First, thank Alexander Kalinkin for helping me for threading issue that I posted.
Now I have another issue in my sparse matrix solver benchmark. In my benchmark, PARDISO is somehow very slow. My problem is obtained from high order FEM. Thus, it contains some large dense blocks inside. While other multifrontal solvers work fairly well, PARDISO does not well behave for such matrices including dense blocks compared to multifrontal solvers. I guess that this may be due to left-looking algorithm in PARDISO (for dense matrices, right looking algorithm is mostly beneficial from a large bulk of GEMM updates). However, this argument is not still convincing. Does anyone have similar experience that PARDISO is relatively slower than others for specific numerical methods ?