- Move endpoint code back up to BTL
- Use opal_pointer_array_t for bounce buffer to identify local smsg completions.
- Update and reenable sendi
- Create a new endpoint for FMA/BTE transactions (keep local smsg/fma transactions seperate)
- Move reverse get code into btl_ugni_put.c
- Move eager get code into btl_ugni_get.c
- Handle remote SMSG overruns correctly
- Added support for inplace sends
- etc
This commit was SVN r26307.
1. New mpifort wrapper compiler: you can utilize mpif.h, use mpi, and use mpi_f08 through this one wrapper compiler
1. mpif77 and mpif90 still exist, but are sym links to mpifort and may be removed in a future release
1. The mpi module has been re-implemented and is significantly "mo' bettah"
1. The mpi_f08 module offers many, many improvements over mpif.h and the mpi module
This stuff is coming from a VERY long-lived mercurial branch (3 years!); it'll almost certainly take a few SVN commits and a bunch of testing before I get it correctly committed to the SVN trunk.
== More details ==
Craig Rasmussen and I have been working with the MPI-3 Fortran WG and Fortran J3 committees for a long, long time to make a prototype MPI-3 Fortran bindings implementation. We think we're at a stable enough state to bring this stuff back to the trunk, with the goal of including it in OMPI v1.7.
Special thanks go out to everyone who has been incredibly patient and helpful to us in this journey:
* Rolf Rabenseifner/HLRS (mastermind/genius behind the entire MPI-3 Fortran effort)
* The Fortran J3 committee
* Tobias Burnus/gfortran
* Tony !Goetz/Absoft
* Terry !Donte/Oracle
* ...and probably others whom I'm forgetting :-(
There's still opportunities for optimization in the mpi_f08 implementation, but by and large, it is as far along as it can be until Fortran compilers start implementing the new F08 dimension(..) syntax.
Note that gfortran is currently unsupported for the mpi_f08 module and the new mpi module. gfortran users will a) fall back to the same mpi module implementation that is in OMPI v1.5.x, and b) not get the new mpi_f08 module. The gfortran maintainers are actively working hard to add the necessary features to support both the new mpi_f08 module and the new mpi module implementations. This will take some time.
As mentioned above, ompi/mpi/f77 and ompi/mpi/f90 no longer exist. All the fortran bindings implementations have been collated under ompi/mpi/fortran; each implementation has its own subdirectory:
{{{
ompi/mpi/fortran/
base/ - glue code
mpif-h/ - what used to be ompi/mpi/f77
use-mpi-tkr/ - what used to be ompi/mpi/f90
use-mpi-ignore-tkr/ - new mpi module implementation
use-mpi-f08/ - new mpi_f08 module implementation
}}}
There's also a prototype 6-function-MPI implementation under use-mpi-f08-desc that emulates the new F08 dimension(..) syntax that isn't fully available in Fortran compilers yet. We did that to prove it to ourselves that it could be done once the compilers fully support it. This directory/implementation will likely eventually replace the use-mpi-f08 version.
Other things that were done:
* ompi_info grew a few new output fields to describe what level of Fortran support is included
* Existing Fortran examples in examples/ were renamed; new mpi_f08 examples were added
* The old Fortran MPI libraries were renamed:
* libmpi_f77 -> libmpi_mpifh
* libmpi_f90 -> libmpi_usempi
* The configury for Fortran was consolidated and significantly slimmed down. Note that the F77 env variable is now IGNORED for configure; you should only use FC. Example:
{{{
shell$ ./configure CC=icc CXX=icpc FC=ifort ...
}}}
All of this work was done in a Mercurial branch off the SVN trunk, and hosted at Bitbucket. This branch has got to be one of OMPI's longest-running branches. Its first commit was Tue Apr 07 23:01:46 2009 -0400 -- it's over 3 years old! :-) We think we've pulled in all relevant changes from the OMPI trunk (e.g., Fortran implementations of the new MPI-3 MPROBE stuff for mpif.h, use mpi, and use mpi_f08, and the recent Fujitsu Fortran patches).
I anticipate some instability when we bring this stuff into the trunk, simply because it touches a LOT of code in the MPI layer in the OMPI code base. We'll try our best to make it as pain-free as possible, but please bear with us when it is committed.
This commit was SVN r26283.
algorithm to pairwise exchange instead of the default one. This might
improve the scheduling and relax the pressure on the network.
This commit was SVN r26246.
Roll in the ORTE state machine. Remove last traces of opal_sos. Remove UTK epoch code.
Please see the various emails about the state machine change for details. I'll send something out later with more info on the new arch.
This commit was SVN r26242.
event on the ME. The events we're likely to see are LINK (the ME was
added to the match list), PUT (weird to see first, but means that the ME
was linked to the match list and then matched), or PUT_OVERFLOW, meaning
the message was unexpected.
This commit was SVN r26199.
together, so implement all functions in the MTL interface for all
MTLs. The only places NULL was still being set was for add_comm/del_comm,
and matched probe, both of which are straight forward to implement (or
return ERROR_NOT_IMPLEMENTED, since the PML can't emulate matched probe).
This commit was SVN r26194.
the ompi_message_t structure to properly initialize convertor (the peer
is available in the request in OB1, and wasn't needed when I did the
original implementation).
* Implement matched probe for the Portals4 MTL and add NULL function pointers
for the other MTLs.
* Add add_comm and del_comm functions to portals4 MTL so that direct call
almost works again.
* Add NEWS item that we've implemented matched probe
This commit was SVN r26180.
* split eq into send and receive eqs so that we can control the number
of outstanding events in send eq and ensure we never lose an ack
* Shouldn't ever truncate on short unexpected receive bocks, so don't set
the truncate bit
* Track active vs. waiting for free short unexpected receive blocks so
to ensure an active short unexpected receive block is posted coming out
of flow control. Also allow creation of "temporary" blocks which should
be released once FREE event is received.
* Slight reorganization of some code in preparation for more flow control
work.
This commit was SVN r26174.
This feature can be enabled at compile time with --with-cma passed
to configure.
At runtime it is also necessary to add "--mca btl btl_sm_use_cma 1"
to the mpirun command.
If both CMA and KNEM are compiled in and enabled at runtime then
KNEM will take precedence and CMA will disable itself
This commit was SVN r26134.
- MAJOR! get src descriptor leaks if mca_bml_base_send fails
- minor. descriptor leaked in mca_pml_send_request_start_copy if the btl returns OMPI_ERR_RESOURCE_BUSY.
This commit was SVN r26077.
Uses new CUDA IPC support. Also, a few minor changes in PML to take
advantage of it.
This code has no effect unless user asks for it explicitly via
configure arguments. Otherwise, it is either #ifdef'ed out or
not compiled.
This commit was SVN r26039.
- re-enable sendi
- move smsg common code into btl_ugni_smsg.h
- added new parameters for smsg/eager frags
- use get for frags larger than the smsg_limit
- bug fixes
- code cleanup
This commit was SVN r25897.
Adds a lock to protect the sm pending_sends list from concurrent access
Fixes bug where btl_sm_process_pending_sends would return an item to
the free list and then continue to use it for a little while
cmr:v1.6
This commit was SVN r25878.
The following Trac tickets were found above:
Ticket 2998 --> https://svn.open-mpi.org/trac/ompi/ticket/2998
definitely should not be linking to more than libmpi.la! (remember
that libmpi.la now wholly contains libopen-rte.la, which wholly
contains libopen-pal.la).
This commit was SVN r25843.
So provide a new parameter (can't have too many!) that handles this situation by stripping the prefix from the returned node name. Also do a little cleanup to ensure we cleanly exit from errors, without generating too many annoying messages.
This commit was SVN r25562.
Turns out, this isn't necessarily true. The Cray, for example, launches processes in a toroidal pattern, thus causing the daemons to wind up somewhere other than what we thought. Other environments (e.g., slurm) are also capable of such behavior, depending upon the default mapping algorithm they are told to use.
Resolve this problem by making the daemon-to-node assignment in the affected environments when the daemon calls back and tells us what node it is on. Order the nodes in the mapping list so they are in daemon-vpid order as opposed to the order in which they show in the allocation. For environments that don't exhibit this mapping behavior (e.g., rsh), this won't have any impact.
Also, clean up the vm launch procedure a little bit so it more closely aligns with the state machine implementation that is coming, and remove some lingering "slave" code.
This commit was SVN r25551.
supposed to. I.e., half-baked/not complete stuff.
This commit backs out all of r25545. Sorry folks!
This commit was SVN r25546.
The following SVN revision numbers were found above:
r25545 --> open-mpi/ompi@7f9ae11faf
to make MPI_IN_PLACE (and other sentinel Fortran constants) work on OS
X, we need to use the following compiler (linker) flag:
-Wl,-commons,use_dylibs
So if we're compiling on OS X, test to see if that flag works with the
compiler. If so, add it to the wrapper FFLAGS and FCFLAGS (note that
per a future update, we'll only have one Fortran compiler anyway).
Fixes trac:1982.
This commit was SVN r25545.
The following Trac tickets were found above:
Ticket 1982 --> https://svn.open-mpi.org/trac/ompi/ticket/1982
Brian dealt with this in the past by creating platform files and using "no-build" to block the components. This was clunky, but acceptable when only one organization was using that option. However, that number has now expanded to at least two more locations.
Accordingly, make --without-rte-support actually work by adding appropriate configury to prevent components from building when they shouldn't. While doing so, remove two frameworks (db and rmcast) that are no longer used as ORCM comes to a close (besides, they belonged in ORCM now anyway). Do some minor cleanups along the way.
This commit was SVN r25497.
https://svn.open-mpi.org/trac/ompi/wiki/ProcessPlacement
The wiki page is incomplete at the moment, but I hope to complete it over the next few days. I will provide updates on the devel list. As the wiki page states, the default and most commonly used options remain unchanged (except as noted below). New, esoteric and complex options have been added, but unless you are a true masochist, you are unlikely to use many of them beyond perhaps an initial curiosity-motivated experimentation.
In a nutshell, this commit revamps the map/rank/bind procedure to take into account topology info on the compute nodes. I have, for the most part, preserved the default behaviors, with three notable exceptions:
1. I have at long last bowed my head in submission to the system admin's of managed clusters. For years, they have complained about our default of allowing users to oversubscribe nodes - i.e., to run more processes on a node than allocated slots. Accordingly, I have modified the default behavior: if you are running off of hostfile/dash-host allocated nodes, then the default is to allow oversubscription. If you are running off of RM-allocated nodes, then the default is to NOT allow oversubscription. Flags to override these behaviors are provided, so this only affects the default behavior.
2. both cpus/rank and stride have been removed. The latter was demanded by those who didn't understand the purpose behind it - and I agreed as the users who requested it are no longer using it. The former was removed temporarily pending implementation.
3. vm launch is now the sole method for starting OMPI. It was just too darned hard to maintain multiple launch procedures - maybe someday, provided someone can demonstrate a reason to do so.
As Jeff stated, it is impossible to fully test a change of this size. I have tested it on Linux and Mac, covering all the default and simple options, singletons, and comm_spawn. That said, I'm sure others will find problems, so I'll be watching MTT results until this stabilizes.
This commit was SVN r25476.
Modify the configure logic and the PMI components to accommodate Cray's approach. Refactor the PMI error reporting code so it resides in only one place. Cray actually decided -not- to define the PMI-2 error codes, so we have to use the PMI-1 codes instead. More fun.
This commit was SVN r25348.
zeroes);
if so, use it for bit-operations like opal_cube_dim and opal_hibit.
Implement two versions of power-of-two.
In case of opal_next_poweroftwo, this reduces the average execution
time from 83 cycles to 4 cycles (Intel Nehalem, icc, -O2, inlining,
measured rdtsc, with loop over 2^27 values).
Numbers for other functions are similar (but of course heavily depend
on the usage, e.g. opal_hibit() with a start of 4 does not save
much). The bsr instruction on AMD Opteron is also not as fast.
- Replace various places where the next power-of-two is computed.
Tested on Intel Nehalem Cluster with openib, compilers GNU-4.6.1 and
Intel-12.0.4 using mpi_testsuite -t "Collective" with 128 processes.
This commit was SVN r25270.
* Only print returnable errors when verbose=1. Still print errors when
we're going to abort, since those obviously aren't returnable
This commit was SVN r25213.
* hdr_data now includes opcount and length for all messages, which is the match
bits for long and rndv messages
* Re-add probe implementation
This commit was SVN r25207.
- add 2 new device ids.
- default rq depth to 64, which proved good for large runs.
This commit should be added to cmr:v1.4:reviewer=jsquyres and
cmr:v1.5:reviewer=jsquyres
This commit was SVN r25145.
Global rdmacm_resolve_timeout defaults to 1000 (1000 ms), which is way
too small for even a 16 node x 12 core iwarp cluster in the presence
of drops. Bump up the default to 30000ms.
This commit fixes trac:2860 and should be added to cmr:v1.4:reviewer=jsquyres
and cmr:v1.5:reviewer=jsquyres
This commit was SVN r25144.
The following Trac tickets were found above:
Ticket 2860 --> https://svn.open-mpi.org/trac/ompi/ticket/2860
- slight change in the selection logic of the fs module, which makes
the ompio independent of the file system type (otherwise ompio
would also have required a configure script).
This commit was SVN r25118.
Don't juse include pre-processor macros between two strins ("s1" #if 0 ... "s2")...
Rather print out the epoch as 0 always...
This commit was SVN r25110.
To enable the epochs and the resilient orte code, use the configure flag:
--enable-resilient-orte
This will define both:
ORTE_ENABLE_EPOCH
ORTE_RESIL_ORTE
This commit was SVN r25093.
libs don't seem to propagate correctly under certain circumstances. This makes
hopefully the nightly tests pass.
also, remove the files that should not have been committed in the first place
:-)
This commit was SVN r25085.
specify btl_tcp_if_include because btl_tcp_if_exclude is defaulted to
the loopback devices.
This commit does a few things:
* Introduce a new OPAL MCA base function:
mca_base_param_check_exclusive_string(). It checks to see that the
''user'' does not set two MCA parameters that are mutually
exclusive by checking the source of those MCS param values.
* Use the above function in many BTLs (and the OOB TCP) to ensure
that <foo>_if_include and <foo>_if_exclude are not both specified
''by the user''.
* Re-arrange many of these BTLs to move their MCA registration code
into a separate component_register() function (vs. the
component_open() function).
This code has been nominally reviewed and checked by Ralph, George,
Terry, and Shiqing.
This commit was SVN r25043.
The following SVN revision numbers were found above:
r24976 --> open-mpi/ompi@8f4ac54336
- Added dynamic SL support to xoob
- Fixed seg fault in finalization
- All the code has been moved to separate files: connect/btl_openib_connect_sl.{c,h}
- The new files compilation is conditionalized
This commit was SVN r24991.
btl_tcp_if_include and btl_tcp_if_exclude are specified.
This commit was SVN r24976.
The following Trac tickets were found above:
Ticket 2838 --> https://svn.open-mpi.org/trac/ompi/ticket/2838
they only add the context id to the tag selection of the underlying
messaging meachinsm.
We would like to enable an MTL to maintain its own context data
per-communicator. This way an MTL will be able to queue incoming eager
messages and rendezvous requests per-communicator basis.
The MTL will be allowed to override comm->c_pml_comm member,
since it's unused in pml_cm anyway.
This commit was SVN r24858.
request completion callback
* Use the completion callback pointer to remove all need for opal_progress
calls in the one-sided layer
This commit was SVN r24848.
- Added enable/disable configuration parameter for dynamic SL
- All the dynamic SL code is conditionalized
- Removed libibmad dependency
- Using only one include - ib_types.h (part of opensm-devel package)
- Removed all the macro and data types definitions, using the
existing definitions from ib_types.h instead
- general cleaning here and there
The async mode is not implemented yet - stay tuned...
This commit was SVN r24830.
Everyone will be starting at MIN anyway (until we implement restart of course)
so there's no reason to set the epoch to INVALID and then immediately reset them
to MIN. This way there's less room to make mistakes later.
This commit was SVN r24829.
if-elseif statements to make it compile with OpenMPi again.
Fixes trac:2808
This commit was SVN r24768.
The following Trac tickets were found above:
Ticket 2808 --> https://svn.open-mpi.org/trac/ompi/ticket/2808
Rather than try to support a bunch of lightweight environments like I did
with the Portals3 code, always use the "modex" and hack the grpcomm for
the SHMEM implementation to return the right nid/pid for a remote
process by "magic".
This commit was SVN r24733.
Instead of returning MPI_SUCCESS every time they are called regardless of the status of the call, they should return a value representative of the action. So similar to MPI_Wait/MPI_Test they will return MPI_SUCCESS if the action was successfull, or the value that matches status.MPI_ERROR for the operation if it is unsuccessful.
This was discussed on the [http://www.open-mpi.org/community/lists/devel/2011/03/9109.php ompi-devel list]
This commit was SVN r24551.
No need for any CMRs to 1.5... that was already done in CMR 2728.
This commit was SVN r24545.
The following SVN revision numbers were found above:
r22841 --> open-mpi/ompi@b400b84162
* rename ib_path_rec_service_level -> ib_path_record_service_level
* use mad.h and ib_types.h
* free all resources
* move ibv_post_recv to be just before ibv_post_send
* cleanup and beatify code
This commit was SVN r24507.
be applied. Correct the MPI validation process of the
MPI_Accumulate arguments.
Fix another potential problem not yet reported. If we convert the
MPI datatypes direclty into OPAL datatypes, we will restrict their
number to the locally different types. Which might not be identical
on the remote node, if we are in a heterogeneous environment. So,
for MPI One sided only deal with MPI level types, never simplify
them on OPAL types (at least on the args). The unfortunate
outcome is that we need to create the args for all datatypes.
This commit was SVN r24466.
Update the CMake script for checking mca subdirs.
Add windows support for __attribute__ packed structures.
Define usleep and posix_memalign with equivalent windows functions.
And a few minor fixes, type casts.
This commit was SVN r24429.
- poll() can return POLLRDNORM even if not requested (Solaris bug)
- MIN macro not defined in btl_openib.c
and while we're at it, we clean up the MIN definition in ad_bgl_pset.h
- btl_openib_connect_rdmacm.c was calling rdma_destroy_id() twice
leading to undefined behavior (a hang on Solaris)
This commit was SVN r24356.
code from upper level into btl configure.m4. Changed
prefix from "OMPI" to "BTL" in preprocessor macro. Add
an mca param that shows it has been configured in.
This commit was SVN r24270.
either direct link to these basic predefined types, or a combination of them.
Anyway, the first items in the datatype list belong to OPAL, the second round
are MPI datatypes created by composing basic OPAL datatypes, and the last
batch are mapped datatype (direct correspondance between an OMPI datatype and
an OPAL one such as int -> int32_t).
Modify the op to fit this new scheme.
This commit was SVN r24247.
Apparently, gcc 4.4.x and 4.5.x complain about the ''possibility'' of
us calling free() on a non-heap variable. We know that this case can
never happen because the refcount will absolutely not go to zero
here. We think it may be gcc being a bit too aggressive on the
warnings.
However, since this happens with gcc 4.4.x and 4.5.x, and since gcc
4.5.x ship in RHEL6 and Fedora 14 (and others), someone '''will'''
complain about this in the future, so we might as well code around it
so that we don't have to keep explaining "despite the warning, it's
really ok."
The workaround is pretty simple: just OBJ_RELEASE the values from
ompi_mpi_comm_parent before it is re-assigned to the new
intercommunicator. Then the compiler's static code analysis can't
possibly tell that it's not a heap variable, and we're ok.
So yes, we are still calling OBJ_RELEASE on a non-heap variable. But
free() '''will never be called''' on it because of the refcount.
This commit was SVN r24214.
The following Trac tickets were found above:
Ticket 2669 --> https://svn.open-mpi.org/trac/ompi/ticket/2669
on some systems caused by the definition of malloc in
opal_config_bottom.h getting expanded in the system malloc.h when
OPAL_ENABLE_MEM_DEBUG is set to 1.
This commit was SVN r24210.
verbose statement that shows up when you --mca btl_base_verbose 100.
It clearly states that the openib BTL disqualifies itself when
MPI_THREAD_MULTIPLE is used.
This commit was SVN r24209.
{{{
base/paffinity_base_service.c: In function ‘opal_paffinity_base_cset2mapstr’:
base/paffinity_base_service.c:623: warning: unused variable ‘range_last’
base/paffinity_base_service.c:623: warning: unused variable ‘range_first’
base/paffinity_base_service.c:622: warning: unused variable ‘count’
base/paffinity_base_service.c:622: warning: unused variable ‘m’
}}}
{{{
connect/btl_openib_connect_oob.c: In function ‘init_ud_qp’:
connect/btl_openib_connect_oob.c:1111: warning: control reaches end of non-void function
connect/btl_openib_connect_oob.c: In function ‘init_device’:
connect/btl_openib_connect_oob.c:1235: warning: unused variable ‘i’
connect/btl_openib_connect_oob.c: In function ‘get_pathrecord_sl’:
connect/btl_openib_connect_oob.c:1323: warning: unused variable ‘i’
}}}
This commit was SVN r24196.
New mca parameter is added (ib_path_rec_service_level) - positive value means that we should get the SL from the SA.
This is usable for torus topologies where different SL value is used for different endpoints.
A cache is kept of ib queue pairs used to communicate with the SA for a particular device and port and path record SL values retrieved from that SA.
The interaction with the cache assumes that there are no recursive calls to these routines. This must be solved either by code flow, by using higher level locks, or by adding a locking mechanism to these routines along with some method for avoiding deadlock.
This code use a UD queue pair to talk to the SA, and not need to chmod /dev/infiniband/umad* for use by normal users.
The request to the SA is a SubnAdmGet(), not a SubnAdmGetTable().
In the future we might add a support of a SubnAdmGetTable(), but it will require implementing RMPP (Reliable Multi-Packet Transaction Protocol) and I'm not sure we want to do that.
This patched is based on the work of David McMillen <davem@systemfabricworks.com>.
This commit was SVN r24195.
The patch includes the following:
* Add new mca parameter - btl_openib_max_hw_msg_size - Maximum size (in bytes) of a single fragment of a long message when using the RDMA protocols (must be > 0 and <= hw capabilities).
* If btl_openib_max_hw_msg_size is larger than the maximum hw limitation print error message.
* Change the default openib flags to include only PUT and not GET.
* Print error message if user choose manually GET flag in openib btl.
* In prepare_dst: limit the message size to be the minimum of both endpoint's hw_limitation and the user limitation (if requested).
This commit was SVN r24191.
Somehow they got fixed in the pt2pt implementation, but not the RDMA
implementation. Thanks to Guillaume Thouvenin for finding this issue.
This commit was SVN r24188.
If specified, a comma-delimited list of TCP interfaces. Interfaces
will be assigned, one to each MPI process, in a round-robin fashion
on each server. For example, if the list is "eth0,eth1" and four
MPI processes are run on a single server, then local ranks 0 and 2
will use eth0 and local ranks 1 and 3 will use eth1.
This feature is only useful for environments with virtual ethernet
interfaces on the same network. For example, if eth0 and eth1 are
virtual interfaces to the same NIC on the same subnet, and if the NIC
provides different hardware resources to eth0 and eth1 (not just
different kernel resources), some HOL blocking and congestion issues
can be eased in a modest fashion.
This commit was SVN r24181.
