pardiso out of memory (-2) in phase 33

[may_ka]

Hi,

pardiso stops with error message -2 in phase 33 when increasing the number of rhs from 600 to 700. 600 rhs runs fine with a memory use  of 146.2GB ("RES") reported by "top". The system has 256GB of RAM, so there is still +40% left. What is the issue when increasing #rhs by 100??

Thanks

[Gennady_F_Intel]

This may happen because of Pardiso consumes too many memory in the case of rhs.  How many threads do you use? what is the problem size?  Please check what iparm[17] returns in the case of 500 and 600 threads for example.

[may_ka]

Hi

first it is 700 to 800 rhs, not 500 to 600 ............. my fault.

iparm with 700 rhs is:

1,2,36,0,0,0,0,0,0,0,0,0,0,0,5321589,3979573,477759815,126792401,612333,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,5321589,4845387,877558,0

iparm with 800 rhs is:

1,2,36,0,0,0,0,0,0,0,0,0,0,0,5321589,3979573,1727119,126792401,612333,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1,0,0,0,5321589,4845387,877558,0

so iparm[17] for 700 is 477,759,815, where iparm[17] for 800 is 1,727,119

That all doesn't make sense to me especially since the manual states that iparm[17] is calulated during phase 1.

Further if that is the memory usage in kbyte then 700 rhs was at 477,759,815*1000/10^9 GByte which is 477.75 GB. But the computer has only 256GB and I am 100% certain that it never went into swap.

The factor contains 126,792,401 non-zero elements and has a row/column dimension of 2,421,403. I need to solve for ~50,000 rhs, which I supply in chunks of up to 700 rhs. Because from what we discussed here

I want to keep the number of RHSs as many as possible.

CPU is Intel(R) Xeon(R) CPU E5-2697 v4 @ 2.30GHz, so effectively 36 cores and threads.

Note that I don't store a rhs matrix of 2.4Miox50,000. The rhs are generated by a pre-mult step with a sparse rectangular matrix before they go to the solver.

Cheers

[Kirill_V_Intel]

Hello,

Can you specify the Intel MKL version which you are using? I believe we have improved the memory consumption for the multi-threaded execution of the Intel MKL Pardiso in MKL 2019u2.

Thanks,
Kirill

[may_ka]

Hi,

the posted results were obtained with 17.08. I'll checkout 19.02.

[Gennady_F_Intel]

some small correction - actually the fix of the problem will be available into the next ( nearest) version MKL 2019 update 3. the current ( latest) available update 2 doesn't contains such kind of fixes...