Effectively tuning PLASMA

[dobson156]

Hi all

The PLASMA user guide for 2.4.5 directs (6.3 Tuning Howto) you to LAWN #217 (Comparative Study of One-Sided Factorizations with Multiple Software Packages on Multi-Core Hardware) for information on tuning PLASMA.

Having read the paper section 3.1 (Tuning.PLASMA) details the tuning process, my understanding is as follows:

• Iterate though values of NB (40...500) and IB (factors or NB)
• run sequential core_blas routines used in factorisations (with N=NB) with the select IB and NB values
• Select the best performing combinations (from across the NB range)
• Run full parallel factorisations using IB and NB values from the selection
• Select best performing combination as "best performing" for architecture @ problem size.

So my first question, is my understanding of the pruned tuning process correct?

My second is the graphs in that paper show that the core_blas routines used by the factorisation (DPORTF - dgemm-seq, DGEQRF - dssrfb-seq, DETRF - dsssm-seq) seem to have changed since the paper was written (dssrfb isn't even included with my version of plasma) which routines should I tune against instead.

Many Thanks

[admin]

I think that you understand it correctly.
Yes, naming changed - young project evolved, sorry for that.
For QR, LQ or anything that uses Householder reflectors, you need to tune xTSMQR.
Jakub

[dobson156]

Thanks for the reply.

The provided timing programs appear to act on top level `PLASMA_` functions. For the step which involves, running pairs of IB and NB against `core_blas` routines (so N=nb) should I either write my own timing programs calling the core_blas or call the closest top level routine setting the --nb and --n to be the same

I guess my question boils down too will PLASMA_dgemm, ans similar, with N=NB only make one Core_dgemm call? (and there fore in essence have the same performance).

Thanks again