[gridengine users] Problems with dmtcp migration between processor versions

Reuti reuti at staff.uni-marburg.de
Mon Nov 19 09:39:36 UTC 2012


Am 19.11.2012 um 00:19 schrieb Dave Love:

> Reuti <reuti at staff.uni-marburg.de> writes:
> 
>>> I don't know how proprietary libraries, such as for BLAS adjust to
>>> different ABIs, but they may set up dispatch tables initially, not on
>>> each call, so compilation options might not make any difference.
>> 
>> If it would be compiled for only one platform, then there wouldn't be
>> any alternate execution path. But it's indeed strange that a better
>> processor had this problem.
> 
> I meant that they may dispatch on something like cpuid to adjust to the
> architecture at runtime, and presumably maintain the state for the
> architecture on which they start.  I guess you could get the same sort
> of thing for dynamic compilation ("JIT"?).
> 
>>> Presumably it would be a reasonable feature to have an option only to
>>> restart with the same resources as granted originally.
>> 
>> If it's only the CPU type (and a fixed PE incase a wildcard request
>> was used initially), it should be possible to use `qalter` before
>> killing the process to limit it to such type of nodes/PEs.
> 
> How do you access the list of granted resources (as apposed to requested
> ones)?  I thought that needed (a free) implementation, which I want for
> other reasons.

`qstat -r` show the requested and granted PE. For other resource requests: maybe in this particular case it's just necessary to check the hostgroup you are running in.

- Reuti



More information about the users mailing list