dsyev eigenvalue solver

[nchilton]

Right....I've got pretty limited experience with c/c++, so would it be easiest to install the netlib lapacke and use that? Can I use the netlib lapacke in conjunction with the mkl blas, or do I need to use the entire netlib lapack/lapacke/blas suite solely?

Nick

[mateo70]

Hello Nick,

No problem to use lapacke from Netlib in conjunction of MKL BLAS. That's what the plasma installer is doing when you configure it with MKL. To be sure to use the one from netlib and not the functions included in MKL, you have to make sure the -llapacke is placed just before the list of -lmkl_... that you need.

Mathieu

[nchilton]

Mathieu,

Fantastic, compilation achieved!

Now to debug! First question - the PLASMA_dsyev routine needs integers for jobz and uplo, compared to character for lapack/mkl etc. What are the integer versions for 'N' or 'V' and 'U' or 'L', respectively?

Cheers!
Nick

[nchilton]

Anyone?

[mateo70]

Hello,

sorry for the late answer, I havn't seen the post before.
You have all those infomations in PLASMA headers files in include directory.
In your case, it will be PlasmaNoVec/PlasmaVec and PlasmaUpper/PlasmaLower

Mathieu

[nchilton]

Mathieu, thank you very much, all sorted! PLASMA is certainly quick, I'm getting good speed-ups! It seems irrespective of my matrix size, DYSEV works optimally with 8 cores. Any ideas why?

Cheers!
Nick

[haidar]

Mathieu, thank you very much, all sorted! PLASMA is certainly quick, I'm getting good speed-ups! It seems irrespective of my matrix size,
DYSEV works optimally with 8 cores. Any ideas why?
Cheers!
Nick

Hi Nick,
could you specify what did you mean by irrespective?
It should scale up to 96 cores that was what we already test it. However your matrix need to be large enough.
what is your matrix size? and what kind of system are you running on?
Thanks
Azzam

[nchilton]

Hi Azzam,

I was timing the diagonalization of various matrices ranging in size from 2000 to 10,000 square. All of the results showed essentially flat results (only very slight increases) past 8 cores. The machine has 2 Quad-core Intel Xeon L5520 @ 2.27 GHz with a total of 48 GB DDR3 @ 1066 MHz and now that I think about it, it might be due to hyper threading... Do you use MPI to access large numbers of cores?

Cheers,
Nick

[admin]

PLASMA only uses Pthreads.
No MPI.
If your machine is a dual quad-core, then you only have eight physical cores.
You will not get any speedup from using more threads than physical cores.
In fact I am surprised that you are not seeing a slowdown.
Jakub

[mateo70]

Hello Nick,

Can you post the result of the following command on your system, please ?
pkg-config --libs hwloc

Tio check that your threads are correctly binded, and also what value to you give to PLASMA_Init ?
If you give something bigger than 8, your threads won't be binded to the cores and the results shouldn't be as good as with only 8 threads as mentionned by Jakub.

Mathieu