coll/tuned: fix allgather bruck algorithm

[ggouaillardet]

No description provided.

[ggouaillardet]

@bosilca could you please review this ?

i wrote a test program available at https://gist.github.com/ggouaillardet/d0cf43616bd99ffb7821

in order to avoid a crash and/or incorrect output, i had to replace the lower bound (rlb) with the true lower bound (true_lb)
shall the discussion is coll_basic_reduce.c updated to clarify that LB is the true lower bound ?

[mellanox-github]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/3/

Build Log
last 50 lines

Test FAILed.

[bosilca]

I thought about using true_lb instead of lb in the shift of the base pointer. And the more I think about this the more I have the impression that something is fishy altogether. Imagine a datatype representing a column in a 2D matrix stored by row. Such a datatype will have an extent equal to 1 element, while it will have a size of N elements and a span of N*(N-1) elements. The current way we handle this, by using the extent to compute how much memory we need, cannot handle this case. More thinking needed I guess...

[ggouaillardet]

At first glance, the ddt has an extent of 1 and a true extent of n_(n-1)+1
So if we allocate the full matrix e.g. n columns, we allocate
true_extent+(n-1)_extent = n^2 elements, which is correct.
Did i miss something ?

[mike-dubman]

testing jenkins PR commander:
retest this please

[mellanox-github]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/26/

Build Log
last 50 lines

[...truncated 23083 lines...]
WARNING: The mechanism by which environment variables are explicitly
passed to Open MPI has changed!

Specifically, beginning in the 1.9.x/2.0.x series, using "-x" to set
environment variables is deprecated.  Please use the
"mca_base_env_list" MCA parameter, instead.

For example, this invocation using the old "-x" mechanism:

    mpirun -x env_foo1=bar1 -x env_foo2=bar2 -x env_foo3 ...

is equivalent to this invocation using the new "mca_base_env_list"
mechanism:

    mpirun -mca mca_base_env_list 'env_foo1=bar1;env_foo2=bar2;env_foo3' ...
5/8 dst = 7 8 9
3/8 dst = 7 8 9
4/8 dst = 7 8 9
2/8 dst = 7 8 9
6/8 dst = 7 8 9
1/8 dst = 7 8 9
7/8 dst = 7 8 9
0/8 dst = 7 8 9
+ '[' yes == yes ']'
+ timeout -s SIGKILL 3m oshrun -np 8 --bind-to core -x SHMEM_SYMMETRIC_HEAP_SIZE=1024M --mca spml yoda -mca pml ob1 -mca btl self,vader /var/lib/jenkins/jobs/gh-ompi-master-pr/workspace-3/ompi_install1/examples/oshmem_max_reduction
WARNING: The mechanism by which environment variables are explicitly
passed to Open MPI has changed!

Specifically, beginning in the 1.9.x/2.0.x series, using "-x" to set
environment variables is deprecated.  Please use the
"mca_base_env_list" MCA parameter, instead.

For example, this invocation using the old "-x" mechanism:

    mpirun -x env_foo1=bar1 -x env_foo2=bar2 -x env_foo3 ...

is equivalent to this invocation using the new "mca_base_env_list"
mechanism:

    mpirun -mca mca_base_env_list 'env_foo1=bar1;env_foo2=bar2;env_foo3' ...
--------------------------------------------------------------------------
SHMEM_ABORT was invoked on rank 3 (pid 16071, host=jenkins01) with errorcode -1.
--------------------------------------------------------------------------
[jenkins01:16065] 7 more processes have sent help message help-shmem-api.txt / shmem-abort
[jenkins01:16065] Set MCA parameter "orte_base_help_aggregate" to 0 to see all help / error messages
Build step 'Execute shell' marked build as failure
[BFA] Scanning build for known causes...

[BFA] Done. 0s

Test FAILed.

[mike-dubman]

retest this please

[mellanox-github]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/34/
Test PASSed.

[ggouaillardet]

On 2014/11/14 0:29, bosilca wrote:

When the count is larger than 1 things seems to nicely fit together. But what if the count is 1?

in this case we use only true_extent, which is correct ... unless i
missed something (again)

[mellanox-github]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/42/
Test PASSed.

[ggouaillardet]

@bosilca is it legal to MPI_Type_create_resized with a negative extent ?
if yes, then there is some work ...

the test program at https://gist.github.com/ggouaillardet/9b4deebb3f508be07912 emits warnings and crashes on 2 tasks or more with Open MPI
(fwiw, mpich is ok with 2 tasks, crashes with 3 tasks and produce incorrect results with 4 tasks ...)

[bosilca]

Well ... based on MPI 3.1 page 107 Section 4.1 it is legal to have a
negative extent. This sounds like a lot of fun in perspective (!)

George.

On Fri, Nov 14, 2014 at 3:27 AM, Gilles Gouaillardet <
[email protected]> wrote:

@bosilca https://github.com/bosilca is it legal to
MPI_Type_create_resized with a negative extent ?
if yes, then there is some work ...

the test program at
https://gist.github.com/ggouaillardet/9b4deebb3f508be07912 emits warnings
and crashes on 2 tasks or more with Open MPI
(fwiw, mpich is ok with 2 tasks, crashes with 3 tasks and produce
incorrect results with 4 tasks ...)

—
Reply to this email directly or view it on GitHub
#262 (comment).

[mellanox-github]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/59/
Test PASSed.