massive threads slow down

[kaji331]

In my AMD Zen5 9755 server, no SMT, there are 256 cores from 2 CPUs. I tried bwa 0.7.17/0.7.19 with GCC8/11/15, 8/16/32/64/128/256 threads for bwa mem (-t parameter). 8/16/32/64/128 threads give me a reasonable speed, but 256 threads are very very slow, 21 seconds for 128 threads vs 19938 seconds for 256 threads. I also find the bwa process is S status in top running only ~100% CPU for 256 threads. All tests run on nvme solid storage.

[lh3]

I don't know. You may try tcmalloc or mimalloc.

[kaji331]

I don't know. You may try tcmalloc or mimalloc.

Thanks. I tried mimalloc for bwa 0.7.19 with GCC 15.2, same problem... T_T

[kaji331]

I tried again, -t 130/192/200/201/210/212/213/214 threads, the CPU/MEM of bwa could reach the setting values in top. But 215/220/225/250 threads only kept about 100-105% CPU and ~5G MEM.

[lh3]

Oh, this should be fixed by #155