[petsc-dev] GAMG error with MKL

[Jeff Hammond]

On Fri, Jul 6, 2018 at 4:28 PM, Smith, Barry F. <bsmith at mcs.anl.gov> wrote:

>
>   Richard,
>
>     The problem is that OpenMP is too large and has too many different
> programming models imbedded in it (and it will get worse) to "support
> OpenMP" from PETSc.
>

This is also true of MPI.  You can write CSP, BSP, PGAS, fork-join,
agent-based, etc. in MPI.  Just like MPI, you don't have to use all the
features.  PETSc doesn't use MPI_Comm_spawn, MPI_Rget_accumulate, or
MPI_Neighborhood_alltoallv, does it?


>     One way to use #pragma based optimization tools (which is one way to
> treat OpenMP) is to run the application code in a realistic size problem,
> using the number of threads/MPI process they prefer with profiling and
> begin adding #pragmas to the most time consuming code fragments/routines,
> measuring the (small) improvement in performance as they are added. This is
> the way I would proceed. The branch generated will not have very many
> pragmas in it so would likely be acceptable to be included into PETSc. It
> would also give a quantitative measure of the possible performance with the
> #prama approach.
>

This is the textbook Wrong Way to write OpenMP and the reason that the
thread-scalability of DOE applications using MPI+OpenMP sucks.  It leads to
codes that do fork-join far too often and suffer from death by Amdahl,
unless you do a second pass where you fuse all the OpenMP regions and
replace the serial regions between them with critical sections or similar.

This isn't how you'd write MPI, is it?  No, you'd figure out how to
decompose your data properly to exploit locality and then implement an
algorithm that minimizes communication and synchronization.  Do that with
OpenMP.


>    Note: that for BLAS 1 operations likely the correct thing to do is turn
> on MKL BLAS threading (being careful to make sure the number of threads MKL
> uses matches that used by other parts of the code). This way we don't need
> to OpenMP optimize many parts of PETSc's vector operations (norm, dot,
> scale, axpy). In fact, this is the first thing Mark should do, how much
> does it speed up the vector operations?
>

BLAS1 operations are all memory-bound unless running out of cache (in which
case one shouldn't use threads) and compilers do a great job with them.
Just put the pragmas on and let the compiler do its job.


>   The problem is how many ECP applications actually use OpenMP just as a
> #pragma optimization tool, or do they use other features of OpenMP. For
> example I remember Brian wanted to/did use OpenMP threads directly in
> BoxLib and didn't just stick to the #pragma model. If they did this then we
> would need custom PETSc to match their model.
>

If this implies that BoxLib will use omp-parallel and then use explicit
threading in a manner similar to MPI (omp_get_num_threads=MPI_Comm_size and
omp_get_thread_num=MPI_Comm_rank), then this is the Right Way to write
OpenMP.

Unfortunately, the Right Way to use OpenMP makes it hard to use MPI unless
you use MPI_THREAD_MULTIPLE and endpoints.  ECP projects should be pushing
the MPI folks harder to ratify and implement endpoints.  I don't know if
the proposal is even active right now, but that doesn't prevent DOE from
compelling Open-MPI and MPICH to support it.

To end on a positive note, OpenMP tasking is a relatively composable model
and supports DAG-based parallelism.  I suspect the initial results in a
code like PETSc will be worse than with traditional implicit OpenMP
(omp-for-simd on all the loops) but it eventually wins own because it
doesn't require any unnecessary barriers and makes it much easier to fuse
parallel regions.

Jeff


>
>   Barry
>
>
> > On Jul 6, 2018, at 3:07 PM, Mills, Richard Tran <rtmills at anl.gov> wrote:
> >
> > True, Barry. But, unfortunately, I think Jed's argument has something to
> it because the hybrid MPI + OpenMP model has become so popular. I know of a
> few codes where adopting this model makes some sense, though I believe
> that, more often, the model has been adopted simply because it is the
> fashionable thing to do. Regardless of good or bad reasons for its
> adoption, I do have some real concern that codes that use this model have a
> difficult time using PETSc effectively because of the lack of thread
> support. Like many of us, I had hoped that endpoints would make it into the
> MPI standard and this would provide a reasonable mechanism for integrating
> PETSc with codes using MPI+threads, but progress on this seems to have
> stagnated. I hope that the MPI endpoints effort eventually goes somewhere,
> but what can we do in the meantime? Within the DOE ECP program, the
> MPI+threads approach is being pushed really hard, and many of the ECP
> subprojects have adopted it. I think it's mostly idiotic, but I think it's
> too late to turn the tide and convince most people that pure MPI is the way
> to go. Meanwhile, my understanding is that we need to be able to support
> more of the ECP application projects to justify the substantial funding we
> are getting from the program. Many of these projects are dead-set on using
> OpenMP. (I note that I believe that the folks Mark is trying to help with
> PETSc and OpenMP are people affiliated with Carl Steefel's ECP subsurface
> project.)
> >
> > Since it looks like MPI endpoints are going to be a long time (or
> possibly forever) in coming, I think we need (a) stopgap plan(s) to support
> this crappy MPI + OpenMP model in the meantime. One possible approach is to
> do what Mark is trying with to do with MKL: Use a third party library that
> provides optimized OpenMP implementations of computationally expensive
> kernels. It might make sense to also consider using Karl's ViennaCL library
> in this manner, which we already use to support GPUs, but which I believe
> (Karl, please let me know if I am off-base here) we could also use to
> provide OpenMP-ized linear algebra operations on CPUs as well. Such
> approaches won't use threads for lots of the things that a PETSc code will
> do, but might be able to provide decent resource utilization for the most
> expensive parts for some codes.
> >
> > Clever ideas from anyone on this list about how to use an adequate
> number of MPI ranks for PETSc while using only a subset of these ranks for
> the MPI+OpenMP application code will be appreciated, though I don't know if
> there are any good solutions.
> >
> > --Richard
> >
> > On Wed, Jul 4, 2018 at 11:38 PM, Smith, Barry F. <bsmith at mcs.anl.gov>
> wrote:
> >
> >    Jed,
> >
> >      You could use your same argument to argue PETSc should do
> "something" to help people who have (rightly or wrongly) chosen to code
> their application in High Performance Fortran or any other similar inane
> parallel programming model.
> >
> >    Barry
> >
> >
> >
> > > On Jul 4, 2018, at 11:51 PM, Jed Brown <jed at jedbrown.org> wrote:
> > >
> > > Matthew Knepley <knepley at gmail.com> writes:
> > >
> > >> On Wed, Jul 4, 2018 at 4:51 PM Jeff Hammond <jeff.science at gmail.com>
> wrote:
> > >>
> > >>> On Wed, Jul 4, 2018 at 6:31 AM Matthew Knepley <knepley at gmail.com>
> wrote:
> > >>>
> > >>>> On Tue, Jul 3, 2018 at 10:32 PM Jeff Hammond <
> jeff.science at gmail.com>
> > >>>> wrote:
> > >>>>
> > >>>>>
> > >>>>>
> > >>>>> On Tue, Jul 3, 2018 at 4:35 PM Mark Adams <mfadams at lbl.gov> wrote:
> > >>>>>
> > >>>>>> On Tue, Jul 3, 2018 at 1:00 PM Richard Tran Mills <
> rtmills at anl.gov>
> > >>>>>> wrote:
> > >>>>>>
> > >>>>>>> Hi Mark,
> > >>>>>>>
> > >>>>>>> I'm glad to see you trying out the AIJMKL stuff. I think you are
> the
> > >>>>>>> first person trying to actually use it, so we are probably going
> to expose
> > >>>>>>> some bugs and also some performance issues. My somewhat limited
> testing has
> > >>>>>>> shown that the MKL sparse routines often perform worse than our
> own
> > >>>>>>> implementations in PETSc.
> > >>>>>>>
> > >>>>>>
> > >>>>>> My users just want OpenMP.
> > >>>>>>
> > >>>>>>
> > >>>>>
> > >>>>> Why not just add OpenMP to PETSc? I know certain developers hate
> it, but
> > >>>>> it is silly to let a principled objection stand in the way of
> enabling users
> > >>>>>
> > >>>>
> > >>>> "if that would deliver the best performance for NERSC users."
> > >>>>
> > >>>> You have answered your own question.
> > >>>>
> > >>>
> > >>> Please share the results of your experiments that prove OpenMP does
> not
> > >>> improve performance for Mark’s users.
> > >>>
> > >>
> > >> Oh God. I am supremely uninterested in minutely proving yet again that
> > >> OpenMP is not better than MPI.
> > >> There are already countless experiments. One more will not add
> anything of
> > >> merit.
> > >
> > > Jeff assumes an absurd null hypothesis, Matt selfishly believes that
> > > users should modify their code/execution environment to subscribe to a
> > > more robust and equally performant approach, and the MPI forum
> abdicates
> > > by stalling on endpoints.  How do we resolve this?
> > >
> > >>> Also we are not in the habit of fucking up our codebase in order to
> follow
> > >>>> some fad.
> > >>>>
> > >>>
> > >>> If you can’t use OpenMP without messing up your code base, you
> probably
> > >>> don’t know how to design software.
> > >>>
> > >>
> > >> That is an interesting, if wrong, opinion. It might be your
> contention that
> > >> sticking any random paradigm in a library should
> > >> be alright if its "well designed"? I have never encountered such a
> > >> well-designed library.
> > >>
> > >>
> > >>> I guess if you refuse to use _Pragma because C99 is still a fad for
> you,
> > >>> it is harder, but clearly _Complex is tolerated.
> > >>>
> > >>
> > >> Yes, littering your code with preprocessor directives improves almost
> > >> everything. Doing proper resource management
> > >> using Pragmas, in an environment with several layers of libraries, is
> a
> > >> dream.
> > >>
> > >>
> > >>> More seriously, you’ve adopted OpenMP hidden behind MKL
> > >>>
> > >>
> > >> Nope. We can use MKL with that crap shutoff.
> > >>
> > >>
> > >>> so I see no reason why you can’t wrap OpenMP implementations of the
> PETSc
> > >>> sparse kernels in a similar manner.
> > >>>
> > >>
> > >> We could, its just a colossal waste of time and effort, as well as
> > >> counterproductive for the codebase :)
> > >
> > > Endpoints either need to become a thing we can depend on or we need a
> > > solution for users that insist on using threads (even if their decision
> > > to use threads is objectively bad).  The problem Matt harps on is
> > > legitimate: OpenMP parallel regions cannot reliably cross module
> > > boundaries except for embarrassingly parallel operations.  This means
> > > loop-level omp parallel which significantly increases overhead for
> small
> > > problem sizes (e.g., slowing coarse grid solves and strong scaling
> > > limits).  It can be done and isn't that hard, but the Imperial group
> > > discarded their branch after observing that it also provided no
> > > performance benefit.  However, I'm coming around to the idea that PETSc
> > > should do it so that there is _a_ solution for users that insist on
> > > using threads in a particular way.  Unless Endpoints become available
> > > and reliable, in which case we could do it right.
> >
> >
>
>


-- 
Jeff Hammond
jeff.science at gmail.com
http://jeffhammond.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20180709/a44b54c2/attachment.html>