PARDISO Scalability

Paweł_J_ — Wed, 13 Aug 2014 09:08:47 GMT

Hi,

I have a question about scalability of the PARDISO solver. I'm using Intel MKL PARDISO with my Finite Element code. I tested simple linear elastic problem with around 800 000 unknowns. The time results are:

threads	pardiso phase 22	pardiso phase 33	sum pardiso	speed up pardiso phase 22	speedup pardiso phase 33	speed up sum pardiso
1	1433.922	18.8455	1452.7675	1	1	1
2	720.3288	10.3364	730.6652	1.9906492702	1.8232169808	1.988280679
4	381.09201	7.709	388.80101	3.7626661341	2.4446101959	3.7365322173
6	285.75699	8.2753	294.03229	5.0179769881	2.2773192513	4.9408434019
8	234.9064	7.7265	242.6329	6.1042270453	2.4390733191	5.9875124107
10	225.3418	9.33	233.6718	6.3633200764	2.0198821008	6.2171280403
12	205.16	7.4491	212.6091	6.9892864106	2.5299029413	6.8330447756

I've tested it on a node with 2 processors with 6 cores Xeon X5650 with 2,66Ghz and 24 GB RAM DDR3 1333MHz .

I'm wondering if speed up around 7 is ok or I may get better speed up, when I play a little bit with input parameters for PARDISO?

Is there some articles or papers about scalability of the PARDISO solver? Could you pass me some links to them?

My input parameters are:

     iparm(1) = 1
      iparm(3) = 0
      iparm(4) = 0
      iparm(5) = 0
      iparm(6) = 0

      iparm(7) = 0
      iparm(8) =10
      iparm(9) = 0
      iparm(10) = 8
      iparm(11) = 1
      iparm(12) = 0

      iparm(13) = 1
      iparm(14) = 0
      iparm(15) = 0
      iparm(16) = 0
      iparm(17) = 0
      iparm(18) = -1
      iparm(19) = 0
      iparm(20) = 0
      iparm(21) = 1
      iparm(22) = 0
      iparm(23) = 0
   iparm(24) = 0
      iparm(25) = 0
      iparm(27) = 0
      iparm(28) = 0
   iparm(30) = 0
   iparm(31) = 0
   iparm(35) = 0
      iparm(60) =0

Thanks for all advices.

best regards,

Pawel J.

Sorry my table was unreadable

Paweł_J_ — Wed, 13 Aug 2014 09:20:00 GMT

Sorry my table was unreadable. I've attached screenshot from Calc.

best regards,
Pawel J.

Pawel,

VipinKumar_E_Intel — Thu, 14 Aug 2014 04:36:22 GMT

Pawel,

The speedup of around 7 seems to be fine. Since the matrix size is large, you may try the Cluster PARDISO from MKL. Please find details here.

https://software.intel.com/en-us/articles/intel-math-kernel-library-parallel-direct-sparse-solver-for-clusters

--Vipin

topic PARDISO Scalability in Intel® oneAPI Math Kernel Library

PARDISO Scalability

Sorry my table was unreadable

Pawel,