PARDISO's reordering step looks very slow. Here's some statistics produced by PARDISO itself:
================ PARDISO: solving a real nonsymmetric system ================ Summary PARDISO: ( reorder to factorize ) ================ Times: ====== Time fulladj: 0.011505 s Time reorder: 47.586437 s Time symbfct: 0.035378 s Time A to LU: 0.000000 s Time numfct : 0.033240 s Time malloc : 0.032921 s Time total : 47.733567 s total - sum: 0.034086 s Statistics: =========== < Parallel Direct Factorization with #processors: > 1 < Numerical Factorization with Level-3 BLAS performance > < Linear system Ax = b> #equations: 188224 #non-zeros in A: 215330 non-zeros in A (%): 0.000608 #right-hand sides: 0 < Factors L and U > #columns for each panel: 128 #independent subgraphs: 0 < Preprocessing with state of the art partitioning metis> #supernodes: 178361 size of largest supernode: 5 number of nonzeros in L 212862 number of nonzeros in U 4320 number of nonzeros in L+U 217182 gflop for the numerical factorization: 0.000066 gflop/s for the numerical factorization: 0.002000
I run eight instances of PARDISO simultaneously on a dual-socket Xeon E5530 machine. Each of them works on a sparse banded matrix (bandwidth=61) similar to what produced the above. Is it normal that METIS takes nearly 50 secs to finish?