RE: [pooma-dev] Problems running in parallel

[Stephen Smith]

Dave,

Almost all the of tests in NewField/tests create fields that are
distributed in parallel.  Running ExpressionTest and StencilTests
with -shmem -np 4 verifies that you are running on top of Cheetah
correctly.  FieldTour1 creates a layout with ReplicatedTag() instead
of LayoutTag_t(), which is why it fails (the test has a bug).

If you based your parallel code on FieldTour1, verify that you are
using Remote<Brick> or Remote<CompressibleBrick> as the patch tag,
and DistributeTag() as the tag used to initialize the layout.
The fact that you are not getting reproducible results is disturbing.
Could you compile with --bounds as well as --messaging to verify that
enough guard layers are used everywhere?  If the code still runs without
errors, then we probably have a problem with the evaluation somewhere.
A useful serial test would be to compile with "--sched serialAsync" which
uses data-flow evaluation on one processor.  If you see problems
with serialAsync, then we would know that the error is coming from
some basic ordering problem (like boundary conditions).

    Stephen

-----Original Message-----
From: Dave Nystrom [mailto:address@hidden]
Sent: Monday, June 25, 2001 10:22 PM
To: pooma-dev
Subject: [pooma-dev] Problems running in parallel

     John and I are trying to get our stuff running in parallel and we are
having runtime problems.  We are using KCC-4.0e with the 4.0f prelinker on
RedHat Linux 7.1.  We are using Cheetah 1.0 sitting on top of mm-1.1.1
i.e. we used --shmem --nompi on the Cheetah configure script.  We
successfully ran the tests for mm and Cheetah.  We built Pooma 2 using the
--messaging option to select the use of Cheetah.  We built tests in the Array
and NewField directory and ran them.  For these tests, we were doing debug
builds with KCC without exceptions and for the Cheetah case.  We ran about
half of the Array tests and they all reported passing.  We ran all of the
tests for the NewField case and all passed except one.  The one that failed
was FieldTour1 which is the only one that seemed to have any Cheetah specific
code in it.  The command line that we were using was:

     FieldTour1 -shmem -np 4

Also, the following command line failed:

     FieldTour1 -shmem -np 1

How should we proceed on this?  Are there examples in the Pooma 2 source tree
which demonstrate how to use the NewField Fields in parallel that we can look
at?  Are there any Pooma 2 tests that exercise running in parallel that pass?
John and I would like to get our stuff running in parallel before I leave for
vacation.  Right now, we only have 3 more days in which our schedules
overlap.

     Also, it turns out that John used the FieldTour1.cpp example as a guide
for building our Fields for the parallel case.  When we run our test problem,
we are getting some pretty strange behavior - it seems that we are not even
getting reproducible results from one run to the next and only about half of
the field is getting the correct answer.  Our test problems run correctly for
the serial case i.e. no Cheetah and CompressibleBrick.

--
Dave Nystrom                    email: address@hidden
LANL X-3                        phone: 505-667-7913     fax: 505-665-3046