Hi Andreas
Without s_vmem=1G complex lfc-mkdir works fine, we will test it setting
h_stack.
Cheers and thanks
Alvaro
> Hi Alvaro,
>
> you might try to limit h_stack to 10M for all jobs (or as a queue
> limit). SGE (imho wrongly - I think, it's a bug...) sets the stack size
> limit to the same value as the virtual memory limit (h_vmem). All jobs
> using threads (python, java, ...) show exactly your mentioned problem
> with those shell limits.
>
> Cheers,
> Andreas
>
> On Fri, 2008-10-10 at 09:20 +0200, Alvaro Simon Garcia wrote:
>
>> Hi Dennis
>>
>> Our SGE batch system is configured to limit user memory, It's set to
>> 1GB, but as you said it seems that lfc-mkdir needs more... We will open
>> a bug regarding this.
>>
>> Cheers and thanks
>> Alvaro
>>
>>> Alvaro Simon Garcia schreef:
>>>
>>>> Hi
>>>>
>>>> We have found some diferences...
>>>>
>>>> on WN
>>>> #ulimit -a
>>>>
>>>> virtual memory (kbytes, -v) unlimited
>>>>
>>>> with batch system:
>>>> #ulimit -a
>>>> virtual memory (kbytes, -v) 1048576
>>>>
>>>> our strace output:
>>>>
>>> [...]
>>>
>>>> mmap2(NULL, 1073745920, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS,
>>>> -1, 0) = -1 ENOMEM (Cannot allocate memory)
>>>>
>>> [...]
>>>
>>> You've got all the evidence right there. Set ulimit of virtual memory to
>>> the amount used by the batch system, this should result in the same error.
>>>
>>> I don't know who is responsible for trying to mmap over a GB of memory,
>>> but I would consider that a bug ;-)
>>>
>>> Cheers,
>>>
>>> Dennis van Dok
>>>
|