We find that some CPU cores running OMP code are saturated ever since we migrated to Studio 2017 Update 3. The code involving
kmp_set_blocktime(0), "omp parallel sections", "omp parallel for"
was running fine with Studio 2016.
With 2017, calling
kmp_set_library_throughput(); kmp_set_blocktime(0)
does not avoid the saturation. However calling
kmp_set_defaults("KMP_BLOCKTIME=0")
or
kmp_set_defaults("OMP_WAIT_POLICY=PASSIVE")
avoids it.
Can anyone explain what is going wrong? Is this a bug in 2017?