Am 22.08.2012 um 15:54 schrieb Dave Love: > Erik Soyez <[email protected]> writes: > >> Good day users, >> >> does anybody know if this bug >> >> http://www.mentby.com/Group/grid-engine/wildcarded-pe-name-circumvents-queue-sorting.html >> >> has been fixed in any son of gridengine version? > > I don't think there have been any relevant changes. The only > possibility I can think of is the one which fixed the lack of sorting by > load value for parallel jobs. I can't easily test it, though I thought > seqno sorting was basically working here. There doesn't seem to be such > a fix in the Oracle 6.2u6 list either. > > I wonder if it's really even a bug. It's not clear to me what should > happen generally in selecting PEs with various types of preference and > with reservation in effect <https://arc.liv.ac.uk/trac/SGE/ticket/1354>. > Has anyone looked closely at what's happening? > > For what it's worth, explicitly sorting PEs might be useful (and fairly > clear) in such cases <https://arc.liv.ac.uk/trac/SGE/ticket/1425>. It > should be straightforward to implement.
This is not what I had in mind when I suggested: https://arc.liv.ac.uk/trac/SGE/ticket/206 (well, some time ago...) The admin should define the pe_seqno, not just "descending" of their name or alike. Andy started already on a specification, but I think it's lost in the old mailing list. The problem I see with using the name to sort: for one type of PE you would like it ascending, for another descending andd hence you have to assemble it in the names. I didn't check the relevant code, but it should be just one entry in addition in the PE. If the argument is to reflect the sort order in the name, you could say the same for the queues. As for queue instances it could also be defined on a host(-group) level, but it would be much work to introduce it also for PE-instances. Instead the entry in the PE could be a number, or the special entries QUEUE to inherit it from the queue (and stay in this queue), or HOSTGROUP to stay in one hostgroup in addition to inherit the seqno (and the seqno of the hostgroup is already defined on the queue level). Maybe this way the wildcard specification to get always machine from a single rack could be eliminated too: you need only one PE and specify HOSTGROUP therein. And: if I get it right, using the name for it still doesn't allow to fill the cluster from the one side with serial jobs, from the other side with parallel ones in an easy way. You would need one PE for host(n), one for host(n) & host(n-1), one for host(n) & host(n-1) & host(n-2) until enough slots can be collected. -- Reuti > Changes, analysis, or comments welcome... > > -- > Community Grid Engine: http://arc.liv.ac.uk/SGE/ > _______________________________________________ > users mailing list > [email protected] > https://gridengine.org/mailman/listinfo/users _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
