Thank you Carles and Moe, I had FastSchedule to 0, and It worked also. Now I switched to FastSchedule=1 and 2 and it works ok.
I use task affinity and also ok. I was using FastSchedule to 0 because after an upgrade to Slurm 2.3.3, the real memory reported by the daemons was different and all my nodes went down. I don't know why. When you have FastSchedule to 1, the cores, sockets and cpus are also sent from the nodes? Regards 2012/2/13 Moe Jette <je...@schedmd.com> > Felip, > > That will work if you configure FastSchedule=1 or 2. If you also use task > affinity, that should work, but I would recommend testing it. > > Moe Jette > SchedMD > > > Quoting Felip Moll <lip...@gmail.com>: > > Hello all, >> >> I am wondering if I can set the number of cpus to a lower number than the >> real number of cpus. I want to do this to limit the number of concurrent >> cores accessing memory. >> >> Is it correct to set something like this on slurm.conf?: >> >> NodeName=pez001 RealMemory=32111 Sockets=2 CoresPerSocket=2 >> ThreadsPerCore=1 >> >> By this way, the node pez001 will display 4 cores, but it has 8 physical >> cores. >> >> >> I am asking it because if you set the memory to a level different than the >> real value, the node becames down. >> >> >> Thank you a lot. >> >> > > >
