> Am 16.11.2015 um 14:14 schrieb Winkler, Ursula (ursula.wink...@uni-graz.at) > <ursula.wink...@uni-graz.at>: > >>> Am 16.11.2015 um 11:41 schrieb Winkler, Ursula (ursula.wink...@uni-graz.at) >>> <ursula.wink...@uni-graz.at>: >>> >>> Dear gridengine members, >>> >>> when there are two queues and one is a subordinate of the other, then >>> the release/shift of slots to the superordinate jobs is no problem, >>> but (other) consumeable resources like h_vmem are. Does somebody have an >>> idea to bypass that? >> >> As the memory is still occupied by the suspended job, I'm not sure whether >> it's wise to bypass it. Even in case >> the job starts, it may run out of memory at one time. >> >> Although I usually use only 2 GB of swap nowadays, you could try to have >> swap as large as the built-in memory. >> The result should be, that any suspended job can be swapped out completely. >> >> -- Reuti > > So long I just lowered the default value of h_vmem (assigned when it isn't > ordered explicit with qsub). > That works in some limited way because many users just overestimate the true > memory demand of their jobs. > > But the idea with swapping out sounds at least worth to be considered in some > cases. How do I implement it within the > gridengine?
In case the swap space on disk has e.g. 64 GB as the real memory is 64 GB, one could set h_vmem to an overall size of 128 GB. Per queue it would still be necessary to limit it with an RQS to 64 GB. -- Reuti _______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
