Hi, Yes, cpu == core. The legacy naming is somewhat unfortunate. I'll update the page to clarify.
The 1 slot - 1 core approach does not work for us because: * each task may use more than one core (e.g., openmp, multithread) * specifying ncpus allows for grouping that better matches the application needs (e.g., geometry, grid size) * we want the user to be able to select based on various tradeoffs Our job requests take the form of: #$ -l pe dev <nslots> #$ -l ncpus=<count>,mem=<Mbytes>,tmpfs=<Mbytes> This way each slot gets a specific amount of cpus, memory, and dedicated tmpfs. Each of the slots is supported by an LXC container with its own contstraints, networking, etc. Where we do the above, the cpus are dedicated. So, there is no overallocation. The cpus are either available or not. John On Fri, 2018-08-31 at 12:58 +0200, Reuti wrote: Hi John, Am 31.08.2018 um 12:27 schrieb Marshall2, John (SSC/SPC) <john.marsha...@canada.ca<mailto:john.marsha...@canada.ca>>: Hi, When gridengine calculates cpu usage (based on wallclock) it uses: cpu usage = wallclock * nslots This does not account for the number of cpus that may be used for each slot, which is problematic. What was the motivation for implementing it this way? I mean: usually one SLOT represents one CORE in GirdEngine. Hence, to get a proper allocation and accounting while not oversubscribe the nodes you have to request the overall amount of cores in case you want to combine processes (like for MPI) and threads (like for Open MP). In your case, it looks to me that you assume that necessary cores are available, independent from the actual usage of each node? -- Reuti PS: I assume with CPUS you refer to CORES. I have written up an article at: https://expl.info/display/MISC/Slot+Multiplier+for+Calculating+CPU+Usage+in+Gridengine which explains the issue and provides a patch (against sge-8.1.9) so that: cpu usage = wallclock * nslots * ncpus_per_slot This makes the usage information much more useful/accurate when using the fair share. Have others encountered this issue? Feedback is welcome. Thanks, John _______________________________________________ users mailing list users@gridengine.org<mailto:users@gridengine.org> https://gridengine.org/mailman/listinfo/users
_______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
