Community
cancel
Showing results for 
Search instead for 
Did you mean: 
may_ka
Beginner
53 Views

pardiso performance drop when decreasing the number of right hand sides

Hi there

I struggle a bit to understand (circumvent if possible) the pardiso performance drop measured as "seconds per right hand side" when decreasing the number of right hand sides.

The sparse matrix feed into pardiso has 2,562,698 rows/columns and the factorization yields 168,330,156 non-zero elements (iparm(18)).

After calling pardiso with phase=12, it is called with phase=33. The iparm vector was left with default values. When using 10 r-h-s, the solution was obtained after 2.21 seconds (phase 33 only), when using 750 r-h-s, the solution was obtained after 16.7 seconds (phase 33 only). Thus, in the first case pardiso needed 0.221 seconds per r-h-s, in the second case 0.022 seconds per r-h-s, which is a decrease of almost factor 10.

The timing was obtained on a Intel(R) Xeon(R) CPU E5-2697 v4 @ 2.30GHz with 36 real cores and 256GB ram. The number of MKL cores was 36.

The iparm vector for the 10 r-h-s case was:

1,2,36,0,0,0,0,0,0,0,0,0,0,0,5732781,4282495,8583732,168330156,1014847,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,5732781,5153806,885183,0

for the 750 r-h-s case:

1,2,36,0,0,0,0,0,0,0,0,0,0,0,5732781,4282495,541945253,168330156,1014847,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,5732781,5153806,885183,0

 

Any suggestion about how I can achieve a time per r-h-s for the 10 r-h-s case similar to that for the 750 r-h-s case are highly appreciated.

Thanks.

 

0 Kudos
5 Replies
Alice_H_Intel
Employee
53 Views

Hello,

Thanks for your question. I'll investigate it and get back to you soon.

Thanks,

Alice

may_ka
Beginner
53 Views

Hi there,

did you find out anything yet?

Thanks.

Gennady_F_Intel
Moderator
53 Views

this is the current implementation. In the case if You require to see the improvements, please submit the request to the intel online service center. 

may_ka
Beginner
53 Views

Hi, does that mean there are some improvements underway?? If so what will be the mkl version shipping it?

Thanks

Kirill_V_Intel
Employee
53 Views

Hi,

I've got an idea (unfortunately, not yet the out-of-box pardiso improvement). Can you try to set iparm[24] = 2 and tell us if you see any improvement? This option would allow pardiso to parallelize over matrix, not over the righthand sides.

I'd expect this should help improve the performance.

Thanks,
Kirill