option to set processor affinity in addprocs #10645 Closed amitmurthy opened this Issue on Mar 26, 2015 · 7 comments Projects None yet Labels parallel Milestone No milestone Assignees No one assigned 4 participants @amitmurthy @ViralBShah @jakebolewski @ihnorton Notifications You’re not receiving notifications from this thread. @amitmurthy The Julia Language member amitmurthy commented on Mar 26, 2015 The ability to lock julia workers to specific cores results in substantial gains for certain workloads. See #10427 (comment) for some test results. addprocs should provide a means for the same. @ViralBShah The Julia Language member ViralBShah commented on Mar 27, 2015 I don't fully recollect, but there was some issue with openblas and affinity - we disable it by default. Do see the openblas build options, and the earlier discussion - I will try to find the relevant links. @amitmurthy The Julia Language member amitmurthy commented on Mar 27, 2015 With threaded openblas it may not make much sense. But in a parallel run, openblas threads are set to 1. @jakebolewski The Julia Language member jakebolewski commented on Mar 27, 2015 Wouldn't this be best handled by the cluster manager. SGE, Slurm, MPI, etc. all have options to do this. @amitmurthy The Julia Language member amitmurthy commented on Mar 27, 2015 What about folks running on largish machines like julia.mit.edu? I think the built in managers, LocalManager and SSHManager should support it. … @ViralBShah The Julia Language member ViralBShah commented on Mar 28, 2015 That is probably the most common use case. @amitmurthy The Julia Language member amitmurthy commented on Mar 28, 2015 Another example - https://groups.google.com/d/msg/julia-dev/BYVCyUlNR8c/_9XGgjolNNIJ - a 192 core system. @ihnorton ihnorton added the parallel label on Mar 28, 2015 @amitmurthy amitmurthy referenced this issue on Apr 3, 2015 Closed support basic core affinities in LocalManager and SSHManager #10733 @amitmurthy The Julia Language member amitmurthy commented on Apr 30, 2015 Have added this feature to ClusterManagers.jl - JuliaParallel/ClusterManagers.jl@9afe252 @amitmurthy amitmurthy closed this on Apr 30, 2015