In multithreaded runs, ScaLAPACK calls to pzhevgx are becoming a significant non-threaded bottleneck. At least the Intel mkl implementation of ScaLAPACK does not gain performance by adding threads to mkl calls. If we want to eliminate this bottleneck, we would have to look for alternatives to pzhevgx.
In multithreaded runs, ScaLAPACK calls to
pzhevgx
are becoming a significant non-threaded bottleneck. At least the Intel mkl implementation of ScaLAPACK does not gain performance by adding threads to mkl calls. If we want to eliminate this bottleneck, we would have to look for alternatives topzhevgx
.List of ideas: