“Tuning packages for a CPU micro-architecture”

“Tuning packages for a CPU micro-architecture”

[Ludovic Courtès]

Hello there, and happy new year where applicable!

This blog post discusses the new ‘--tune’ option and “package
multi-versioning” as a way of achieving performance without giving up on
reproducibility:

  https://hpc.guix.info/blog/2022/01/tuning-packages-for-a-cpu-micro-architecture/

I’ve found reproducibility to be often depicted as being detrimental to
performance in HPC circles, so this is a contribution to suggest
otherwise.

Ludo’.

Re: “Tuning packages for a CPU micro-architecture”

[zimoun]

Hi Ludo,

Thanks for this nice blog post!

On Thu, 06 Jan 2022 at 17:22, Ludovic Courtès <ludovic.courtes@inria.fr> wrote:

> I’ve found reproducibility to be often depicted as being detrimental to
> performance in HPC circles, so this is a contribution to suggest
> otherwise.

Who knows if an “HPC fool” would not write a ’package-with-tune’
rewriting the ’tune’ properties to all the packages and then run a
world-rebuild. ;-)

Well, the conscious choice of letting ’--tune’ operates only on packages
explicitly marked as “tunable” appears to me a clean design.  It makes a
an explicit and clear separation between “portable-optimized” packages
(almost all) vs “micro-optimized” packages.  The best of both worlds: as
much as possible reproducible, sacrificing for few corner-cases, keeping
the performance for all.


Cheers,
simon

Re: “Tuning packages for a CPU micro-architecture”

[Ludovic Courtès]

Hello,

zimoun <zimon.toutoune@gmail.com> skribis:

> On Thu, 06 Jan 2022 at 17:22, Ludovic Courtès <ludovic.courtes@inria.fr> wrote:
>
>> I’ve found reproducibility to be often depicted as being detrimental to
>> performance in HPC circles, so this is a contribution to suggest
>> otherwise.
>
> Who knows if an “HPC fool” would not write a ’package-with-tune’
> rewriting the ’tune’ properties to all the packages and then run a
> world-rebuild. ;-)

Heh.  :-)

> Well, the conscious choice of letting ’--tune’ operates only on packages
> explicitly marked as “tunable” appears to me a clean design.  It makes a
> an explicit and clear separation between “portable-optimized” packages
> (almost all) vs “micro-optimized” packages.  The best of both worlds: as
> much as possible reproducible, sacrificing for few corner-cases, keeping
> the performance for all.

Yeah, I think so.

A question that remains: even as a “regular user” on my laptop, I use
things that depend for instance on Eigen.  Often Eigen is not critical
to the overall performance (and energy efficiency) of those things, but
it still plays its part; it would be tempting from that perspective to
have ‘--tune’ turned on by default.  I think it’s too early to take that
route but might be something to consider in the future.

Thanks for your feedback!

Ludo’.