PARDISO performance is not consistent

xian-zhong_guous_cd- — Thu, 24 Apr 2014 00:24:32 GMT

I have a SPD matrix (see below for detail). Running 8 threads and 16G memory, factorization sometimes takes 314 s (e.g. Run 1) and sometime 36 s (e.g. Run 2) on same machine (see below for hardware detail). Any idea what's going on?

Matrix has been attached:

line 1: # of equations

line 2: index base

line 3: # of nonzeros

followed by triplet format (i,j,x)

Run 1:

=== PARDISO: solving a symmetric positive definite system ===
MKLPARDISO::numericalFact starts Wed Apr 23 15:57:22 2014

The local (internal) PARDISO version is : 103911000
1-based array indexing is turned ON
PARDISO double precision computation is turned ON
METIS algorithm at reorder step is turned ON
Scaling is turned ON
Matching is turned ON

Summary: ( reordering phase )
================

Times:
======
Time spent in calculations of symmetric matrix portrait (fulladj): 0.510406 s
Time spent in reordering of the initial matrix (reorder) : 10.586863 s
Time spent in symbolic factorization (symbfct) : 4.795883 s
Time spent in data preparations for factorization (parlist) : 0.045271 s
Time spent in allocation of internal data structures (malloc) : 0.072302 s
Time spent in additional calculations : 3.339262 s
Total time spent : 19.349987 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 0

< Factors L and U >
number of columns for each panel: 192
number of independent subgraphs: 0
< Preprocessing with state of the art partitioning metis>
number of supernodes: 171522
size of largest supernode: 10602
number of non-zeros in L: 1336768222
number of non-zeros in U: 1
number of non-zeros in L+U: 1336768223
=== PARDISO is running in In-Core mode, because iparam(60)=1 and there is enough RAM for In-Core ===
Percentage of computed non-zeros for LL^T factorization
0 % 1 % 2 % 3 % 4 % 5 % 6 % 7 % 8 % 9 % 10 % 11 % 12 % 13 % 14 % 15 % 16 % 17 % 18 % 19 % 20 % 21 % 22 % 23 % 24 % 25 % 26 % 27 % 28 % 29 % 30 % 31 % 32 % 33 % 34 % 35 % 36 % 37 % 38 % 39 % 40 % 41 % 42 % 43 % 44 % 45 % 46 % 47 % 48 % 49 % 50 % 51 % 52 % 53 % 54 % 55 % 56 % 57 % 58 % 59 % 60 % 61 % 62 % 63 % 64 % 65 % 66 % 67 % 68 % 69 % 70 % 71 % 72 % 73 % 74 % 75 % 76 % 77 % 78 % 79 % 80 % 81 % 83 % 84 % 85 % 87 % 88 % 89 % 90 % 91 % 92 % 93 % 94 % 95 % 96 % 97 % 98 % 99 % 100 %
MKLPARDISO::solve starts Wed Apr 23 16:02:37 2014

=== PARDISO: solving a symmetric positive definite system ===
Single-level factorization algorithm is turned ON

Summary: ( factorization phase )
================

Times:
======
Time spent in copying matrix to internal data structure (A to LU): 0.000000 s
Time spent in factorization step (numfct) : 314.881940 s
Time spent in allocation of internal data structures (malloc) : 0.000198 s
Time spent in additional calculations : 0.000002 s
Total time spent : 314.882140 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 0

gflop/s for the numerical factorization: 14.910589

=== PARDISO: solving a symmetric positive definite system ===

MKLPARDISO::solve ends Wed Apr 23 16:02:49 2014

Summary: ( solution phase )
================

Times:
======
Time spent in direct solver at solve step (solve) : 4.046562 s
Time spent in additional calculations : 8.368093 s
Total time spent : 12.414655 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 1

gflop/s for the numerical factorization: 14.910589

Run 2:

=== PARDISO: solving a symmetric positive definite system ===
The local (internal) PARDISO version is : 103911000
MKLPARDISO::numericalFact starts Mon Apr 21 14:49:46 2014

1-based array indexing is turned ON
PARDISO double precision computation is turned ON
METIS algorithm at reorder step is turned ON
Scaling is turned ON
Matching is turned ON

Summary: ( reordering phase )
================

Times:
======
Time spent in calculations of symmetric matrix portrait (fulladj): 0.426923 s
Time spent in reordering of the initial matrix (reorder) : 10.688333 s
Time spent in symbolic factorization (symbfct) : 3.410566 s
Time spent in data preparations for factorization (parlist) : 0.046608 s
Time spent in allocation of internal data structures (malloc) : 0.072816 s
Time spent in additional calculations : 3.617951 s
Total time spent : 18.263197 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 0

< Factors L and U >
number of columns for each panel: 192
number of independent subgraphs: 0
< Preprocessing with state of the art partitioning metis>
number of supernodes: 171522
size of largest supernode: 10602
number of non-zeros in L: 1336768222
number of non-zeros in U: 1
number of non-zeros in L+U: 1336768223
=== PARDISO is running in In-Core mode, because iparam(60)=1 and there is enough RAM for In-Core ===
Percentage of computed non-zeros for LL^T factorization
0 % 1 % 2 % 3 % 4 % 5 % 6 % 7 % 8 % 9 % 10 % 11 % 12 % 13 % 14 % 15 % 16 % 17 % 18 % 19 % 20 % 21 % 22 % 23 % 24 % 25 % 26 % 27 % 28 % 29 % 30 % 31 % 32 % 33 % 34 % 35 % 36 % 37 % 38 % 39 % 40 % 41 % 42 % 43 % 44 % 45 % 46 % 47 % 48 % 49 % 50 % 51 % 52 % 53 % 54 % 55 % 56 % 57 % 58 % 59 % 60 % 61 % 62 % 63 % 64 % 65 % 66 % 67 % 68 % 69 % 70 % 71 % 72 % 73 % 74 % 75 % 76 % 77 % 78 % 80 % 81 % 82 % 83 % 85 % 86 % 87 % 88 % 89 % 91 % 92 % 93 % 94 % 95 % 96 % 97 % 98 % 99 % 100 %
MKLPARDISO::solve starts Mon Apr 21 14:50:23 2014

=== PARDISO: solving a symmetric positive definite system ===
Single-level factorization algorithm is turned ON

Summary: ( factorization phase )
================

Times:
======
Time spent in copying matrix to internal data structure (A to LU): 0.000001 s

Time spent in factorization step (numfct) : 36.840021 s
Time spent in allocation of internal data structures (malloc) : 0.000220 s
Time spent in additional calculations : 0.000002 s
Total time spent : 36.840244 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 0

gflop/s for the numerical factorization: 127.444969

=== PARDISO: solving a symmetric positive definite system ===
MKLPARDISO::solve ends Mon Apr 21 14:50:25 2014

Summary: ( solution phase )
================

Times:
======
Time spent in direct solver at solve step (solve) : 0.829546 s
Time spent in additional calculations : 1.730757 s
Total time spent : 2.560303 s

Statistics:
===========
< Parallel Direct Factorization with number of processors: > 8
< Numerical Factorization with BLAS3 and O(n) synchronization >

< Linear system Ax = b >
number of equations: 1576740
number of non-zeros in A: 56993250
number of non-zeros in A (%): 0.002292

number of right-hand sides: 1

< Factors L and U >
number of columns for each panel: 192
number of independent subgraphs: 0
< Preprocessing with state of the art partitioning metis>

number of supernodes: 171522
size of largest supernode: 10602
number of non-zeros in L: 1336768222
number of non-zeros in U: 1
number of non-zeros in L+U: 1336768223
gflop for the numerical factorization: 4695.075325

gflop/s for the numerical factorization: 127.444969

Hardware:

Operating System: Linux 2.6.32-431.3.1.el6.x86_64 (CentOS 6.3)

Average Load: 1.01 1.39 2.14 (average over last 1min, 5min & 15min)
CPU Type: Intel(R) Xeon(R) CPU E5-2650 v2 @ 2.60GHz (x86_64)
CPU Addressability: 64bit
CPU Count: 16 (8 cores/socket, Hyper-threading enabled)
CPU Clock: 1200 MHz
CPU Cache: 20480 KB (L2)
Physical Memory: 64367 MB
Swap Space: 48000 MB
Graphics device: NVIDIA Device 11fa (rev a1)
ls: cannot access /proc/ide: No such file or directory
SCSI CD 0,0,0,0: /dev/scd0 ()
SCSI Disk 0,0,0,0: /dev/sda ()
()
SCSI Disk 0,0,0,0: /dev/sda ()
()

thanks for the issue. We will

Gennady_F_Intel — Thu, 24 Apr 2014 13:11:14 GMT

thanks for the issue. We will check the issue on our side. Is that lp64 mode?

Yes.

xian-zhong_guous_cd- — Thu, 24 Apr 2014 16:19:15 GMT

Yes.

I tried the production

xian-zhong_guous_cd- — Thu, 04 Sep 2014 23:57:41 GMT

I tried the production version of mkl 11.2. Unfortunately, it still produces poor performance (see mkl11.2time.txt). Interestingly, mkl 11.2 beta produces good result (see mkl11.2betatime.txt).

Any progress on this issue?

xian-zhong_guous_cd- — Mon, 10 Nov 2014 23:13:21 GMT

Any progress on this issue?

topic Any progress on this issue? in Intel® oneAPI Math Kernel Library

PARDISO performance is not consistent

thanks for the issue. We will

Yes.

I tried the production

Any progress on this issue?