The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
This merge adds Checkpoint/Restart support to Open MPI. The initial
frameworks and components support a LAM/MPI-like implementation.
This commit follows the risk assessment presented to the Open MPI core
development group on Feb. 22, 2007.
This commit closes trac:158
More details to follow.
This commit was SVN r14051.
The following SVN revisions from the original message are invalid or
inconsistent and therefore were not cross-referenced:
r13912
The following Trac tickets were found above:
Ticket 158 --> https://svn.open-mpi.org/trac/ompi/ticket/158
some issues with the C #defines SEEK_{SET, END, POS}. The workaround
involves some hackery that should work in almost every common use case
for the C stdio constants (and all the legal issues of the MPI constants).
The one issue is that the C stdio constants are now const ints instead
of #defines, which means that #ifdef checks will fail for the constants.
Behavior can be disabled at either configure time or build time.
Refs trac:387
This commit was SVN r12121.
The following Trac tickets were found above:
Ticket 387 --> https://svn.open-mpi.org/trac/ompi/ticket/387
Add --enable-orterun-prefix-by-default (and a synonym:
--enable-mpirun-prefix-by-default) to make orterun always behave as if
"--prefix $prefix" was given on the command line (where $prefix is the
value given to the --prefix option to configure). This prevents many
rsh/ssh users from needing to modify their shell startup files to set
the LD_LIBRARY_PATH for Open MPI (they will still need to set PATH or
otherwise find the OMPI executables to mpicc/mpirun/etc. their MPI
applications).
Also added --noprefix option to orterun to disable this behavior.
Finally, note that even if --enable-orterun-prefix-by-default is
specified, if the user specifies --prefix or /path/to/mpirun, these
options will override the default value of the prefix ($prefix).
This commit was SVN r11669.
The following Trac tickets were found above:
Ticket 377 --> https://svn.open-mpi.org/trac/ompi/ticket/377
- Make the F90 bindings compile and link properly with gfortran 4.0,
4.1, Intel 9.0, PGI 6.1, Sun (don't know version offhand -- the most
current as of this writing, I think), and NAG 5.2, although some
have limitations (e.g., NAG can't seem to handle the medium and
large sizes)
- Building the F90 "small" module size is now the default, even for
developers
- Split up mpif.h into multiple files because parts of it were toxic
to the F90 bindings
- Properly specify unsized/unshaped arrays to make the bindings work
on all known compilers
- Make ompi_info show Fortran 90 bindings size
- XML somewhat lags the generated scripts as of this commit, but
functionality was my main goal -- the XML can be updated later (if
at all).
This commit was SVN r10118.
in San Jose. Allow the configure option --disable-binaries to build OMPI,
but not build or install the support binaries (so basically, just build
the libraries).
This commit was SVN r9777.
svn merge -r 9453:9609 https://svn.open-mpi.org/svn/ompi/tmp/f90-stuff .
Several improvements over the current F90 MPI bindings:
- The capability to make 4 sizes of the F90 bindings:
- trivial: only the F90-specific MPI functions (sizeof and a few
others)
- small: (this is the default) all MPI functions that do not take
choice buffers
- medium: small + all MPI functions that take one choice buffer
(e.g., MPI_SEND)
- large: all MPI functions, but those that take 2 choice buffers
(e.g., MPI_GATHER) only allow both buffers to be of the same type
- Remove all non-standard MPI types (LOGICAL*x, CHARACTER*x)
- Remove use of selected_*_kind() and only use MPI-defined types
(INTEGER*x, etc.)
- Decrease complexity of the F90 configure and build system
This commit was SVN r9610.
complete, but stable enough that it will have no impact on general development,
so into the trunk it goes. Changes in this commit include:
- Remove the --with option for disabling MPI-2 onesided support. It
complicated code, and has no real reason for existing
- add a framework osc (OneSided Communication) for encapsulating
all the MPI-2 onesided functionality
- Modify the MPI interface functions for the MPI-2 onesided chapter
to properly call the underlying framework and do the required
error checking
- Created an osc component pt2pt, which is layered over the BML/BTL
for communication (although it also uses the PML for long message
transfers). Currently, all support functions, all communication
functions (Put, Get, Accumulate), and the Fence synchronization
function are implemented. The PWSC active synchronization
functions and Lock/Unlock passive synchronization functions are
still not implemented
This commit was SVN r8836.
--enable-mpi-{cxx,f77,f90} so that people aren't confused about what
they are actually disabling.
This should go to the 1.0 branch
This commit was SVN r7851.
- Add --enable-trace which turns on some internal tracing and dumps a
file per process in the session directory tree. Meant for internal
developer tracing, *not* for tracing MPI applications in the
traditional sense.
This commit was SVN r7229.
which case, skip it, since it takes so bloody long to compile)
* Dsiable the XGrid PLS when compiling in 64 bit mode, as Tiger only
ships with XGrid libraries for 32bit apps
* ompi_config.h and orte_config.h (and supporting headers) are now only
installed if --with-devel-headers is enabled. Since they are no
longer needed for MPI applications, it doesn't make sense to install
them if we are only installing mpi.h and mpif.h.
Also, since we are no longer including ompi_config.h in mpi.h, there
is no longer a need to do the dumb sed trick on install
This commit was SVN r7042.
- fix the --with-f90-max-array-dim configure switch
- fix configure test to find the supported f90 linker switch to find
fortran modules
- Unbelievably, some versions of sh (cough cough Solaris 9 cough
cough) actually seem to internally perform a "cd" into a
subdirectory when you run "./foo/bar", such that if you try to
source a script in the top-level directory in the bar script (i.e.,
". some_script" in the bar script), it will try to run it in the
"foo" subdirectory, rather than the top-level directory! #$@#$%#$%
So we have to pass in the pwd to the scripts so that they know where
some_script is.
- Reworked much of ompi/mpi/f90/Makefile.am for lots of reasons. See
the internal comments (mostly having to do with dependency stuff --
Libtool does not apparently support F90, so we can only build the
F90 library statically.
This commit was SVN r6993.
of flags to configure. Now don't need to specify the contrib/platform
part of the path if you don't want to
* Add "optimized" platform setting that will undo all the performance-
affecting things that a developer build sets up.
This commit was SVN r6946.
support in OMPI. Currently only enables/disables the architecture
sharing modex in ob1 pml.
* Add sds framework to ompi_info
* Figure out table ids to use for Portals BTL at configure time, since
we should use 30 & 31 on Red Storm, but the reference implementation
only supports 0-8.
* Some bug fixes in Portals UTCP sds
This commit was SVN r6650.
* Add ability to completely disable libltdl (the dlopen code to load
dynamic shared objects) to configure: --disable-dlopen
* Added MCA param (component_disable_dlopen) to disable DSO loading
at runtime
* Made the event library behave in some not-completely-erroneous way
on platforms where it has absolutely no eventops support (ie, no
select, poll, or epoll)
* Disabled orte_wait, opal_few, and opal_daemon_init code on
platforms without fork, waitpid support. All non-init functions
will return OPMI_ERR_NOT_SUPPORTED
* Disable orteprobe tool when fork or pipe aren't supported
This commit was SVN r6490.
sockaddr_in - seems to be a good indicator)
* disable util/if code if no inet devices (again, no sockaddr_in)
* add enable/disable flag to disable stacktrace pretty-print code
(defaults to enabled). Seems there's something funky going on with
the preprocessor on Red Storm that was causing problems - this was
the easiest fix
* clean up a bunch of the configure.m4 files to remove bogus comments,
properly comment them, fix the dumb logic for happy/unhappy
* Create a macro for testing both header and library for a package,
since we seem to do this kind of test quite often. Handles the
-I and -L search paths properly (including stripping out /usr and
/usr/local if not needed)
* Converted mvapi components to configure.m4, using the nice new
ompi_check_package macro (above)
This commit was SVN r6454.
having it in config/ompi_setup_cxx.m4
2. Adding --enable-coverage option. This will add teh flags -ftest-coverage
and -fprofile-arcs to the flags. Also, one needs to compile with
debug and static only to enable code coverage
3. Adding the coverage flag options to WRAPPER_*_FLAGS so that mpicc and co.,
will add these to teh executables when they are compiled
This commit was SVN r5416.
dimensions the f90 MPI bindings should support (they are strongly
typed, and the number of dimensions of choice arguments must be
specified -- it cannot be arbitrary). The default is four.
Note that even though increasing this value has essentially a linear
effect on the code, the multiplier constant is fairly large (only a
small number of functions have 2 choice buffers, so the exponential
factor is relatively small). Increasing this value increases the
amount of time f90 compilers will spend compiling src/mpi/f90/mpi.f90
(some compilers will crash if it is too big).
This commit was SVN r5268.
#if OMPI_WANT_MPI2_ONE_SIDED and some automake conditionals. Also had
to add some AC_SUBSTs to eliminate part of mpif.h (otherwise the
"external" statements would have made undefined symbols).
All the MPI-2 one-sided functionality (including the skeleton
top-level MPI API functions that only invoke an MPI exception) can be
re-enabled with --enable-mpi2-one-sided.
This commit was SVN r3802.
have some nasty side effect we don't know about, make it a configure
option for now. Also add a harmless signal handler to the pcm open
(since pcm_open will have a signal handler eventually for SIGCHLD, I
think). Use --enable-event-signals / --disable-event-signals to
control behavior.
This commit was SVN r2748.
