[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [HTCondor-users] how to announce/schedule heterogeneous CPUs? [hypothetical question]

On 11/18/20 6:43 AM, Thomas Hartmann wrote:
Hi all,

I.e., how could one schedule such a heterogeneous CPU with a big.little-like design, i.e., separate(?) the 'big cores' from the 'little cores' - assuming that with HTC the whole CPU w/should be utilized ~100% for most of the time? Binding jobs with cpuset to specific cores (no idea, if it is possible wih ARMs?) and have a node offer each core class as different slots?

This is an interesting question to ponder, but we'll need to wait until we see what the kernels can give us. These BIG-little ARM designs can be controlled by their kernels in several ways. One of which has the kernel just turn off either the big or little cores, so condor wouldn't have any input in the scheduling of jobs to cores. Also, assuming the little cores are 1/10th the speed of the big cores, it is unclear how many users would want to run on them, if they cause a "long-tail" problem. On systems that run both at once, maybe one idea would be to pin the condor daemons and the rest of the user-space system to the little cores, and reserve the big cores exclusively for batch jobs.