First revision of the Locatation Aware Mapping Algorithm (LAMA) RMAPS
component. This component is used to effect many different types of
regular of process/processor affinity patterns. Although quite
flexible in the patterns that it provides, it is ''not'' a
fully-arbitrary, rankfile-like solution for process/processor
affinity.
Inspiried by !BlueGene-like network specifications, LAMA has a core
algorithm that is quite good at specifying regular patterns in
multiple "dimensions" (where "dimensions" are expressed in terms of
different hardware elements: processor hardware threads, cores,
sockets, ...etc.). The LAMA core algorithm is described here:
http://www.open-mpi.org/papers/cluster-2011-lama/
= LAMA Usage Levels =
LAMA allows specifying affinity multiple different ways:
1. None: Speciying no affinity options to mpirun results in exactly
the same behavior as today: no affinity is used.
1. Simple: Using the mpirun options "--bind-to <WIDTH>" and "--map-to
<LEVEL>" to indicate how "wide" each process should be bound
(i.e., bind to a processor core, or to a processor socket, etc.)
and how to lay out the processes (i.e., round robin by cores,
sockets, etc.).
1. Expert: Using four new MCA parameters to effect process mapping
and binding to processors. These options are a bit complex, and
are not for the faint at heart, but offer a high degree of
(regular pattern) flexibility (each of these are described more
fully below):
* rmaps_lama_map: a sequence of characters describing how to lay
out processes
* rmaps_lama_bind: a sequence of characters describing the
resources to bind to each process
* rmaps_lama_mppr: a sequence of characters describing the maximum
number of processes to allow per resource (i.e., a specific
definition of "oversubscription")
* rmaps_lama_ordering: once all processes are in place, how to
order the ranks in MPI_COMM_WORLD
We anticipate that most users will utilize the "None" and "Simple"
levels of affinity, and they continue to work just as they do with the
v1.6 series and SVN trunk.
The Expert level was designed for two purposes:
1. To provide a precise definition for the "Simple" level (i.e.,
every
--bind-to/--map-by option in the "Simple" level has a
corresponding
precise specification in the "Expert" level)
1. As modern computing platforms become more complex, we simply
cannot predict what application developers will need in terms of
processor affinity. LAMA is an attempt to provide a highly
flexible mechanism that allows applications to utilize a variety
of complex, unique affinity patterns beyond the common "bind to
core" and "bind to socket" patterns.
= LAMA Simple Level =
The "Simple" level is pretty much the same as what Open MPI has
offered for years. It supports the same --bind-to and --map-by
options that Open MPI has supported for a while, but expands their
scope a bit.
Specifically, the following options are available for both --bind-to
and --map-by:
* slot
* hwthread
* core
* l1cache
* l2cache
* l3cache
* socket
* numa
* board
* node
= LAMA Expert Level =
The "Expert" level requires some explanation. I'll repeat my
disclaimer here: the LAMA Expert level is not for the meek. It is
flexible, but complex. '''Most users won't need the Expert level.'''
LAMA works in three phases: mapping, binding, and ordering. Each is
described below.
== Expert: Mapping ==
Processes are paired with sets of resources. For example, each
process may be paired with a single processor core. Or each process
may be paired with an entire processor socket. LAMA performs this
mapping, obeying the Max Processes Per Resource ("MPPR", pronounced
"mipper") limits. More on MPPR, below.
Mapping can be performed across multiple hardware levels:
* h: Hardware thread
* c: Processor core
* s: Processor socket
* L1: L1 cache
* L2: L2 cache
* L3: L3 cache
* N: NUMA node
* b: Processor board
* n: Server node
If the act of mapping is that of pairing MPI processes to the
resources that have been allocated to a job, one can easily imagine
looping through all the resources and assigning processes to them.
But to effect different process process layout patterns across those
resources, one may want to loop over those resources ''in a different
order.'' That is, if the above-mentioned nine hardware resources
(hardware thread, processor core, etc.) can be thought of as an
nine-dimensional space, you can imagine nine nested loops to traverse
all of them. And you can imagine that changing the order of nesting
would change the traversal pattern.
LAMA accepts a sequence of tokens representing the above-mentioned
nine hardware resources to specify the order of looping when mapping
resources to processes.
For example, consider a "simple" traversal: csL1L2L3Nbnh. Reading
that sequence of letters from left-to-right, it specifies mapping by
processor core, processor socket, L1 cache, L2 cache, L3 cache, NUMA
node, processor board, server node, and finally hardware thread.
Wait... what? That string specifies resources from "smallest" to
"largest" -- with the exception of hardware threads. Why are they
tacked on to the end?
In short, this string of letters means "map by round robin by core" --
(indeed, it exactly corresponds to the Simple level "--map-by core").
Specifically, LAMA traverses the string from left-to-right and maps
processes to all the resources indicated by that token (e.g., "c" for
processor core). When there are no more resources indicated by that
token, it goes on to the next token.
Hence, in this case, LAMA will map the first process to the first
core, then it will map the second process to the second core, and so
on.
Once all the cores are exhausted, LAMA effectively ignores all the
other letters until "h" (because all the other resources are made up
of cores; when cores are exhausted, those resources are exhausted,
too).
If there are still more processes to be mapped, LAMA will then
traverse all the hyperthreads -- meaning that the next process will be
mapped to the second hyperthread on the first core. And the next
process will be mapped to the second hyperthread on the second core.
And so on.
Keep in mind that the cores involved may span many server nodes; we're
not just talking about the cores (etc.) in a single machine.
As another example, the sequence "sL1L2L3Nbnch" is exactly equivalent
to "--map-by socket" (i.e., LAMA maps the first process to the first
socket, the second process to the second socket, and so on).
The sequence of letter can be combined in many, many different ways to
produce many different regular mapping patterns.
=== Max Processes Per Resource (MPPR) ===
The MPPR is an expression that precisely defines the maximum number of
processes that can be mapped to any single resource. In effect, it
defines the concept of "oversubscription." Specifically, traditional
HPC wisdom is that "oversubscription" is when there is more than one
MPI process per processor core.
This conventional defintion is expressed in a MPPR string of "1:c"
(one process per core).
But what if your MPI processes are multi-threaded, and they need
multiple processes per core? You'd need a different description of
"oversubscription" in this case. Perhaps you want to have one MPI
process per socket. This would be expressed in a MPPR string of
"1:s".
The general form of an individual MPPR specification is an integer
follow by a colon, followed by any of the tokens from mapping can be
used in the MPPR specification. For example "1:c" is pronounced "one
process per core."
Multiple MPPR specifications can be strung together into a
comma-delimited list, too. All of these MPPR values and then taken
into account when mapping. Here's some examples:
* 1:c -- allow, at most, one process per processor core (i.e., don't
schedule by hyperthread)
* 1:s -- allow, at most, one process per processor socket (e.g.,
that process may be multithreaded, or wants exclusive use of the
socket's caches)
* 1:s,2:n -- only allow one process per processor socket, but, at
most, two processes per server node (e.g., if the two MPI processes
will consume all the RAM on the server node, even if there are more
processor cores available)
If mapping all processes to resources would exceed a MPPR limit, this
job is ruled to be oversubscribed. If --oversubscribe was specified
on the mpirun command line, the job continues. Otherwise, LAMA will
abort the job.
Additionally, if --oversubscribe is specified, LAMA will endlessly
cycle through the mapping token string untill all processes have been
mapped.
== Expert: Binding ==
Once processes have been paired with resources during the Mapping
stage, they are optionally bound to a (potentially different) set of
resources. For example, processes may be mapped round robin by
processor socket, but bound to an individual processor core.
To be clear: if binding is not used, then mapping is effectively
reduced to "counting how many processes end up on each server node."
Without binding, there's no enforcement that a process will stay where
LAMA thinks it was placed.
With binding, however, processes are bound to a set of hardware
threads. The number of threads to which the process is bound is
sometimes referred to as the "binding width". For example, if a
process is bound to all the hardware threads in a processor socket,
its "width" is the processor socket.
(note that we specifically do not say that the hardware threads are
sequential, even if they are all within a single resource such as a
processor core or socket. BIOS ordering of hardware threads can be
wonky; so we only refer to "sets of hardware threads")
Bindings are expressed as an integer and a token from the mapping
string. For example "1s" means "bind each process to one processor
socket" (there is no ":" in the binding string because the ":" is
pronounced as "per" when reading the MPPR string).
Note that it only makes sense to bind processes to a single resource
specification (unlike the MPPR specification, where multiple limits
can be specified).
== Expert: Ordering ==
Finally, processes are assigned a rank in MPI_COMM_WORLD. LAMA
currently offers two ordering modes: sequential or natural:
* Sequential: if you laid out all the hardware resources in a single
line, and then overlaid all the MPI processes on top of them, they
are ordered from 0 to (N-1) from left-to-right.
* Natural: the ordering of ranks follows the mapping ordering. For
example, consider a server node with two processor sockets, each
containing four cores. The command line "mpirun -np 8 --bind-to
core --map-by socket --order n a.out" would result in MCW ranks
that look like this: [0 2 4 6] [1 3 5 7].
= Execution =
At this point, the job is fully mapped, optionally bound, and its
ranks in MPI_COMM_WORLD are ordered. It now starts its execution.
= Final Notes =
Note that at this point, lama is not the default mapper. It must be
activiated with "--mca rmaps lama". We'll continue to do further
testing and comparitive analysis with the current set of ORTE mappers.
Also, note that the LAMA algorithm can handle heterogeneity between
hardware resources (e.g., an MPI job spanning server nodes with
differing numbers of processor sockets). For lack of a longer
explanation (this commit message already long enough!), LAMA considers
each server node individually during mapping and binding.
See the LAMA paper for more details:
http://www.open-mpi.org/papers/cluster-2011-lama/
This commit was SVN r27206.
Remove some stale configure.m4's we no longer need.
Optimize the nidmaps a bit by only sending info that has changed each time, instead of sending a complete copy of everything. Makes no difference for the typical MPI job - only impacts things like staged execution where we are sending multiple (possibly many) launch messages.
This commit was SVN r27165.
"num_app_ctx" - the number of app_contexts in the job
"first_rank" - the MPI rank of the first process in each app_context
"np" - the number of procs in each app_context
Still need clarification on the MPI_Init portion of the ticket. Specifically, does the ticket call for returning an error is someone calls MPI_Init more than once in a program? We set a flag to tell us that we have been initialized, but currently never check it.
This commit was SVN r27005.
gone. So exit with 0, not ORTE_ERROR_DEFAULT_EXIT_CODE (which is 1).
This fixes a race condition in the rsh launcher upon termination,
where ORTE would sometimes think that a daemon failed to launch.
This commit was SVN r26935.
* Fix a typo introduced by the removal of the notifier framework
* Fix to flush the modex cached data correctly using the orte DB API.
This commit was SVN r26773.
Among other things, this patch deals with the following issues:
* fix ompi-checkpoint argument parsing
* ompi-restart -showme prints an extraneous "Restarted child with PID"
message. Move around the debug statement to avoid this.
* fixes for the state machine changes
This commit was SVN r26770.
Update all the orte ess components to remove their associated APIs for retrieving proc data. Update the grpcomm API to reflect transfer of set/get modex info to the db framework.
Note that this doesn't recreate the old GPR. This is strictly a local db storage that may (at some point) obtain any missing data from the local daemon as part of an async methodology. The framework allows us to experiment with such methods without perturbing the default one.
This commit was SVN r26678.
no longer using the runtime provided by the reference implementation.
Remove the Catamount support from ORTE, since we're no longer supporting
Catamount. Left the Catamount timer component, because I'm not sure whether
it's used on the XTs running CNL.
This commit was SVN r26677.
* Add new configure command line options and deprecate some old ones:
* --with-verbs replaces --with-openib
* --with-verbs-libdir replaces --with-openib-libdir
* If you specify --with-openib[-libdir] without
--with-verbs[-libdir], you'll get a "these options have been
deprecated!" warning, but then they'll act just like
--with-verbs[--libdir].
'''Sidenote:''' Note that we are not renaming any components at this
time, nor are we renaming the top-level OMPI_CHECK_OPENIB m4 macro
(which is pretty strongly tied to the openib BTL and is bastaridzed
by the ofud BTL). Note that there will likely be more changes in
this area coming soon (next week?) when some long-standing changes
move to the SVN trunk: some openib BTL infrastructure will move to
ompi/mca/common, and its configury gets split up / refactored.
We extend our philosophy of other --with-<foo> configure options of
--with-verbs to ''all'' verbs-lovin components:
* If you specify --with-verbs, then all verbs-lovin' components must
configure successfully (or abort). This currently means: OOB ud,
BTL ofud, BTL openib.
* If you specify --with-verbs=DIR, then all verbs-lovin' component
must configure successfully (or abort), and will use DIR to find
verbs headers and libraries.
* If you specify --without-verbs, then all verbs-lovin' components
will be ignored.
This commit also fixes a problem where the --with-openib=DIR form
would not use DIR for ''all'' verbs-lovin' components (I think only
BTL openib and BTL ofud used that DIR). Now all of them do, as does
hwloc (because hwloc has some !OpenFabrics helper functions that
require ibv types from verbs.h).
There's a little new m4 infrastructure worth mentioning:
* If you create a new verbs-lovin' component (i.e., a component that
need verbs), your configure.m4 should
AC_REQUIRE([OPAL_CHECK_VERBS_DIR]).
* You can then use three global shell variables: $opal_want_verbs,
$opal_verbs_dir, $opal_verbs_libdir, which will be set as follows:
* opal_want_verbs will be "yes" and opal_verbs_dir and
opal_verbs_libdir will both be set to directory values, '''OR'''
* opal_want_verbs will be "no" and opal_verbs_dir and
opal_verbs_libdir will both be set empty
This commit was SVN r26640.
1. they modified the OMPI-ORTE interface, which is something I promised to avoid doing unless absolutely necessary, and
2. the framework ident is already in the component name key provided to the modex db. What is missing is the project ident, but as Jeff and I discussed last week, we really need to add that field to the component struct anyway to avoid multi-project collisions on framework names. That will be done over the next couple of weeks as a separate effort.
This commit was SVN r26613.
The following SVN revision numbers were found above:
r26600 --> open-mpi/ompi@5ba4deff07
r26601 --> open-mpi/ompi@0e3094c318
Restore enable-static-ports option by default - the Cray will have to disable it to get around their library issues, but that's just a warning problem as opposed to blocking the build.
This commit was SVN r26606.
readable by normal human beings, vs. having a bitmap of physical
PU's). Use the new hwloc base prettyprint functions to generate the
output.
This commit was SVN r26533.
exits with status 77, the whole job will be killed, but mpirun will
still return an exit status of 77, so MTT will report it as a skip
anyway.
This commit was SVN r26445.
The following SVN revision numbers were found above:
r26413 --> open-mpi/ompi@02aa36f2e5
nonzero status (we polled other MPI implementations since one one in
the OMPI community had a concrete opinion on what behavior to do here
-- all other MPI's seem to adhere to this behavior, too).
This commit adds an MCA parameter that allows us to tell ORTE to
''not'' kill jobs when a process exits with a status of 77, meaning
the GNU testing standard of "this test was skipped". In all the OMPI
tests, all procs will either return 77 or not. So if they all return
77, mpirun won't consider it an error, but will still return an exit
status of 77 (so that MTT can know that the test was cleanly skipped).
This commit was SVN r26413.
* Remove paffinity, maffinity, and carto frameworks -- they've been
wholly replaced by hwloc.
* Move ompi_mpi_init() affinity-setting/checking code down to ORTE.
* Update sm, smcuda, wv, and openib components to no longer use carto.
Instead, use hwloc data. There are still optimizations possible in
the sm/smcuda BTLs (i.e., making multiple mpools). Also, the old
carto-based code found out how many NUMA nodes were ''available''
-- not how many were used ''in this job''. The new hwloc-using
code computes the same value -- it was not updated to calculate how
many NUMA nodes are used ''by this job.''
* Note that I cannot compile the smcuda and wv BTLs -- I ''think''
they're right, but they need to be verified by their owners.
* The openib component now does a bunch of stuff to figure out where
"near" OpenFabrics devices are. '''THIS IS A CHANGE IN DEFAULT
BEHAVIOR!!''' and still needs to be verified by OpenFabrics vendors
(I do not have a NUMA machine with an OpenFabrics device that is a
non-uniform distance from multiple different NUMA nodes).
* Completely rewrite the OMPI_Affinity_str() routine from the
"affinity" mpiext extension. This extension now understands
hyperthreads; the output format of it has changed a bit to reflect
this new information.
* Bunches of minor changes around the code base to update names/types
from maffinity/paffinity-based names to hwloc-based names.
* Add some helper functions into the hwloc base, mainly having to do
with the fact that we have the hwloc data reporting ''all''
topology information, but sometimes you really only want the
(online | available) data.
This commit was SVN r26391.
Fix the state machine to support multiple jobs being simultaneously launched as this is not only required for mapreduce, but can happen under comm-spawn applications as well.
This commit was SVN r26380.
it succeeds and run $1 or $2, accordingly. This allows "make dist" to
run properly on machines that do not have OpenFabrics stuff installed
(e.g., the nightly tarball build machine).
There's still more to be done here -- it doesn't check for non-uniform
directories where the OpenFabrics headers/libraries might be
installed. We might need to re-tool/combine
ompi/config/ompi_check_openib.m4 (which checks for way more than
oob/ud needs) and move it up to config/ompi_check_ofa.m4, or
something...?
This commit was SVN r26350.
