Am 21.06.2012 um 15:12 schrieb Mazouzi: > On Thu, Jun 21, 2012 at 2:09 PM, Reuti <re...@staff.uni-marburg.de> wrote: > Nevertheless, what I remember is that using MPI you need per node one data > server plus the the computing instances. I.e. in principle you are starting > more processes than granted slots, as some processes are used for > communication only. > > I don't know if this the best way to do, we managed this as follow : > > 1 - user need to request a pair of slots (2*n) > 2- modify the script so that it don't multiply n by 2 > 3- gamess needs two process by node as mentioned by reuti, so we created a > new pe with allocation rules=2 > > The ideal is to have pe that ensure a pair of process on a given host. But I > don't know how to define the allocation rules for that.
To get multiple of a fixed amount of slots per exechost is an RFE but not implemented yet: https://arc.liv.ac.uk/trac/SGE/ticket/444 Although it's some time old, it never made it into any release. In Torque one can define whether ppn=2 allows to be taken multiple times from an exechost or not. This is not possible in SGE though. In the past with only 2 or 4 cores in machine this wasn't much of an issue, but with 64 cores you get multiple PEs if you want to cover all multiple of N possibilities per exechost. -- Reuti _______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
