Am 07.06.2011 um 15:22 schrieb Dave Love: > Reuti <[email protected]> writes: > >>> Apologies for the confusion. However, that doesn't deaql with >>> multi-level sorting on seqno and load. In an inhomogeneous cluster, >>> you typically want to sort by seqno (node type) first. At >>> least in that case, the second-level packing using slots for the load >>> doesn't work for us in the serial case either. There's a >>> one-liner fix for the parallel case in the Univa repo, but I don't know >>> what's the problem in the serial case. >> >> To use "seqno" and as a second sort criteria "load" should work >> automatically. What behavior do you notice instead: >> >> - only sorted by "slots"? >> >> - only sorted by "seqno" ignoring the load? (Is it the same with the >> load_formula being np_load_avg or only with the consumable complex?) > > Seqno seems to work, but with slots for the load, jobs seem to end up on > random hosts. I don't know about with np_load_avg, though I guess its' > worth trying inverting that to pefer most loaded (which won't work too > wel with poorly-behaved jobs)
It would be interesting whether it's a general problem when using consumables instead of a real load. There was an issue about using consumables as thresholds, but that was fixed some time ago already. -- Reuti _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
