oversubscribed on a node. And thus whether to call sched_yield or not.
The value of node->node_slots_inuse does not currently represent the number of
slots actually in use, at the moment. This is actually a bug in the RAS/RMAPS
base components, but the fix for that specific bug is bigger than we want to
address at the moment (but will certianly do so in the near future).
Since we cannot trust this value, use the total number of mapped processes
(which was properly set by the RMAPS component upon mapping -- Just not
properly propagated back to the registry's node segment) from the process
mapping.
In addition to this change I cleaned up a couple of the debug messages. It
seems that TM and RSH are the only two directly effected by this. SLURM
would be if that section of code wasn't currently inactive, but put the fix
in for prosparity.
This commit was SVN r7743.
too distant past
* work around apparently broken handling of max_slots somewhere along
the line by just setting it to 0
Both changes should go to the trunk.
This commit was SVN r7710.
this loop if "nodes" is an empty list. get_first, in this loop context,
allows us to do just that, while get_begin doesn't.
This fixes a --host problem that appeared on the Linux PPC64 build.
This commit was SVN r7703.
Also revised the callbacks to store and utilize local variables to avoid problems where threads modify the global structures. Not sure this totally fixes the problem, but it's a shot - suggested by Josh (and Jeff, I believe).
This commit was SVN r7694.
command:
svn merge -r 7567:7663 https://svn.open-mpi.org/svn/ompi/tmp/jjhursey-rmaps .
(where "." is a trunk checkout)
The logs from this branch are much more descriptive than I will put
here (including a *really* long description from last night). Here's
the short version:
- fixed some broken implementations in ras and rmaps
- "orterun --host ..." now works and has clearly defined semantics
(this was the impetus for the branch and all these fixes -- LANL had
a requirement for --host to work for 1.0)
- there is still a little bit of cleanup left to do post-1.0 (we got
correct functionality for 1.0 -- we did not fix bad implementations
that still "work")
- rds/hostfile and ras/hostfile handshaking
- singleton node segment assignments in stage1
- remove the default hostfile (no need for it anymore with the
localhost ras component)
- clean up pls components to avoid duplicate ras mapping queries
- [possible] -bynode/-byslot being specific to a single app context
This commit was SVN r7664.
it's always ORTE_SUCCESS and sometimes masks real !=ORTE_SUCCESS rc
values.
- Add MCA param pls_tm_want_path_check. If nonzero (the default),
check for the orted in the PATH before each tm_spawn()'ing (doing a
little caching so that we don't hammer on the filesystem -- remember
all the PATH's where we successfully found the orted so that we
don't have to query the filesystem multiple times for a PATH where
we previously found the orted)
- Be sure to opal_argv_split() the pls_tm_orted MCA param
This commit was SVN r7625.
will fire some subscriptions that will eventually result in invoking
terminate_job (i.e., terminate anything that may have been
successfully started by launch).
This commit was SVN r7622.
If you use --prefix and then "-x LD_LIBRARY_PATH", the rsh pls would
take great pains to ensure that PATH and LD_LIBRARY_PATH were setup
correctly on the local and remote nodes, but then the fork pls would
blitely overwrite LD_LIBRARY_PATH with what the user exported (i.e.,
most likely without our prefix). This patch takes care of that -- the
fork pls examines the incoming environment, and if it sees PATH or
LD_LIBRARY_PATH, it re-prefixes those variables.
This commit was SVN r7566.
it is possible that if the receive has been arrived the callback will
be called before recv_buffer_nb() returns. This causes deadlock
as we try to acquire the lock, but already hold it.
This was causing orterun and orteds to stall in certian situations.
Became evident when stress testing dynamics with remote nodes.
This commit was SVN r7543.
- Fix bug identified by users: --prefix may also apply on the local
node; we need to prefix the PATH and LD_LIBRARY_PATH environment
variables before invoking execve()
This commit was SVN r7541.
some orted's to stall on locks in the MPI Dynamics cases. Since it
is not essentual that we call these functions, they can so away.
Unlock the peer lock when aborting. This causes a potential deadlock
in do_waitall [see comment in code]. This was causing orteds to
deadlock at times when the seed had terminated. With proper interleaving
and timing the orted was deadlocking. This seems to have fixed this in
my stress testing with MPI 2 Dynamics.
This commit was SVN r7539.
The NS replica should give out tags that are over ORTE_RML_TAG_DYNAMIC
or it will overlap with other outstanding tags. This overlap was killing
MPI_Comm_spawn when a program tried to use it multiple times (> 3).
With this fix MPI_Comm_spawn is behaving properly.
A program can call it many times in a row with out problem.
NOTE: Not tested for multi-threaded build yet
(A long time debugging for a one liner... :/)
This commit was SVN r7529.
Have the ras_base_schedule_policy MCA parameter working once again. before it
would only do slot based allocation, even if the MCA parameter was set properly.
Currently you can specify to orterun a node allocation by either:
-mca ras_base_schedule_policy node
-bynode
and slot allocation (which is the default) by:
-mca ras_base_schedule_policy slot
-byslot
This commit was SVN r7513.
