NOTE: launch performance will be absolutely awful if you do this with BTLs that aren't configured to modex_recv on first message!
Even with "modex on demand", we still have to do a barrier in place of the modex - we simply don't move any data around, which does reduce the time impact. The barrier is required to ensure that the other proc has in fact registered all its BTL info and therefore is prepared to hand over a complete data package. Otherwise, you may not get the info you need. In addition, the shared memory BTL can fail to properly rendezvous as it expects the barrier to be in place.
This behavior will *only* take effect under the following conditions:
1. launched via mpirun
2. #procs is greater than ompi_hostname_cutoff, which defaults to UINT32_MAX
3. mca param rte_orte_direct_modex is set to 1. At the moment, we are having problems getting this param to register properly, so only the first two conditions are in effect. Still, the bottom line is you have to *want* this behavior to get it.
The planned next evolution of this will be to make the direct modex be non-blocking - this will require two fixes:
1. if the remote proc doesn't have the required info, then let it delay its response until it does. This means we need a way for the MPI layer to tell the RTE "I am done entering modex data".
2. adjust the SM rendezvous logic to loop until the required file has been created
Creating a placeholder to bring this over to 1.7.5 when ready.
cmr=v1.7.5:reviewer=hjelmn:subject=Enable direct modex at scale
This commit was SVN r30259.
pkg{data,lib,includedir}, use our own ompi{data,lib,includedir}, which is
always set to {datadir,libdir,includedir}/openmpi. This will keep us from
having help files in prefix/share/open-rte when building without Open MPI,
but in prefix/share/openmpi when building with Open MPI.
This commit was SVN r30140.
This change contains a non-mandatory modification
of the MPI-RTE interface. Anyone wishing to support
coprocessors such as the Xeon Phi may wish to add
the required definition and underlying support
****************************************************************
Add locality support for coprocessors such as the Intel Xeon Phi.
Detecting that we are on a coprocessor inside of a host node isn't straightforward. There are no good "hooks" provided for programmatically detecting that "we are on a coprocessor running its own OS", and the ORTE daemon just thinks it is on another node. However, in order to properly use the Phi's public interface for MPI transport, it is necessary that the daemon detect that it is colocated with procs on the host.
So we have to split the locality to separately record "on the same host" vs "on the same board". We already have the board-level locality flag, but not quite enough flexibility to handle this use-case. Thus, do the following:
1. add OPAL_PROC_ON_HOST flag to indicate we share a host, but not necessarily the same board
2. modify OPAL_PROC_ON_NODE to indicate we share both a host AND the same board. Note that we have to modify the OPAL_PROC_ON_LOCAL_NODE macro to explicitly check both conditions
3. add support in opal/mca/hwloc/base/hwloc_base_util.c for the host to check for coprocessors, and for daemons to check to see if they are on a coprocessor. The former is done via hwloc, but support for the latter is not yet provided by hwloc. So the code for detecting we are on a coprocessor currently is Xeon Phi specific - hopefully, we will find more generic methods in the future.
4. modify the orted and the hnp startup so they check for coprocessors and to see if they are on a coprocessor, and have the orteds pass that info back in their callback message. Automatically detect that coprocessors have been found and identify which coprocessors are on which hosts. Note that this algo isn't scalable at the moment - this will hopefully be improved over time.
5. modify the ompi proc locality detection function to look for coprocessor host info IF the OMPI_RTE_HOST_ID database key has been defined. RTE's that choose not to provide this support do not have to do anything - the associated code will simply be ignored.
6. include some cleanup of the hwloc open/close code so it conforms to how we did things in other frameworks (e.g., having a single "frame" file instead of open/close). Also, fix the locality flags - e.g., being on the same node means you must also be on the same cluster/cu, so ensure those flags are also set.
cmr:v1.7.4:reviewer=hjelmn
This commit was SVN r29435.
Fix two problems that surfaced when using direct launch under SLURM:
1. locally store our own data because some BTLs want to retrieve
it during add_procs rather than use what they have internally
2. cleanup MPI_Abort so it correctly passes the error status all
the way down to the actual exit. When someone implemented the
"abort_peers" API, they left out the error status. So we lost
it at that point and *always* exited with a status of 1. This
forces a change to the API to include the status.
cmr:v1.7.3:reviewer=jsquyres:subject=Fix MPI_Abort and modex_recv for direct launch
This commit was SVN r29405.
So we now allow singletons to start on their own, only spawning an HNP when initiating an operation that actually requires it.
cmr:v1.7.4:reviewer=jsquyres
This commit was SVN r29354.
Create a new required key in the OMPI layer for retrieving a "node id" from the database. ALL RTE'S MUST DEFINE THIS KEY. This allows us to compute locality in the MPI layer, which is necessary when we do things like intercomm_create.
cmr:v1.7.4:reviewer=rhc:subject=Cleanup handling of modex data
This commit was SVN r29274.
* paccept - establish a persistent listening port for async connect requests
* pconnect - async connect to remote process that has posted a paccept port. Provides a timeout mechanism, and allows the underlying implementation to retry until timeout
* pclose - shuts down a prior paccept posting
Includes example programs paccept.c and pconnect.c in orte/test/mpi. New MPI extension interfaces coming...
This commit was SVN r29063.
*** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE ***
Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro.
***************************************************************************************
I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week.
The code is in https://bitbucket.org/rhc/ompi-oob2
WHAT: Rewrite of ORTE OOB
WHY: Support asynchronous progress and a host of other features
WHEN: Wed, August 21
SYNOPSIS:
The current OOB has served us well, but a number of limitations have been identified over the years. Specifically:
* it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code)
* we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface.
* the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients
* there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort
* only one transport (i.e., component) can be "active"
The revised OOB resolves these problems:
* async progress is used for all application processes, with the progress thread blocking in the event library
* each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on")
* multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC.
* a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions.
* opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object
* NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions
* obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel
* the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport
* routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active
* all blocking send/recv APIs have been removed. Everything operates asynchronously.
KNOWN LIMITATIONS:
* although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline
* the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker
* routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways
* obviously, not every error path has been tested nor necessarily covered
* determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost.
* reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways
* the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC
This commit was SVN r29058.
This creates a really bad scaling behavior. Users have found a nearly 20% launch time differential between mpirun and PMI, with PMI being the slower method. Some of the problem is attributable to poor exchange algorithms in RM's like Slurm and Alps, but we make things worse by calling "get" so many times.
Nathan (with a tad advice from me) has attempted to alleviate this problem by reducing the number of "get" calls. This required the following changes:
* upon first request for data, have the OPAL db pmi component fetch and decode *all* the info from a given remote proc. It turned out we weren't caching the info, so we would continually request it and only decode the piece we needed for the immediate request. We now decode all the info and push it into the db hash component for local storage - and then all subsequent retrievals are fulfilled locally
* reduced the amount of data by eliminating the exchange of the OMPI_ARCH value if heterogeneity is not enabled. This was used solely as a check so we would error out if the system wasn't actually homogeneous, which was fine when we thought there was no cost in doing the check. Unfortunately, at large scale and with direct launch, there is a non-zero cost of making this test. We are open to finding a compromise (perhaps turning the test off if requested?), if people feel strongly about performing the test
* reduced the amount of RTE data being automatically fetched, and fetched the rest only upon request. In particular, we no longer immediately fetch the hostname (which is only used for error reporting), but instead get it when needed. Likewise for the RML uri as that info is only required for some (not all) environments. In addition, we no longer fetch the locality unless required, relying instead on the PMI clique info to tell us who is on our local node (if additional info is required, the fetch is performed when a modex_recv is issued).
Again, all this only impacts direct launch - all the info is provided when launched via mpirun as there is no added cost to getting it
Barring objections, we may move this (plus any required other pieces) to the 1.7 branch once it soaks for an appropriate time.
This commit was SVN r29040.
This commit reintroduces key compression into the pmi db. This feature
compresses the keys stored into the component into a small number of
PMI keys by serializing the data and base64 encoding the result. This
will avoid issues with Cray PMI which restricts us to ~ 3 PMI keys per
rank.
This commit was SVN r28993.
Features:
- Support for an override parameter file (openmpi-mca-param-override.conf).
Variable values in this file can not be overridden by any file or environment
value.
- Support for boolean, unsigned, and unsigned long long variables.
- Support for true/false values.
- Support for enumerations on integer variables.
- Support for MPIT scope, verbosity, and binding.
- Support for command line source.
- Support for setting variable source via the environment using
OMPI_MCA_SOURCE_<var name>=source (either command or file:filename)
- Cleaner API.
- Support for variable groups (equivalent to MPIT categories).
Notes:
- Variables must be created with a backing store (char **, int *, or bool *)
that must live at least as long as the variable.
- Creating a variable with the MCA_BASE_VAR_FLAG_SETTABLE enables the use of
mca_base_var_set_value() to change the value.
- String values are duplicated when the variable is registered. It is up to
the caller to free the original value if necessary. The new value will be
freed by the mca_base_var system and must not be freed by the user.
- Variables with constant scope may not be settable.
- Variable groups (and all associated variables) are deregistered when the
component is closed or the component repository item is freed. This
prevents a segmentation fault from accessing a variable after its component
is unloaded.
- After some discussion we decided we should remove the automatic registration
of component priority variables. Few component actually made use of this
feature.
- The enumerator interface was updated to be general enough to handle
future uses of the interface.
- The code to generate ompi_info output has been moved into the MCA variable
system. See mca_base_var_dump().
opal: update core and components to mca_base_var system
orte: update core and components to mca_base_var system
ompi: update core and components to mca_base_var system
This commit also modifies the rmaps framework. The following variables were
moved from ppr and lama: rmaps_base_pernode, rmaps_base_n_pernode,
rmaps_base_n_persocket. Both lama and ppr create synonyms for these variables.
This commit was SVN r28236.
A few changes were required to support this move:
1. the PMI component used to identify rte-related data (e.g., host name, bind level) and package them as a unit to reduce the number of PMI keys. This code was moved up to the ORTE layer as the OPAL layer has no understanding of these concepts. In addition, the component locally stored data based on process jobid/vpid - this could no longer be supported (see below for the solution).
2. the hash component was updated to use the new opal_identifier_t instead of orte_process_name_t as its index for storing data in the hash tables. Previously, we did a hash on the vpid and stored the data in a 32-bit hash table. In the revised system, we don't see a separate "vpid" field - we only have a 64-bit opaque value. The orte_process_name_t hash turned out to do nothing useful, so we now store the data in a 64-bit hash table. Preliminary tests didn't show any identifiable change in behavior or performance, but we'll have to see if a move back to the 32-bit table is required at some later time.
3. the db framework was a "select one" system. However, since the PMI component could no longer use its internal storage system, the framework has now been changed to a "select many" mode of operation. This allows the hash component to handle all internal storage, while the PMI component only handles pushing/pulling things from the PMI system. This was something we had planned for some time - when fetching data, we first check internal storage to see if we already have it, and then automatically go to the global system to look for it if we don't. Accordingly, the framework was provided with a custom query function used during "select" that lets you seperately specify the "store" and "fetch" ordering.
4. the ORTE grpcomm and ess/pmi components, and the nidmap code, were updated to work with the new db framework and to specify internal/global storage options.
No changes were made to the MPI layer, except for modifying the ORTE component of the OMPI/rte framework to support the new db framework.
This commit was SVN r28112.
Reasoning: The old behavior was a little confusing. mca_base_components_open does not open an output stream so it is a little unexpected that mca_base_components_close does. To add to this several frameworks (that don't use mca_base_components_close) failed to close their output in the framework close function and others closed their output a second time. This change is an improvement to the symantics of mca_base_components_open/close as they are now symetric in their functionality.
This commit was SVN r27570.
As a secondary cleanup, the HNP doesn't need to update its nidmap during an xcast as it already has an up-to-date picture of the situation. So just dump that data and move along.
This commit was SVN r27318.
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.
