ScaLAPACK Archives

[Scalapack] question on Scalapack / Magma


Duncan,

We have a proof of concept for many of the routines in ScaLAPACK. 
We have a number of reports describing the approaches, e.g.,
1) In this paper we show that quick enhancements of ScaLAPACK to support GPUs 
are possible but
   more is needed to get top performance
   http://www.netlib.org/lapack/lawnspdf/lawn272.pdf

2) In this paper we used a runtime system and heterogeneous data splitting to 
show perfect scalability
   of one-sided factorizations for GPU-accelerated clusters. These results are 
on tiled algorithms
   http://www.netlib.org/lapack/lawnspdf/lawn250.pdf

3) We explored an approach on eigensolvers, directly based on ScaLAPACK, 
extended for latest two-stage
     approaches, and using parallel GPU BLAS from Cray (the libsci_acc library)
     (not published yet; can send a copy)

4) The approach in DPLASMA (http://icl.utk.edu/dplasma/) is based on tiled 
algorithms (similar to #2) and
     the PaRSEC runtime system, but doesn't have the heterogeneous data layout 
of (#2)

5) Recently we enhanced the PLASMA/DPLASMA approach by extending the QUARK 
runtime system 
     and used it to design one-sided factorizations (prototype for Cholesky 
factorization) that match in 
     performance & scalability #2 but on standard Lapack/ScaLAPACK algorithms 
and data-layout
      (not published yet; can send a copy)

My follow on question is what CPU+GPU combinations are of interest.
Right now I?m thinking both X86-64 + GPU and Power + GPU

Yes, these are definitely of interest. We also have a few of the Jetson TK1s 
that we can play with.
I would also say that of interest are GPU-only software solutions (e.g., to 
reduce energy 
consumption, simplify tuning, etc.).

Stan




On May 19, 2015, at 3:33 PM, Duncan Poole wrote:

Stan and the scalapack team, please forgive me for this ignorant question.
Can Scalapack leverage Magma to take advantage of GPUs on a hybrid cluster?
What is the best way to gpu enable scalapack?

My follow on question is what CPU+GPU combinations are of interest.
Right now I?m thinking both X86-64 + GPU and Power + GPU

Thanks for any insights - Duncan

This email message is for the sole use of the intended recipient(s) and may 
contain confidential information.  Any unauthorized review, use, disclosure 
or distribution is prohibited.  If you are not the intended recipient, please 
contact the sender by reply email and destroy all copies of the original 
message.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
<http://lists.eecs.utk.edu/mailman/private/scalapack/attachments/20150520/76808ae7/attachment.html>

<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or