On Fri, Apr 5, 2019 at 10:02 PM Aubrey Li <aubrey.li@xxxxxxxxxxxxxxx> wrote: > AVX-512 components use could cause core turbo frequency drop. So > it's useful to expose AVX-512 usage elapsed time as a heuristic hint > for the user space job scheduler to cluster the AVX-512 using tasks > together. > > Tensorflow example: > $ while [ 1 ]; do cat /proc/pid/status | grep AVX; sleep 1; done > AVX512_elapsed_ms: 4 > AVX512_elapsed_ms: 8 > AVX512_elapsed_ms: 4 > > This means that 4 milliseconds have elapsed since the AVX512 usage > of tensorflow task was detected when the task was scheduled out. > > Or: > $ cat /proc/pid/status | grep AVX512_elapsed_ms > AVX512_elapsed_ms: -1 (Very nitpicky, feel free to ignore: If you change the /proc/pid to /proc/tid in the commit message, it becomes clearer that this status is really per-task/thread, not per-process/threadgroup.) [...] > + > +/* > + * Report the amount of time elapsed in millisecond since last AVX512 > + * use in the task. > + */ > +static void avx512_status(struct seq_file *m, struct task_struct *task) > +{ > + unsigned long timestamp = task->thread.fpu.avx512_timestamp; This is theoretically a data race, right? Should this have a READ_ONCE() on it? Is there something that zeroes out the avx512_timestamp on fork()/clone(), or will every child inherit the avx512 timestamp? As far as I can tell, the timestamp is inherited; I think it would be nicer to zero it out at that point. Either way, It might be worth documenting this decision.