[GE users] h_vmem as consumable and stack size

andy andy.schwierskott at sun.com
Tue Mar 3 12:41:25 GMT 2009


Hi,

I can provide some more insight :-)

A user running Gaussian was running into a similar problem: Gaussian did not
behave properly when *_stack could not be set > *_vmem. We provided the user
a shepherd binary where the last four lines of the code are *not* executed:

source/daemons/shepherd/setrlimits.c
[snip]
   /*
    * s_vmem > h_vmem
    * data segment limit > vmem limit
    * stack segment limit > vmem limit
    * causes problems that are difficult to find
    * we try to prevent these problems silently
    */
   if (s_vmem > h_vmem) {
      s_vmem = h_vmem;
   }
   s_data = RL_MIN(s_data, s_vmem);
   s_stack = RL_MIN(s_stack, s_vmem);
   h_data = RL_MIN(h_data, h_vmem);
   h_stack = RL_MIN(h_stack, h_vmem);
[snap]

The user confirmed that Gaussian is working properly with such a shepherd
binary.

This evaluation was done a little late for SGE 6.2u2 (comming out these
days), so we plan to integrate this change in SGE 6.2u3 (planned for late
May) and possibly (not yet confirmed) for the next SGE 6.1 patch (no date
planned).

Frankly I don't think it's a SGE bug, but an application bug of Gaussian
(and Matlab which was referenced in this thread). However if it helps to
solve our user's problem I think we scan fix this in SGE than waiting until
the Gaussian and Matlab vendors are able or willing to fix this in their
software.

I would like to know of anyone who is experiencing this problem had raised a
ticket with the Gaussian/Matlab vendors or if there is possibly already an
answer available?

Anyhone who can not wait for the next SGE patches and who need to use limits
other than "infinity" can compile a shepherd binary where the affected four
lines of code are commented out. If you don't want to use CVS, tar.gz
snapshots of many SGE versions and patches are available under

   http://gridengine.sunsource.net/servlets/ProjectDocumentList


Andy


On Tue, 3 Mar 2009, reuti wrote:

> Hi Sabine,
>
> Am 03.03.2009 um 10:12 schrieb s_kreidl:
>
> > Ok, thanks Reuti.
> >
> > Is there a special reason, why you set the stack and data size
> > limits in conjunction with the virtual memory size?
> > And do you plan on leaving the procedure like that in later
> > versions of the SGE?
>
> I don't have more insights about it than the comments in the source:
>
>     /*
>      * s_vmem > h_vmem
>      * data segment limit > vmem limit
>      * stack segment limit > vmem limit
>      * causes problems that are difficult to find
>      * we try to prevent these problems silently
>      */
>
> And hence the limit is for h_data and h_stack is set in addition to
> h_vmem. This ist still the actual behavior in the latest CVS.
>
> -- Reuti
>
>
> > Thanks,
> > Sabine
> >
> > ------------------------------------------------------
> > http://gridengine.sunsource.net/ds/viewMessage.do?
> > dsForumId=38&dsMessageId=119456
> >
> > To unsubscribe from this discussion, e-mail: [users-
> > unsubscribe at gridengine.sunsource.net].
>
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=119506
>
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
>

------------------------------------------------------
http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=119559

To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].



More information about the gridengine-users mailing list