We've noticed this too, but cgroups will help since they can limit on actual memory used rather than virtual address space.
On Tue, Aug 18, 2015 at 09:14:28PM +0000, Brendan Moloney wrote: > Hi, > > I have a small cluster so exclusive node access isn't really an option. We > enforce memory constraints on jobs using h_vmem. I discovered that more > recent versions of CUDA will always allocate a huge amount of virtual memory > (RAM + SWAP + GPU_RAM) so that it can supply a "Unified Virtual Address" > (UVA) space. > > Is there some way to get around this? Would cgroups support help? > > Thanks, > Brendan > _______________________________________________ > users mailing list > [email protected] > https://gridengine.org/mailman/listinfo/users -- -- Skylar Thompson ([email protected]) -- Genome Sciences Department, System Administrator -- Foege Building S046, (206)-685-7354 -- University of Washington School of Medicine _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
