This is really great news!! I'll test the trunk on our cluster.
Thank you,
Saliya
On Fri, Mar 14, 2014 at 4:44 PM, Jeff Squyres (jsquyres) wrote:
> We just fixed the segv (see
> https://svn.open-mpi.org/trac/ompi/changeset/31073, if you care).
>
> The issue was an errant large array on the sta
We just fixed the segv (see https://svn.open-mpi.org/trac/ompi/changeset/31073,
if you care).
The issue was an errant large array on the stack in debug builds, which would
cause JVMs to run out of stack space.
The fix is on the SVN trunk now; it will be on the v1.7 branch shortly.
On Mar 11,
We haven't figured it out yet - it seems somewhat erratic as your observations
don't match anything we are seeing on our machines. We know the coll/ml
component is causing trouble for Java applications (but nothing else, oddly
enough), but that doesn't match your experience.
On Mar 12, 2014, a
Just checking if there's some solution for this.
Thank you,
Saliya
On Tue, Mar 11, 2014 at 10:54 PM, Saliya Ekanayake wrote:
> I forgot to mention that I tried the hello.c version instead of Java and
> it too failed in a similar manner, but
>
> 1. On a single node with --mca btl ^tcp it went up
I forgot to mention that I tried the hello.c version instead of Java and it
too failed in a similar manner, but
1. On a single node with --mca btl ^tcp it went up to 24 procs before
failing
2. On 8 nodes with --mca btl ^tcp it could go only up to 16 procs
On Tue, Mar 11, 2014 at 5:06 PM, Saliya
I just tested with "ml" turned off as you suggested, but unfortunately it
didn't solve the issue.
However, I found that by explicitly setting --mca btl ^tcp the code worked
on upto 4 nodes with each running 8 procs. If I don't specify this it'll
simply fail even on one node with 8 procs.
Thank yo
Looks like we still have a bug in one of our components -- can you try:
mpirun --mca coll ^ml ...
This will deactivate the "ml" collective component. See if that enables you to
run (this particular component has nothing to do with Java).
On Mar 11, 2014, at 1:33 AM, Saliya Ekanayake wrot
Seems odd - the Java code is passing all tests on my Linux boxes. A quick
glance shows it failing on memcpy on your machine during MPI_Init, which would
make one suspect either an uninitialized variable or something not getting
loaded correctly.
Oscar, Jose? Any thoughts?
Ralph
On Mar 10, 201
Just tested that this happens even with the simple Hello.java program given
in OMPI distribution.
I've made a tarball containing details of the error adhering to
http://www.open-mpi.org/community/help/. Please let me know if I have
missed any info necessary.
Thank you,
Saliya
On Mon, Mar 10,
Greetings, and thanks for trying out our Java bindings.
Can you provide some more details? E.g., is there a particular program you're
running that incurs these problems? Or is there even a particular MPI function
that you're using that results in this segv (e.g., perhaps we have a specific
bu
Hi,
I have 8 nodes each with 2 quad core sockets. Also, the nodes have IB
connectivity. I am trying to run OMPI Java binding in OMPI trunk revision
30301 with 8 procs per node totaling 64 procs. This gives a SIGSEV error as
below.
I wonder if you have any suggestion to resolve this?
Thank you,
S
11 matches
Mail list logo