Cray and PETSc

[Barry Smith]

   Adrian,

    Thank you for in the inquiry.
  

On Thu, 16 Aug 2007, Adrian Tate wrote:

> Hello Barry
>
> I'm not sure we met in person - I am the lead of the libraries group
>at Cray. We decided some time ago that our iterative solver strategy
>would be to leverage PETSc, and to hopefully provide some Cray
>specific tunings to improve performance of the KSP solvers (largely
>through our custom sparse BLAS and some parallel tuning for our
>interconnect). I believe that John Lewis has been in communication
>with you with respect to the tuning of Sparse BLAS and their
>integration with the PETSc build.
>
>We are however, considering packaging and supplying PETSc along with
>the scientific library that we provide (libSci). This allows for
>better integration of our internal kernels and also it means that we
>are no longer requiring users (including benchmakers and internal
>users) to build their own PETSc. I see from your online page that
>doing so is acceptable as long as we use a copyright switch during
>configure. By applying this switch, do we make our PETSc library
>unsupported from your perspective?

   The GNU copyright code is tiny; it would not effect the usability
or support of PETSc

> We do not expect to be able to
>provide anywhere near the degree of support that your team provide,
>and I was hoping to supply a pre-built library whose users could still
>seek assistance through your normal support channels - is this
>realistic?
>
   Yes. 

   The two isses that concern us with pre-packaged versions of PETSc are 
1) keeping up to date on our releases. We generally make two releases
a year and much prefer that users use the most recently release.
If they are using an older release it means we are less able to help
them. 
2) keeping up to date on our patches. We may make several bug patches
to each release. Users with a pre-packaged version have trouble keeping
up with the source code patches we provide.

>  
> 
>
>
> Also, I would be interested to know your degree of interest in the
>Cray-specific modifications that we make - would you prefer those to
>be channeled back into the PETSc library? 

  If they involved directly changing PETSc code, we much prefer to get
it channeled back into the master copy of the source code. Makes it
much easier to debug user code. If it is auxiliary code, like a faster
ddot() etc then it is more appropriate to not try to include it.

>Any other comments you have
>on the way that Cray can contribute to the PETSc project, I would be
>very glad to hear.

  PETSc has a variety of "tuning factors" that could theoretically be
set optimally for a particular machine, these range from simple things
like compiler options, to a choice between C and Fortran versions of
the same code (what we call them Fortran kernels), to different loop
unrollings (in the inline.h file), even something like PetscMemzero()
which has five possible forms. Now currently we do not tune these
or even have a test harness for selecting a good tuning. One thing 
you could/would like to do, is determine good choices for these
options on your machines. Just as a simple example, on some Linux systems
the basic libraries are just compiled with the GNU compiler, hence the
system memset() is not particularly effective. A version compiled of PETSc
compiled using a "Fortran memset" may be much faster, or Intel provides
its own _intel_fast_memset() which is better. I've seen a few percent
increase in performance of entire nonlinear PDE solver applications
just from using the non-default memset(). [This was specifically on
an Itanium system, but is likely on other configurations also].


   Barry
 

> 
>  
> 
> Regards,
> 
> 
> Adrian Tate
> 
> ---
> 
> Technical Lead
> 
> Math Software
> Cray Inc.
> 
> (206) 349 5868
> 
>