Closed noobpwnftw closed 6 years ago
Also including transform in reduction. These procedures are generally less time consuming, usually 1-5 minutes after limiting threads.
When called with a NUMA work list, transform should be OK on many threads. I will try to make the compression NUMA aware where easily possible and optionally allow lowering the number of threads for all other cases.
Maybe because they are too fast on the CPU side.
For example, the fix_closs_worker runs for 6 seconds on 64 threads, 5 minutes on 384 threads, NUMA aware. I think with the new Xeon generation, the models with "M" are for their worth.
6 seconds versus 5 minutes? oops...
Yup, when there is congestion, the backfire is really annoying.
I did some performance measurements and these are verified that less threads run faster, without -d option.