Hi Goncalo,
that's absolutely right, I have tried it out, the current kernels ignore s_rss (as well as the corresponding 'ulimit -m' bash setting).
Nevertheless, GlueHostMainMemoryVirtualSize (which should be passed to s_vmem according to the GLUE schema usage guide) was ignored by the original sge_local_submit_attributes.sh. GlueHostMainMemoryRAMSize should set the physical RAM size instead of vmem.
We'll try to check with Univa which alternative settings are supported by the current UGE release.
Cheers,
Manfred
________________________________________
Von: Gonçalo Borges [[log in to unmask]]
Gesendet: Donnerstag, 27. Februar 2014 14:50
An: Alef, Manfred (SCC); LHC Computer Grid - Rollout
Betreff: Re: AW: [LCG-ROLLOUT] CREAM: CERequirements not parsed from jdl to batch system?
Hi Manfred...
Just a comment after looking to your patch...
I think in the latest 2.6 kernels, ulimit does not effectively enforces
resident memory limits anymore.
Check the following thread:
http://stackoverflow.com/questions/3043709/resident-set-size-rss-limit-has-no-effect/6365534#6365534
So, basically, the rss set will not work. I think the new paradigm is to
use climits instead, but I'm not sure
if LRMS are already prepared for that.
So, to conclude, although the "-l s_rss" substitution you have set,
although philosophically correct,
it will not work, and I think that was the reason why "-l s_vmem" was
used also to set GlueHostMainMemoryRAMSize
Cheers
Goncalo
On 02/27/2014 11:56 AM, Alef, Manfred (SCC) wrote:
> Hi Goncalo,
>
> here it is (my first posting got rejected).
>
> Cheers,
> Manfred
>
>
> ________________________________________
> Von: Gonçalo Borges [[log in to unmask]]
> Gesendet: Donnerstag, 27. Februar 2014 12:45
> An: LHC Computer Grid - Rollout; Alef, Manfred (SCC)
> Betreff: Re: [LCG-ROLLOUT] CREAM: CERequirements not parsed from jdl to batch system?
>
> Hi Manfred...
>
> Let me know the patch (case you want to distribute it) and we can add it
> to blahp packages.
>
> Cheers
> Goncalo
>
> On 02/27/2014 11:30 AM, Thomas Hartmann wrote:
>> Hi Goncalo,
>>
>> regarding the memory requirements parsed by
>> sge_local_submit_attributes.sh Manfred asked my to forward his answer
>> (his mail got rejected)
>>
>> Hi Goncalo,
>>
>>> I guess you have been changing sge_local_submit_attributes.sh
>> yes, I have changed sge_local_submit_attributes.sh to implement the
>> proper translation of $GlueHostMainMemoryRAMSize to s_rss (not: r_vmem
>> according to the GLUE schema usage guide), and
>> $GlueHostMainMemoryVirtualSize to s_vmem.
>>
>> Cheers,
>> Manfred
>>
>>
>> On 27.02.2014 11:48, Gonçalo Borges wrote:
>>> Yup.. Can't access even with my CERN account (I'm not in ATLAS group)
>>>
>>> I guess you have been changing sge_local_submit_attributes.sh to
>>> implement those changes, right?! If you send us the patch, we can check
>>> it and try to push it into BLAHP package.
>>>
>>> Cheers
>>> Goncalo
>>>
>>> On 02/27/2014 10:28 AM, Thomas Hartmann wrote:
>>>> Hi Jeff,
>>>>
>>>> I started sketched our experiences with jdl vs. SGE on a twiki page in
>>>> AtlasComputing
>>>>
>>>> https://twiki.cern.ch/twiki/bin/view/AtlasComputing/JobRequirementsSGE
>>>>
>>>> btw: since the page got automatically behind an ACL wall: is there a
>>>> wiki, that is usable without CERN/LHC VO affiliations in case site
>>>> admins without affiliations need access?
>>>>
>>>> Cheers,
>>>> Thomas
>>>>
>>>> On 27.02.2014 09:22, Jeff Templon wrote:
>>>>> Yo,
>>>>>
>>>>> Dennis van Dok's results for torque / maui are here:
>>>>>
>>>>>
>>>>> http://wiki.nikhef.nl/grid/Passing_job_requirements_through_the_WMS
>>>>>
>>>>> AFAICT there is a degree of arbitrariness in which (and how)
>>>>> requirements are translated from the JDL to the batch system. My
>>>>> opinion is that some small team should sit down and decide what the
>>>>> desired behavior is, and then work on implementing this behavior on
>>>>> each of the batch systems. It would be really great if this same
>>>>> behavior is reproduced when passing requirements to an ARC CE or a
>>>>> GRAM 5 CE.
>>>>>
>>>>> The coordination sounds like a job for EGI, however I think the input
>>>>> to this coordination has to come from a discussion between sites and
>>>>> the users; sites should drive the discussion, as the sites are the
>>>>> parties that understand the use cases from multiple user
>>>>> communities. Since the WLCG multicore TF is already working on
>>>>> related issues, we can probably find at least a subset of the right
>>>>> group of people there. Thomas is one of them!
>>>>>
>>>>> JT
>>>>> On Feb 26, 2014, at 19:18 , Thomas Hartmann
>>>>> <[log in to unmask]> wrote:
>>>>>
>>>>>> Hi Goncalo and Maarten,
>>>>>>
>>>>>> after some more testing (and debugging of my script...) we were able to
>>>>>> get CERequirements getting piped to SGE, e.g.
>>>>>>
>>>>>> CERequirements = "other.GlueHostMainMemoryVirtualSize > 2000";
>>>>>> or other.GlueCEPolicyMaxWallClockTime > 30
>>>>>>
>>>>>> So far, we tested successfully the CPUTime, WallTime, Mem and VMem
>>>>>> parameters passing from the jdl to the finally submitted SGE
>>>>>> wrapper/job.
>>>>>>
>>>>>> Do you know, if there are further experiences for SGE and other batch
>>>>>> systems how the implementation/support is for the various parameters?
>>>>>>
>>>>>> Cheers and many thanks for your help,
>>>>>> Thomas
>>>>>>
>
|