used at nce (up to one unique collective module per collective function).
Matches r15795:15921 of the tmp/bwb-coll-select branch
This commit was SVN r15924.
The following SVN revisions from the original message are invalid or
inconsistent and therefore were not cross-referenced:
r15795
r15921
switching:
0 0
/ \ \ / \ \
1 \ \ --> 4 \ \
/ \ \ / \ \
3 2 \ 3 2 \
4 1
(duh). The first form is the bmtree suitable for bcast, but the latter is better for reduce.
Updating default decision function accordingly.
This commit was SVN r15422.
- adding linear algorithm with synchronization for gather.
This algorithm prevents congestion at root process, but introduces
synchronization (serializes non-root processes, but allows messages
to arrive from two processes at the same time).
It performed better than binomial and linear algorithms for large message,
and intermediate and large communicator sizes.
- Updating MPI_Gather decision function to reflect performance results
from MX. I will perform more measurements though - so this one can
change.
This commit was SVN r15165.
This "feature" is disabled by default and it should not affect the current performance.
In case when the message size is large and segment size is smaller than eager size for particular interface,
the leaf nodes in generalized reduce function can overflood parent nodes by sending all segments without
any synchronization. This can cause the parent to have HIGH number of unexpected messages (think 16MB
message with 1KB segments for example). In case of binomial algorithm root node always has at least one
child which is leaf, so this can potentially affect the root's performance significantly [Especially in
large communicators where root may have quite a few children (binomial tree for example)].
When the segment size is bigger than the eager size, rendezvous protocol ensures that this does
not happen so it is not necessary.
Originally, the problem was exposed in "infinite" bucket allocator clean up time for "small" segment sizes
(which may explain some "deadlocks" on Thunderbird tests).
To prevent this, we allow user to specify mca parameter "--mca coll_tuned_reduce_algorithm_max_requests NUM"
this limits number of outstanding messages from a leaf node in generalized reduce to the parent to NUM.
Messages are sent as non-blocking synchrnous messages, so syncronization happens at "wait" time.
The synchronization actually improved performance of pipeline and binomial algorithm for large message sizes
with 1KB segments over MX, but I need to test it some more to make sure it is consistent.
Since there is no easy way to find out what is "the eager" size for particular btl, I set the limit to 4000B.
If message/individual segment size is greater than 4000B - we will not use this feature. This variable may
or may not be exposed as mca parameter later...
I did not have any problems running it and both "default" and "synchronous" tests passed Intel Reduce* tests
up to 80 processes (over MX).
This commit was SVN r14518.
Currently 3 algorithms are available:
- non-overlapping, reduce + scatterv, (works for non-commutative operations)
- recursive halving algorithm (copied from basic module)
- ring algorithm (similar to allreduce ring, for large messages)
This commit was SVN r13929.
Algorithm allows user to specify the segment size to be used for computation/communication overlap.
The additional memory requirement for the algorithm is 2 x segment size.
It performed well for (really) large message sizes over MX and it passed intel Allreduce_c and Allreduce_loc_c tests.
This commit was SVN r13832.
- Allreduce algorithms:
- Recursive doubling is used for small messages (up to 10KB) and can be used for
both commutative and non-commutative operations.
Recursive doubling passed OCC, IMB-3.2, Intel (Allreduce_c, Allreduce_loc_c, and
Allreduce_user_c), mpi_test_suite (Allreduce MIN/MAX, and Allreduce MIN/MAX with
MPI_IN_PLACE) tests on TCP up to 36 nodes and MX up to 64 nodes.
- Ring algorithms performs well for larger messages but cannot be used for
non-commutative operations. It passed the same tests as recursive doubling, except
some of the non-commutative tests in Intel benchmarks Allreduce_loc_c and Allreduce_user_c
(which was expected).
- MPI_Allreduce with new decision function passed all of the tests mentioned above.
- Cleaning up coll_tuned_util. Moving isendrecv to static inline just like sendrecv.
This commit was SVN r13252.
- in allgather algorithms I replaces irecv-isend-waitall sequence with
call to ompi_coll_tuned_sendrecv
- most of the functions in util code and allgather decision function conform to 80 character line width.
-
This commit was SVN r13069.
It contains four algorithms:
Bruck (ciel(logP) steps), Recursive Doubling (log(P) for power-of-2 processes), Ring (P-1 steps),
and Neighbor Exchange (P/2 steps for even number of processes).
All algorithms passed occ, IMB-2.3, and intel verification tests from ompi-tests/ for up to 56 processes.
The fixed decision function is based on results collected over MX on the Grig cluster at
the University of Tennessee at Knoxville.
I have also added (and commented out) copy of MPICH2 decision function for allgather
(from their IJHPCA 2005 paper).
This commit was SVN r12910.
- consistent arguments checking (not allowing to select an algorithm which
is not available)
- consistent way of computing the segcount (number of datatypes by segment).
- small cleanups.
- more informative debugging messages.
This commit was SVN r12545.
the default decision functions (for broadcast, reduce and barrier) are based on a
high performance network (not TCP). It should give good performance (really good) for
any network having the following caracteristics: small latency (5 microseconds) and good
bandwidth (more than 1Gb/s).
+ Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most
of the reduce algorithms use a generic tree based function for completing the reduce.
+ Added macros for computing the trees (they are used for bcast and reduce right now).
+ Allow the usage of all 5 topologies.
+ Jelena's implementation of a binary tree that can be used for non commutative operations.
Right now only the tree building function is there, it will get activated soon.
+ Some others minor cleanups.
This commit was SVN r12326.
yes this means it WAS possible for two nodes to choice two different algorithms
(discovered by Doug Gregor and figured out by George)
Also changed some names like size to comsize so we know which sizes we are using where
This should be updated in al versions
This commit was SVN r10601.
- move files out of toplevel include/ and etc/, moving it into the
sub-projects
- rather than including config headers with <project>/include,
have them as <project>
- require all headers to be included with a project prefix, with
the exception of the config headers ({opal,orte,ompi}_config.h
mpi.h, and mpif.h)
This commit was SVN r8985.
be locally completing. for now using synchronous calls until the new functionality is available. then will change
the code to use the new PML send flags.
This commit was SVN r8867.
this was implemented using a chain (tree followed with pipeline) by setting the chain fanout to a factor of size etc but the chain datastructure was fixed in length and if exceeded the topo create returned a null which isn't helpfull in cid next function of comdup...
Anyway two fixes, first we do have a real linear function so changed the decision function and second altered the
topo chain create to force chain fanouts of less than 1 to 1 and fanouts bigger than max to max.
next check in will change chain to dynamically allocd array (reallocable) but we shouldn't ever use a chain fanout for a linear tree anyway.
(lession must rerun all tests for all data sizes when changing decision functions)
This commit was SVN r8662.
(apparently we've been doing this in opal and orte, but not in ompi
yet). All public symbols begin with "ompi_coll_tuned_" (not
mca_coll_tuned_) except the component struct. Now this component
passes the illegal symbol report with no hits.
This commit was SVN r8589.
Lots of misc fixes: printfs->opal_output, handles fanin/out correctly for forced ops
unused vars, correct calculations on meaning of 'msgsize' for decision functions
(varies depending on algorithm), etc
This commit was SVN r8113.
