topic Threaded version of paradiso OOC in Intel® oneAPI Math Kernel Library

Threaded version of paradiso OOC

Laurence_Page — Fri, 23 Jan 2009 12:28:49 GMT

Is there any work being done/planned to enable threads with the paradiso Out of Core solver?

Re: Threaded version of paradiso OOC

Gennady_F_Intel — Fri, 23 Jan 2009 14:04:56 GMT

Yes, we are working on this problem right now but I cannot provide exact data when this version will be available for external customer.
--Gennady

Re: Threaded version of paradiso OOC

Sergey_K_Intel1 — Mon, 26 Jan 2009 09:41:10 GMT

In fact, out of core PARDISO fromMKL 10.1 Goldreleased recentlyuses parallelism available through MKLBLAS and MKL LAPACK.It should also be noted that starting from the release MKL DSS/PARDISO only uses the number of threads set with the help of any available Intel MKL or OpenMP threading controls (e.g. OMP_NUM_THREADS, MKL_NUM_THREADS environment variables and etc). The value of iparm(3) is ignored.

Gennady meant out-of-core version with parallelism at PARDISO level.

All the best
Sergey

Re: Threaded version of paradiso OOC

Laurence_Page — Wed, 28 Jan 2009 09:11:22 GMT

Quoting - Sergey Kuznetsov (Intel)

In fact, out of core PARDISO from MKL 10.1 Gold released recently uses parallelism available through MKL BLAS and MKL LAPACK. It should also be noted that starting from the release MKL DSS/PARDISO only uses the number of threads set with the help of any available Intel MKL or OpenMP threading controls (e.g. OMP_NUM_THREADS, MKL_NUM_THREADS environment variables and etc). The value of iparm(3) is ignored.

Gennady meant out-of-core version with parallelism at PARDISO level.

All the best
Sergey

I'm using 10.1.019, should that contain the parallel OOC version of paradiso?

It reports that it is using the requested (via MKL_NUM_THREADS) number of threads if I run In-Core. However, by simply changing iparm(60) from 0 to 2 (to force it to run out-of -core) it always reports that it is running on a single thread. There doesn't appear to be any parallel activity from the BLAS-3 level routines. The manual says (not very loudly) that paradiso OOC will only use the one thread.

Would you expect the BLAS/LAPACK threads to make a significant speed improvement?

Re: Threaded version of paradiso OOC

Sergey_K_Intel1 — Thu, 29 Jan 2009 07:13:40 GMT

Quoting - Laurence Page

Dear Laurence,

Yes, 10.1.019 should contain the parallel OOC version of PARDISO. And PARDISO reportsthe number of threads used by PARDISO only but it allows BLAS and LAPACK to use the number of threads set through MKL threading controls like MKL_NUM_THREADS environment variable.

There are no doubts that the high level parallelism at PARDISO level shouldhave better performance characteristicscompared to the usage of parallelism at BLAS and LAPACK level.

Ufortunately, speed imrovement for the PARDISO OOC versionwhich usesBLAS and LAPACK parallelism depends on many factors:OS, hardware, matrix structure, value of MKL_MAX_OOC_CORE_SIZE and etc. I'd recommend to set MKL_MAX_CORE_SIZE close to the size of RAM available on your computer.

All the best
Sergey