[petsc-dev] Kokkos/Crusher perforance

Mon Jan 24 14:30:07 CST 2022

I want to note that crusher is early access hardware, so we should expect performance to not be great right now.  Doing what we can to help identify the performance issues and keeping OLCF informed would be the best.

Note that we cannot make any of the preliminary results publicly available without explicit permission from OLCF; all of the results have to be considered preliminary and the software stack will undergo a rapid churn.

All the best, Todd.

From: petsc-dev <petsc-dev-bounces at mcs.anl.gov> on behalf of Barry Smith <bsmith at petsc.dev>
Date: Monday, January 24, 2022 at 2:24 PM
To: Justin Chang <jychang48 at gmail.com>
Cc: "petsc-dev at mcs.anl.gov" <petsc-dev at mcs.anl.gov>
Subject: Re: [petsc-dev] Kokkos/Crusher perforance

  For this, to start, someone can run

src/vec/vec/tutorials/performance.c

and compare the performance to that in the technical report Evaluation of PETSc on a Heterogeneous Architecture \\ the OLCF Summit System \\ Part I: Vector Node Performance. Google to find. One does not have to and shouldn't do an extensive study right now that compares everything, instead one should run a very small number of different size problems (make them big) and compare those sizes with what Summit gives. Note you will need to make sure that performance.c uses the Kokkos backend.

  One hopes for better performance than Summit; if one gets tons worse we know something is very wrong somewhere. I'd love to see some comparisons.

  Barry

On Jan 24, 2022, at 3:06 PM, Justin Chang <jychang48 at gmail.com<mailto:jychang48 at gmail.com>> wrote:

Also, do you guys have an OLCF liaison? That's actually your better bet if you do.

Performance issues with ROCm/Kokkos are pretty common in apps besides just PETSc. We have several teams actively working on rectifying this. However, I think performance issues can be quicker to identify if we had a more "official" and reproducible PETSc GPU benchmark, which I've already expressed to some folks in this thread, and as others already commented on the difficulty of such a task. Hopefully I will have more time soon to illustrate what I am thinking.

On Mon, Jan 24, 2022 at 1:57 PM Justin Chang <jychang48 at gmail.com<mailto:jychang48 at gmail.com>> wrote:
My name has been called.

Mark, if you're having issues with Crusher, please contact Veronica Vergara (vergaravg at ornl.gov<mailto:vergaravg at ornl.gov>). You can cc me (justin.chang at amd.com<mailto:justin.chang at amd.com>) in those emails

On Mon, Jan 24, 2022 at 1:49 PM Barry Smith <bsmith at petsc.dev<mailto:bsmith at petsc.dev>> wrote:

On Jan 24, 2022, at 2:46 PM, Mark Adams <mfadams at lbl.gov<mailto:mfadams at lbl.gov>> wrote:

Yea, CG/Jacobi is as close to a benchmark code as we could want. I could run this on one processor to get cleaner numbers.

Is there a designated ECP technical support contact?

   Mark, you've forgotten you work for DOE. There isn't a non-ECP technical support contact.

   But if this is an AMD machine then maybe contact Matt's student Justin Chang?

On Mon, Jan 24, 2022 at 2:18 PM Barry Smith <bsmith at petsc.dev<mailto:bsmith at petsc.dev>> wrote:

  I think you should contact the crusher ECP technical support team and tell them you are getting dismel performance and ask if you should expect better. Don't waste time flogging a dead horse.

On Jan 24, 2022, at 2:16 PM, Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>> wrote:

On Mon, Jan 24, 2022 at 2:11 PM Junchao Zhang <junchao.zhang at gmail.com<mailto:junchao.zhang at gmail.com>> wrote:

On Mon, Jan 24, 2022 at 12:55 PM Mark Adams <mfadams at lbl.gov<mailto:mfadams at lbl.gov>> wrote:

On Mon, Jan 24, 2022 at 1:38 PM Junchao Zhang <junchao.zhang at gmail.com<mailto:junchao.zhang at gmail.com>> wrote:
Mark, I think you can benchmark individual vector operations, and once we get reasonable profiling results, we can move to solvers etc.

Can you suggest a code to run or are you suggesting making a vector benchmark code?
Make a vector benchmark code, testing vector operations that would be used in your solver.
Also, we can run MatMult() to see if the profiling result is reasonable.
Only once we get some solid results on basic operations, it is useful to run big codes.

So we have to make another throw-away code? Why not just look at the vector ops in Mark's actual code?

   Matt

--Junchao Zhang

On Mon, Jan 24, 2022 at 12:09 PM Mark Adams <mfadams at lbl.gov<mailto:mfadams at lbl.gov>> wrote:

On Mon, Jan 24, 2022 at 12:44 PM Barry Smith <bsmith at petsc.dev<mailto:bsmith at petsc.dev>> wrote:

  Here except for VecNorm the GPU is used effectively in that most of the time is time is spent doing real work on the GPU

VecNorm              402 1.0 4.4100e-01 6.1 1.69e+09 1.0 0.0e+00 0.0e+00 4.0e+02  0  1  0  0 20   9  1  0  0 33 30230   225393      0 0.00e+00    0 0.00e+00 100

Even the dots are very effective, only the VecNorm flop rate over the full time is much much lower than the vecdot. Which is somehow due to the use of the GPU or CPU MPI in the allreduce?

The VecNorm GPU rate is relatively high on Crusher and the CPU rate is about the same as the other vec ops. I don't know what to make of that.

But Crusher is clearly not crushing it.

Junchao: Perhaps we should ask Kokkos if they have any experience with Crusher that they can share. They could very well find some low level magic.

On Jan 24, 2022, at 12:14 PM, Mark Adams <mfadams at lbl.gov<mailto:mfadams at lbl.gov>> wrote:

Mark, can we compare with Spock?

 Looks much better. This puts two processes/GPU because there are only 4.
<jac_out_001_kokkos_Spock_6_1_notpl.txt>

--
What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/<http://www.cse.buffalo.edu/~knepley/>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20220124/751654ad/attachment-0001.html>