Hello there:
I have found an interesting appearance which I can not explain. Okey, let's go.
I apply "micsmc" to surveiling the offload running state of MIC. The critical code like this:
#pragma offload target(mic:0) inout(XXXX) in (XXXX) { #pragma omp parallel for schedule (dynamic) for( int i = 0; i < num_cluster; i++) // num_cluster from 60 to 300,concentrated on 90~150 { do something.... } }
And then set the environment variables :
export OMP_NUM_THREADS=X
export KMP_AFFINITY=compact
when X = 4
The percent of the first core's user is between 80 to 90.
when X =7
The percent of the first core's user is between 80 to 85.
when X=8
The percent of the first core's user is between 80 to 85.
when X=240
The percent of the first core's user is between 70 to 80.
when X= 241
All core's users are so bad! Once we setting our thread number beyond 240!
Why are the percent of the first core user always below 100%?