dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Aug 09, 2010 at 02:34:45PM +0100, Garth N. Wells wrote:
> On Mon, 2010-08-09 at 15:20 +0200, Anders Logg wrote:
> > On Mon, Aug 09, 2010 at 02:12:49PM +0100, Garth N. Wells wrote:
> > > On Mon, 2010-08-09 at 15:05 +0200, Anders Logg wrote:
> > > > On Mon, Aug 09, 2010 at 01:54:04PM +0100, Garth N. Wells wrote:
> > > > > On Mon, 2010-08-09 at 14:46 +0200, Anders Logg wrote:
> > > > > > On Mon, Aug 09, 2010 at 01:09:47PM +0100, Garth N. Wells wrote:
> > > > > > > On Mon, 2010-08-09 at 13:53 +0200, Anders Logg wrote:
> > > > > > > > On Mon, Aug 09, 2010 at 12:47:10PM +0100, Garth N. Wells wrote:
> > > > > > > > > On Mon, 2010-08-09 at 13:37 +0200, Anders Logg wrote:
> > > > > > > > > > On Sat, Aug 07, 2010 at 01:24:44PM +0100, Garth N. Wells wrote:
> > > > > > > > > > > On Fri, 2010-08-06 at 19:55 +0100, Garth N. Wells wrote:
> > > > > > > > > > > > On Fri, 2010-08-06 at 20:53 +0200, Anders Logg wrote:
> > > > > > > > > > > > > On Fri, Aug 06, 2010 at 07:51:18PM +0100, Garth N. Wells wrote:
> > > > > > > > > > > > > > On Fri, 2010-08-06 at 20:36 +0200, Anders Logg wrote:
> > > > > > > > > > > > > > > On Fri, Aug 06, 2010 at 04:55:44PM +0100, Garth N. Wells wrote:
> > > > > > > > > > > > > > > > On Fri, 2010-08-06 at 08:42 -0700, Johan Hake wrote:
> > > > > > > > > > > > > > > > > On Friday August 6 2010 08:16:26 you wrote:
> > > > > > > > > > > > > > > > > > ------------------------------------------------------------
> > > > > > > > > > > > > > > > > > revno: 4896
> > > > > > > > > > > > > > > > > > committer: Garth N. Wells <gnw20@xxxxxxxxx>
> > > > > > > > > > > > > > > > > > branch nick: dolfin-all
> > > > > > > > > > > > > > > > > > timestamp: Fri 2010-08-06 16:13:29 +0100
> > > > > > > > > > > > > > > > > > message:
> > > > > > > > > > > > > > > > > >   Add simple Stokes solver for parallel testing.
> > > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > >   Other Stokes demos don't run in parallel because MeshFunction io is not
> > > > > > > > > > > > > > > > > >   supported in parallel.
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > Does anyone have an overview of what is needed for this to be fixed. I
> > > > > > > > > > > > > > > > > couldn't find a blueprint on it.
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Here it is:
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >     https://blueprints.launchpad.net/dolfin/+spec/parallel-io
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > I am interested in getting this fixed :)
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Me too! We need to look at all the io since much of it is broken in
> > > > > > > > > > > > > > > > parallel.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > We need to settle on how to handle XML data. I favour (and I know Niclas
> > > > > > > > > > > > > > > > Janson does too) the VTK approach in which we have a 'master file' that
> > > > > > > > > > > > > > > > points to other XML files which contain portions of the vector/mesh,
> > > > > > > > > > > > > > > > etc. Process zero can read the 'master file' and then instruct the other
> > > > > > > > > > > > > > > > processes on which file(s) they should read in.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > This only works if the data is already partitioned. Most of our demos
> > > > > > > > > > > > > > > assume that we have the mesh in one single file which is then
> > > > > > > > > > > > > > > partitioned on the fly.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > The approach does work for data which is not partitioned. Just like with
> > > > > > > > > > > > > > VTK, one can read the 'master file' or the individual files.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > > The initial plan was to support two different ways of reading data in parallel:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > 1. One file and automatic partitioning
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > DOLFIN gets one file "mesh.xml", each process reads one part of it (just
> > > > > > > > > > > > > > > skipping other parts of the file), then the mesh is partitioned and
> > > > > > > > > > > > > > > redistributed.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > 2. Several files and no partitioning
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > DOLFIN get multiple files and each process reads one part. In this
> > > > > > > > > > > > > > > case, the mesh and all associated data is already partitioned. This
> > > > > > > > > > > > > > > should be very easy to fix since everything that is needed is already
> > > > > > > > > > > > > > > in place; we just need to fix the logic. In particular, the data
> > > > > > > > > > > > > > > section of each local mesh contains all auxilliary parallel data.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > This can be handled in two different ways. Either a user specifies the
> > > > > > > > > > > > > > > name of the file as "mesh*.xml", in which case DOLFIN appends say
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > >   "_%d" % MPI::process_number()
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > on each local process.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > The other way is to have a master file which lists all the other
> > > > > > > > > > > > > > > files. In this case, I don't see a need for process 0 to take any kind
> > > > > > > > > > > > > > > of responsibility for communicating file names. It would work fine for
> > > > > > > > > > > > > > > each process to read the master file and then check which file it
> > > > > > > > > > > > > > > should use. Each process could also check that the total number of
> > > > > > > > > > > > > > > processes matches the number of partitions in the file. We could let
> > > > > > > > > > > > > > > process 0 handle the parsing of the master file and then communicate
> > > > > > > > > > > > > > > the file names but maybe that is an extra complication.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > This fails when the number of files differs from the number of
> > > > > > > > > > > > > > processes. It's very important to support m files on n processes. We've
> > > > > > > > > > > > > > discussed this at length before.
> > > > > > > > > > > > >
> > > > > > > > > > > > > I don't remember. Can you remind me of what the reasons are?
> > > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > I perform a simulation using m processes, and write the result to m
> > > > > > > > > > > > files. Later I want to use the result later in another computation using
> > > > > > > > > > > > n processors.
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > I've looked a little into parallel io, and looked at what Trilinos and
> > > > > > > > > > > PETSc do. Both support HDF5, and HDF5 has been developed to work
> > > > > > > > > > > parallel. HDF5 does not advocate the one-file per processes (too awkward
> > > > > > > > > > > and complicated they say), but advocates a one file approach. It has
> > > > > > > > > > > tools that allow different processes to write to different parts of the
> > > > > > > > > > > same file in parallel.
> > > > > > > > > > >
> > > > > > > > > > > >From reading this, what I propose (for now) is:
> > > > > > > > > > >
> > > > > > > > > > > 1. We only ever write one XML file for a given object. This file can be
> > > > > > > > > > > read by different processes, with each reading in only a chunk.
> > > > > > > > > > >
> > > > > > > > > > > 2. We should add an XML format for partitioning data (Trilinos calls
> > > > > > > > > > > this a 'map'). If a map file is present, it is used to define the
> > > > > > > > > > > partitions. It may make sense to have a map file for each process (but
> > > > > > > > > > > no need for a 'master file').
> > > > > > > > > >
> > > > > > > > > > I suggest something slightly different. I'm ok with the one file
> > > > > > > > > > approach, but it would be good to store that data in a partitioned
> > > > > > > > > > way. Our current model for parallel computing is that each process has
> > > > > > > > > > a Mesh and each process has the partitioning data it needs stored in
> > > > > > > > > > the data section of the Mesh. So each process has just a regular mesh
> > > > > > > > > > with some auxilliary data attached to it. That makes it easy to read
> > > > > > > > > > and write using already existing code. (No need for a special parallel
> > > > > > > > > > format.)
> > > > > > > > > >
> > > > > > > > > > But we could easily throw all that data into one big file, something
> > > > > > > > > > like this:
> > > > > > > > > >
> > > > > > > > > > <distributed_mesh num_parts="16">
> > > > > > > > > >   <mesh ...>
> > > > > > > > > >     ...
> > > > > > > > > >   </mesh>
> > > > > > > > > >   <mesh ...>
> > > > > > > > > >     ...
> > > > > > > > > >   </mesh>
> > > > > > > > > >   ...
> > > > > > > > > > </distributed_mesh>
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > I would like to separate mesh and partitioning data. A partitioning of a
> > > > > > > > > given mesh is not unique to a mesh, so it should be separated. A
> > > > > > > > > partition could still go in the same XML file though.
> > > > > > > >
> > > > > > > > It is separate from the mesh, but it is stored locally as part of the
> > > > > > > > Mesh in MeshData. I think this has proved to be a very good way
> > > > > > > > (efficient and simple) to store the data.
> > > > > > > >
> > > > > > >
> > > > > > > This breaks the concept of using a given Mesh on a different number of
> > > > > > > processes. For a given mesh, I may want to store and use different
> > > > > > > partitions.
> > > > > > >
> > > > > > > The same applied to Vectors.
> > > > > > >
> > > > > > > > Or do you suggest that we store it differently on file than we do as
> > > > > > > > part of the DOLFIN data structures?
> > > > > > > >
> > > > > > >
> > > > > > > I suggest that we decouple the Mesh and the partition data for file
> > > > > > > output. It comes back to permitting a variable number of processes.
> > > > > >
> > > > > > I thought that the storage output from m processes would be the mesh
> > > > > > partitioned into m pieces with auxilliary partitioning
> > > > > > information. This would then be read by n processes and repartitioned.
> > > > > >
> > > > > > If you don't want the partitioning into a specific number of
> > > > > > partitions (like m partitions above), then what extra partitioning
> > > > > > data would be required? It would just be a regular mesh file.
> > > > > >
> > > > >
> > > > > It would map entities to processes. It could also be used for Vectors.
> > > >
> > > > Then how is that not coupled to a specific number of processes?
> > > >
> > > > And why would it be better than what we have today, which stores all
> > > > the parallel data we need (not just that mapping but everything else
> > > > that can and must be generated from it) as part of each mesh?
> > > >
> > > > > > Partitioning data is always coupled to a specific number of processes
> > > > > > as far as I can imagine.
> > > > > >
> > > > >
> > > > > Naturally, and so it should be. But the mesh output doesn't have to be,
> > > > > which why I suggest having a Mesh and partitioning data, e.g.
> > > > >
> > > > >   // Let DOLFIN partition and distribute the mesh
> > > > >   Mesh meshA("mesh.xml")
> > > >
> > > > Is this just one regular mesh file?
> > > >
> > >
> > > Yes, and ParMETIS would partition the mesh.
> > >
> > > > >   // Partition the mesh according to my_partition.xml mesh. Throw an
> > > > >   // error if there is a mis-match between my_partition and the number
> > > > >   // of processes
> > > > >   Mesh meshB("mesh.xml", "my_partition.xml")
> > > >
> > > > I don't understand this use-case. The first line above would already
> > > > partition the mesh (using ParMETIS), and then you want to repartition
> > > > it. Is it the case that my_partition.xml contains a better
> > > > partitioning than what ParMETIS can compute?
> > > >
> > >
> > > In the case "meshB("mesh.xml", "my_partition.xml")", ParMETIS would
> > > never be called. The mesh would be distributed according to
> > > "my_partition.xml".
> >
> > I missed that you created two different meshes. ok, that sounds like a
> > good thing to have.
> >
> > But how would the mesh.xml be stored on file? Just one big file?
> >
>
> Yes, just as we have now.

ok.

> > And how is the above example related to the storing of *partitioned*
> > meshes?
> >
>
> >From reading around, I don't think that we should store partitioned
> meshes in the sense that each partition is a separate file. We should
> store an entire mesh in one file, and store the partitioning data if
> desired.

I don't understand why not. It's a good feature to have to be able to
restart a simulation (on the same number of processors) and completely
avoid the partitioning step. Moreover, everything we need is already
implemented since we can read/write meshes including mesh data.

In summary, if I understand correctly that you want things stored in
one file, we would have the following:

1. Reading from one file and automatic partitioning/redistribution

This is already supported today, and it already supports the use case
of first running on m processors and then n processors since in both
cases the mesh will be read from one single file and partitioned.

2. Reading from multiple files where the mesh has already been
partitioned and m = n. This is almost supported today. We just need to
decide on the logic for reading from multiple files.

In addition to the above two cases (which are in principle already in
place) we would add the possibility to supply the partitioning. This
would be simple to do as it would just bypass the call to ParMETIS.

> For scalable parallel io, we should use the parallel features of HDF5.
> Each process starts writing to the same file but at the right position.
> See
>
>     http://www.spscicomp.org/ScicomP14/talks/Yang.ppt

ok.

--
Anders

Attachment: signature.asc
Description: Digital signature