Add a new topo module: Treematch #448
Conversation
|
Sent from
|
Refer to this link for build results (access rights to CI server needed): |
|
this is great addon! |
at Inria Bordeaux. This allows us to take advantage of the remap capability of MPI to rearrange the ranks beased on the weights povided by the application.
counting or not the collective traffic as a separate entity. The need for such a PML is simply because the PMPI interface doesn't allow us to identify the collective generated traffic.
|
Once pushed in the master we can add all the features we want, including additional output information and per rank output files (as long as we update the scripts to be able to concatenate these files). Regarding extending the same mechanism to the coll framework, there is no need as the coll framework has an automatic fallback to the next module, allowing us to implement the same functionality without the need for any tricks. |
|
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
@bosilca Any idea why the monitoring test is failing at Jenkins? |
|
@bosilca This monitoring PML is an interesting concept. Is it conceptually different than a PMPI-based tool? Also, what is the mechanism to select the underlying ("real") PML? E.g., if I want to use the use of cm or ob1, can I do that? |
|
On Fri, Mar 27, 2015 at 8:59 AM, Jeff Squyres notifications@github.com
|
|
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
…g all available credentials, and letting the receiver use the highest priority one they have in common.
sm dstore: set pmix segment size to proper value
|
I'm getting a failure in the monitoring_test when doing make check. Should this test really be run as part of make check? |
Support authentication across security domains
|
@hppritcha +1 -- I'm not sure we do. @bosilca thoughts? |
|
We are using a hidden trick to do some ugly things, and it appears that the extremely picky compile flags expose it out. The long story is that we need (wanted) to expose a flush function, to allow the application to dump files at desired intervals (allowing us to identify the stages in a parallel application). Because there is no "legal" way to interact with a component, we tried to expose this internal function pointer as an read-only MCA parameter, but apparently the compiler disagrees. |
|
It looks like treematch/treematch is a wholesale import of a separate library. As such, my comment about no copyrights in the individual treematch/treematch/*.[ch] files is moot -- but we should have some kind of top-level license/copyright for that directory somewhere. And we do need to make sure that the license is compatible with OMPI's overall license. |
|
George are you talking about the monitor_test? |
|
Let me take this one back. The compiler expose our trick but the compilation step succeed. What fails is that we need at least 2 processes to run this test, and that "make check" doesn't run parallel tests. So I think that @hppritcha is right and we should not run time test as part of "make check". |
|
This library is developed by researchers that signed the OMPI collaboration agreement (Inria Bordeaux). Moreover the Cecill-B license is another derivative of the BSD and 100% compatible with it. The main difference is the requirement for citation (http://www.cecill.info/faq.en.html#bsd). |
|
They seem to be used in the PriorityQueue which is then used in the k-partitioning.c. |
|
k-partitioning.c isn't actually used in tm_kpartitioning.c because USE_KL_KPART is 0. We'll check things more in depth after the week-end and confirm everything. |
|
@bgoglin thanks for the tip. Things seems to work correctly after removing the fibo.[ch], PriorityQueue.[ch] and k-partitioning.[ch]. |
|
|
Refer to this link for build results (access rights to CI server needed): |
|
@bgoglin Any progress on the license issue? |
|
Sorry for the delay, they just confirmed that putting everything in BSD3 (the one in the upstream repo COPYING https://gforge.inria.fr/scm/viewvc.php/COPYING?view=markup&root=treematch) is OK after removing those files. |
|
|
Refer to this link for build results (access rights to CI server needed): |
|
Can this PR be merged now that legal issues have been resolved and problematic code has been removed? |
|
Can one of the admins verify this patch? |
|
Any news about merging this PR ? |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
looks like this PR needs work. Let's see if the MLNX jenkins doesn't like it either. |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
bot:retest |
|
will check why it ignored prev command from @hppritcha |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
Refer to this link for build results (access rights to CI server needed): Build Log Test FAILed. |
|
@bosilca Can you have a look at all these Jenkins failures? They look legit. |
|
@bosilca ping -- these jenkins errors look legit. Can you have a look? |
Bring over the hwloc assembly updates
7 participants
This PR contains two commits that are combined into a single PR as they are related. First, it adds a new topo module allowing for process reordering based on user-provided weights. Second, it adds a new overlay PML (one that works in combination with another PML), allowing to accurately track the amount of data send between processes.