > each running job will have all the resources it needs. If 'JobTemplate' had
>> > a resources dictionary field, it could also be very useful.
>> resources that are common for all schedulers are expressed as
>> JobTemplate attributes, e.g.: minPhysMemory
>> others DRMS specific options (also resources requirements)
>>  should go to:          attribute Dictionary drmsSpecific;
>>   // must be supported
>> >
>> >
>> >
>> > This is it for now, thank for reading it.
>> thanks for providing your comments, and sorry that you lost much of
>> time of reading very old version of the specification (@Peter: maybe
>> it would be better to delete reference to the September 2009, DRMAA2
>> Draft 5)
>> >
>> > I would like to hear what you think.
>> >
>> >
>> >
>> > Best Regards,
>> >
>> > Nadav
>> >
>> > 2010/12/21 Peter Tröger <peter at troeger.eu>
>> >>
>> >> Hi Navad,
>> >>
>> >> Now I saw the documentation of the planned interface for DRMAA2, and I
>> >> find it to be a great improvement, and very useful for my organization. I am
>> >> truly anxious to try it, and have some more questions about its release:
>> >>
>> >> Do you know which distributed resource manager will be the first to
>> >> implement DRMAA2? (SGE maybe?) Also, do you have any estimation on when
>> >> it'll happen, and when will I be able to download a trial version of it?
>> >>
>> >> Since we have the Oracle Grid Engine Product Manager as one of the
>> >> co-chairs, I leave the implementation estimation to you ;-) .... We also
>> >> have very capable people in Poznan, which might take care of non-OGE
>> >> systems. We assume to put out the spec in January, and from there, the group
>> >> can only hope. From experience, I would expect nothing useful before Summer
>> >> 2011.
>> >>
>> >> Is it still possible to suggest ideas that we have about the interface of
>> >> DRMAA2? If so, how is it done? Is it customary to share ideas in this forum,
>> >> or do you prefer it to be done through Wiki?
>> >>
>> >> The best thing is to start a discussion on the list. The Wiki is good as
>> >> reference. Comments on the Wiki pages might get lost ...
>> >> Best regards,
>> >> Peter.
>> >
>> >
>> > --
>> >  drmaa-wg mailing list
>> >  drmaa-wg at ogf.org
>> >  http://www.ogf.org/mailman/listinfo/drmaa-wg
>> >
>> Best Regards,
>> --
>> Mariusz
>> --
>>  drmaa-wg mailing list
>>  drmaa-wg at ogf.org
>>  http://www.ogf.org/mailman/listinfo/drmaa-wg
> --
>  drmaa-wg mailing list
>  drmaa-wg at ogf.org
>  http://www.ogf.org/mailman/listinfo/drmaa-wg

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.ogf.org/pipermail/drmaa-wg/attachments/20110127/77bf2cd1/attachment-0001.html 

More information about the drmaa-wg mailing list