Hi all
We use the PARDISO solver (Parallel Studio 2018) in our FEM code for a symmetric indefinite system and discovered that the time spent in the factorization step may vary significantly. We use the following solver parameters, all other values are set to zero:
iparm [0] = 1;
iparm [1] = 2;
iparm [2] = 4;
iparm [9] = 8;
The solver output of a slow and fast example can be found below. The total time differs almost by a factor of 2 despite similar size of the equation system. This makes PARDISO clearly less attractive than our iterative GPU solver in certain cases. Is this a performance issue/bug? Do we use bad settings? Or is it just normal behavior? Any help is highly appreciated.
Thank you very much and best regards
David
Slow example:
Summary: ( starting phase is reordering, ending phase is solution )
================
Times:
======
Time spent in calculations of symmetric matrix portrait (fulladj): 0.703717 s
Time spent in reordering of the initial matrix (reorder) : 11.242095 s
Time spent in symbolic factorization (symbfct) : 3.675149 s
Time spent in data preparations for factorization (parlist) : 0.107626 s
Time spent in copying matrix to internal data structure (A to LU): 0.000000 s
Time spent in factorization step (numfct) : 133.596461 s
Time spent in direct solver at solve step (solve) : 1.954850 s
Time spent in allocation of internal data structures (malloc) : 0.250255 s
Time spent in additional calculations : 4.518999 s
Total time spent : 156.049152 s
Statistics:
===========
Parallel Direct Factorization is running on 6 OpenMP
< Linear system Ax = b >
number of equations: 2956396
number of non-zeros in A: 79437854
number of non-zeros in A (%): 0.000909
number of right-hand sides: 1
< Factors L and U >
number of columns for each panel: 96
number of independent subgraphs: 0
< Preprocessing with state of the art partitioning metis>
number of supernodes: 338571
size of largest supernode: 17752
number of non-zeros in L: 3101919901
number of non-zeros in U: 1
number of non-zeros in L+U: 3101919902
gflop for the numerical factorization: 20373.427694
gflop/s for the numerical factorization: 152.499756
Fast example:
Summary: ( starting phase is reordering, ending phase is solution )
================
Times:
======
Time spent in calculations of symmetric matrix portrait (fulladj): 0.462895 s
Time spent in reordering of the initial matrix (reorder) : 10.686642 s
Time spent in symbolic factorization (symbfct) : 3.656114 s
Time spent in data preparations for factorization (parlist) : 0.108843 s
Time spent in copying matrix to internal data structure (A to LU): 0.000000 s
Time spent in factorization step (numfct) : 64.537849 s
Time spent in direct solver at solve step (solve) : 1.397830 s
Time spent in allocation of internal data structures (malloc) : 0.240506 s
Time spent in additional calculations : 4.692960 s
Total time spent : 85.783639 s
Statistics:
===========
Parallel Direct Factorization is running on 6 OpenMP
< Linear system Ax = b >
number of equations: 2961080
number of non-zeros in A: 83916044
number of non-zeros in A (%): 0.000957
number of right-hand sides: 1
< Factors L and U >
number of columns for each panel: 96
number of independent subgraphs: 0
< Preprocessing with state of the art partitioning metis>
number of supernodes: 308944
size of largest supernode: 6004
number of non-zeros in L: 2958633165
number of non-zeros in U: 1
number of non-zeros in L+U: 2958633166
gflop for the numerical factorization: 8909.331438
gflop/s for the numerical factorization: 138.048162