On Wed, 26 Nov, 2014 at 10:24 AM, Johan Hake <[email protected]> wrote:


On Wed, Nov 26, 2014 at 10:52 AM, Garth N. Wells <[email protected]> wrote:

On Wed, 26 Nov, 2014 at 9:38 AM, Johan Hake <[email protected]> wrote:
On Wed, Nov 26, 2014 at 10:28 AM, Garth N. Wells <[email protected]> wrote:

On Wed, 26 Nov, 2014 at 9:09 AM, Johan Hake <[email protected]> wrote:
On Wed, Nov 26, 2014 at 9:43 AM, Garth N. Wells <[email protected]> wrote:
On Wed, 26 Nov, 2014 at 8:32 AM, Johan Hake <[email protected]> wrote:
On Wed, Nov 26, 2014 at 9:22 AM, Garth N. Wells <[email protected]> wrote:


On Wed, 26 Nov, 2014 at 7:50 AM, Johan Hake <[email protected]> wrote:
On Wed, Nov 26, 2014 at 8:34 AM, Garth N. Wells <[email protected]> wrote:


On Tue, 25 Nov, 2014 at 9:48 PM, Johan Hake <[email protected]> wrote:
Hello!

I just pushed some fixes to the jit interface of DOLFIN. Now one can jit on different mpi groups.

Nice.

Previously jiting was only done on rank 1 of the mpi_comm_world. Now it is done on rank 1 of any passed group communicator.

Do you mean rank 0?

​Yes, of course.​


There is no demo atm showing this but a test has been added:

  test/unit/python/jit/test_jit_with_mpi_groups.py

Here an expression, a subdomain, and a form is constructed on different ranks using group. It is somewhat tedious as one need to initialize PETSc with the same group, otherwise PETSc will deadlock during initialization (the moment a PETSc la object is constructed).

This is ok. It's arguably a design flaw that we don't make the user handle MPI initialisation manually.

​Sure, it is just somewhat tedious. You cannot start your typical script with importing dolfin.​

The procedure in Python for this is:

1) Construct mpi groups using mpi4py
2) Initalize petscy4py using the groups
3) Wrap groups to petsc4py comm (dolfin only support petsc4py not mpi4py)
4) import dolfin
5) Do group specific stuff:
   a) Function and forms no change needed as communicator
      is passed via mesh
   b) domain = CompiledSubDomain("...", mpi_comm=group_comm)
   c) e = Expression("...", mpi_comm=group_comm)

It's not so clear whether passing the communicator means that the Expression is only defined/available on group_comm, or if group_comm is simply to control who does the JIT. Could you clarify this?

My knowledge is not that good in MPI. I have only tried to access (and construct) the Expression on ranks included in that group. Also when I tried construct one using a group communicator on a rank that is not included in the group, I got an when calling MPI_size on it. There is probably a perfectly reasonable explaination to this. ​​

Could you clarify what goes on behind-the-scenes with the communicator? Is it only used in a call to get the process rank? What do the ranks other than zero do?

​Not sure what you want to know. Instead of using mpi_comm_world to construct meshes you use the group communicator. This communicator has its own local group of ranks​. JITing is still done on rank 0 of the local group, which might and most often is different from rank 0 process of the mpi_comm_word.

I just want to be clear (and have in the docstring) that

   e = Expression("...", mpi_comm=group_comm)

is valid only on group_comm (if this is the case), or make clear that the communicator only determines the process that does the JIT.

​I see now what you mean. I can update the docstring. As far as I understand it should be that the expression is only valid on group_comm, and that rank 0 of that group take care of the JIT.


OK, could you make this clear in the docstring?

​Sure.​

If we required all Expressions to have a domain/mesh, as Martin advocates, things would be clearer.

Sure, but the same question is there for the mesh too. Is it available on ranks that is not in the group?


I think in this case it is clear - a mesh lives only on the processes belonging to its communicator. The ambiguity with an Expression is that is doesn't have any data that lives on processes.

​Sure.​

The group communicator works exactly like the world communicator but now on just a subset of the processes. There were some sharp edges with deadlocks as a consequence, when barriers were taken on the world communicator. This is done by default when dolfin is imported and petcs gets initialized with the world communicator. So we need to initialized petsc using the group communicator. Other than that there are not real differences.

That doesn't sound right. PETSc initialisation does not take a communicator. It is collective on MPI_COMM_WORLD, but each PETSc object takes a communicator at construction, which can be something other than MPI_COMM_WORLD or MPI_COMM_SELF.

​Well, for all I know petsc can be initialized with a mpi_comm​. In petsc4py that is done by:

  import petsc4py
  petsc4py.init(comm=group_1)
  import petsc4py.PETSc as petsc

It turned out that this was required for the Function constructor to not deadlock. The line:

    _vector = factory.create_vector();

initilizes PETSc with world_comm, which obviously deadlocks.

There must be something else wrong. PETScInitialize does not take a communicator:

http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/Sys/PetscInitialize.html


From that web page:

​Collective on MPI_COMM_WORLD or PETSC_COMM_WORLD if it has been set

​So setting PETSC_COMM_WORLD initializes PETSc on a subset of processes.

It is fundamentally wrong that PETSc is 'initialised' with a sub-communicator in DOLFIN to avoid a deadlock for the code use sketched in this thread. Communicators belong to objects. Initialising PETSc on a sub-communicator, it would not be possible to solve a subsequent problem on more processes than contained in the sub-communicator.

If the code is deadlocking, then the underlying problem should be fixed. It's probably a function call somewhere that is erroneously not passing the communicator (since in cases where a communicator is not passed, we assume it to be MPI_COMM_WORLD - not ideal but keeps it simple for most users).

​Found the bugger!

If the first PETScObject is created only on a group, SubSystemManager::​init_petsc is called. This wont initialize PETSc if it is already initialized outside, for example with MPI_COMM_WORLD. However SLEPc will be initialized using PETSc_COMM_WORLD. If that is set to MPI_COMM_WORLD we have a dead lock. If we however, create a PETScVector on all processes before we create the group one. We are fine as SLEPc then gets initialized on all processes.

So the solution is to call:

SubsystemManager.init_petsc()

on all processes before doing anything.

I don't quite follow. Could you post a simple snippet that deadlocks when SubsystemManager.init_petsc() is not called, but works when it is called?

Garth

Well, I learned a bunch of MPI related stuff here :) I will update the tests accordingly.

Johan




Garth



Also see:

​http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/Sys/PETSC_COMM_WORLD.html
​Johan​


Why does petsc4py want one? It doesn't make sense to initialise it with a communicator - a communicator belongs to objects.

Garth


You might say that this could be avoided by initializing PETSc on all ranks with the world communicator before constructing a Function on a group. However it still deadlocks during construction. Here I have just assumed it deadlock at the same line, but I need to double check this. And when I initilized PETSc using the group communicator it just worked. So somewhere a collective call to mpi_world_comm is executed when constructing a PETScVector.

Johan






Garth



​Johan​



Garth



Please try it out and report any sharp edges. A demo would also be fun to include :)

We could run tests on different communicators to speed them up on machines with high core counts!

True!

Johan​


Garth


Johan











_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics

Reply via email to