The bug was a race condition in the barrier operation that caused the barrier in MPI_Finalize to fail on very short programs.
Scalaiblity was improved by using the daemons to aggregate modex and barrier messages before sending them to the rank=0 proc. Improvement is proportional to ppn, of course, but there really wasn't a scaling problem at low ppn anyway. This modification also paves the way for better allgather operations since now all the data for each node is sitting at the daemon level, and the daemons are now aware that a collective operation on the OOB is underway (so they -can- participate in a collective of their own to support it).
Also added better diagnostics to map out the timing associated with MPI_Init - turned on by -mca orte_timing 1.
This commit was SVN r17988.
1. taking advantage of the fact that we no longer create the launch message via a GPR trigger. In earlier times, we had the GPR create the launch message based on a subscription. In that mode of operation, we could not guarantee the order in which the data was stored in the message - hence, we had no choice but to parse the message in a loop that checked each value against a list of possible "keys" until the corresponding value was found.
Now, however, we construct the message "by hand", so we know precisely what data is in each location in the message. Thus, we no longer need to send the character string "keys" for each data value any more. This represents a rather large savings in the message size - to give you an example, we typically would use a 30-char "key" for a 2-byte data value. As you can see, the overhead can become very large.
2. sending node-specific data only once. Again, because we used to construct the message via subscriptions that were done on a per-proc basis, the data for each node (e.g., the daemon's name, whether or not the node was oversubscribed) would be included in the data for each proc. Thus, the node-specific data was repeated for every proc.
Now that we construct the message "by hand", there is no reason to do this any more. Instead, we can insert the data for a specific node only once, and then provide the per-proc data for that node. We therefore not only save all that extra data in the message, but we also only need to parse the per-node data once.
The savings become significant at scale. Here is a comparison between the revised trunk and the trunk prior to this commit (all data was taken on odin, using openib, 64 nodes, unity message routing, tested with application consisting of mpi_init/mpi_barrier/mpi_finalize, all execution times given in seconds, all launch message sizes in bytes):
Per-node scaling, taken at 1ppn:
#nodes original trunk revised trunk
time size time size
1 0.10 819 0.09 564
2 0.14 1070 0.14 677
3 0.15 1321 0.14 790
4 0.15 1572 0.15 903
8 0.17 2576 0.20 1355
16 0.25 4584 0.21 2259
32 0.28 8600 0.27 4067
64 0.50 16632 0.39 7683
Per-proc scaling, taken at 64 nodes
ppn original trunk revised trunk
time size time size
1 0.50 16669 0.40 7720
2 0.55 32733 0.54 11048
3 0.87 48797 0.81 14376
4 1.0 64861 0.85 17704
Condensing those numbers, it appears we gained:
per-node message size: 251 bytes/node -> 113 bytes/node
per-proc message size: 251 bytes/proc -> 52 bytes/proc
per-job message size: 568 bytes/job -> 399 bytes/job
(job-specific data such as jobid, override oversubscribe flag, total #procs in job, total slots allocated)
The fact that the two pre-commit trunk numbers are the same confirms the fact that each proc was containing the node data as well. It isn't quite the 10x message reduction I had hoped to get, but it is significant and gives much better scaling.
Note that the timing info was, as usual, pretty chaotic - the numbers cited here were typical across several runs taken after the initial one to avoid NFS file positioning influences.
Also note that this commit removes the orte_process_info.vpid_start field and the handful of places that passed that useless value. By definition, all jobs start at vpid=0, so all we were doing is passing "0" around. In fact, many places simply hardwired it to "0" anyway rather than deal with it.
This commit was SVN r16428.
This patch also fixes a minor bug discovered along the way: we had "lost" the passing of the oversubscribed condition flag from the mapper to the orteds. Thus, we were not setting sched_yield correctly when in oversubscribed conditions (except when a hostfile was specified - different logic there because we treat the number of slots allocated on the node as "uncertain")
I did not modify the process component in this patch - I will send a proposed patch to the maintainers of that component so they can review it first.
This commit was SVN r16418.
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.
and implementation. This has shown drastic performance benefit when
transferring Many files at roughly the same time.
I tested this for many different filem operations and everything was working
fine. Let me know if you have any problems with this functionality.
Some Notes:
- opal-checkpoint now has a 'quiet' flag to keep it from being too verbose.
- FileM RSH component is fully non-blocking.
- FileM RSH component has incomming connection throttling since by default
ssh only allows 10 concurrent scp connections to any single host. This
default can be adjusted via an MCA parameter.
{{{-mca filem_rsh_max_incomming 10}}}
- There is an MCA parameter for max outgoing connections, but it is currently
not implemented. If someone needs it then it should not be hard to implement.
{{{-mca filem_rsh_max_outgoing 10}}}
- Changed the FileM request structure so that it is a bit more explicit and
flexible.
- Moved the 'preload-binary' and 'preload-files' functionality into odls/base
allowing for code reuse in the 'process' and 'default' ODLS components.
- Fixed a bug in the process name resolution which broke the 'preload-*'
functionality due to GPR table structure changes.
- The FileM RSH component might be able to see even more speedup from using a
thread pool to operate on the work_pool structures, but that is for future
work.
- Added a 'opal-show-help' file to ODLS Base
This commit was SVN r16252.
* General TCP cleanup for OPAL / ORTE
* Simplifying the OOB by moving much of the logic into the RML
* Allowing the OOB RML component to do routing of messages
* Adding a component framework for handling routing tables
* Moving the xcast functionality from the OOB base to its own framework
Includes merge from tmp/bwb-oob-rml-merge revisions:
r15506, r15507, r15508, r15510, r15511, r15512, r15513
This commit was SVN r15528.
The following SVN revisions from the original message are invalid or
inconsistent and therefore were not cross-referenced:
r15506
r15507
r15508
r15510
r15511
r15512
r15513
Cleanup ALL instances of output involving the printing of orte_process_name_t structures using the ORTE_NAME_ARGS macro so that the number of fields and type of data match. Replace those values with a new macro/function pair ORTE_NAME_PRINT that outputs a string (using the new thread safe data capability) so that any future changes to the printing of those structures can be accomplished with a change to a single point.
Note that I could not possibly find outputs that directly print the orte_process_name_t fields, but only dealt with those that used ORTE_NAME_ARGS. Hence, you may still have a few outputs that bark during compilation. Also, I could only verify those that fall within environments I can compile on, so other environments may yield some minor warnings.
This commit was SVN r15517.
You will not see any impact from this change unless you use the syntax described in ticket #1023. I've tried as many of the RAS components as possible and saw no problem - there may be issues with other RAS components that would not compile on any of my systems. Anything that appears should be trivial to fix.
This commit was SVN r15427.
The problem stemmed from no longer launching a local orted on the same node as mpirun. The orted would save and reuse the base environment. Mpirun didn't do that, and the odls was using the orted's globally saved environment (which wasn't being set).
This fix establishes a globally accessible base launch environment that both the orted and mpirun can utilize. Since we now use that, we don't need to pass it to the odls_launch_proc function, so remove that param from the API (and modify all components to handle the change).
This commit was SVN r15405.
Short description: major changes include -
1. singletons now fork/exec a local daemon to manage their operations.
2. the orte daemon code now resides in libopen-rte
3. daemons no longer use the orte triggering system during startup. Instead, they directly call back to their parent pls component to report ready to operate. A base function to count the callbacks has been provided.
I have modified all the pls components except xcpu and poe (don't understand either well enough to do it). Full functionality has been verified for rsh, SLURM, and TM systems. Compile has been verified for xgrid and gridengine.
This commit was SVN r15390.
The default odls has been updated and works fine. The process odls has been updated, but I could not verify its operation. The bproc ODLS has not been updated yet. Ralph will look at it soon.
This commit was SVN r15257.
The SnapC Full local Coordinator used this argument to attach to the job the
daemon would be launching. So once this option was removed C/R support broke.
This commit has the local coordinator attach to the job just before it is
launched by the ODLS module. This is a much cleaner solution, and will
eventually allow the SnapC modules to attach to multiple jobs launched
on a single machine.
This commit fixes the C/R regression introduced in r15007.
This commit was SVN r15121.
The following SVN revision numbers were found above:
r15007 --> open-mpi/ompi@85df3bd92f
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
A bunch of fixes from the /tmp/iof-fixes branch that fix up ''some''
(but not ''all'') of the problems that we have seen with iof:
* Reading very large files via stdin redirected to orteun (Sun saw
this)
* Reading a little bit of a large file redirected to orterun's stdin
and then either closing stdin or exiting the process
The Big Change was to make the proxy iof (the one running in non-HNP
orteds) send back a "I'm closing the stream" ACK back to the service
iof. This tells the HNP that there will be nothing more coming from
that peer, and therefore the iof forward should be removed.
Many other minor cleanups/fixes, terminology changes, and
documentation additions are included in this commit as well. However,
there are still some pretty big outstanding issues with IOF that are
not addressed either by #967 or this commit. A few examples:
* IOF was designed to allow multiple subscribers to a single stream.
We're not entirely sure that this works (for one thing, there is
nothing in the ORTE/OMPI code base that uses this functionality).
* There are also resources leaked when processes/jobs exit (per
Ralph's first comment on this ticket).
* There is no feedback to close orterun's stdin when all subscribers
to the corresponding stream have closed stdin.
This commit was SVN r14967.
The following Trac tickets were found above:
Ticket 967 --> https://svn.open-mpi.org/trac/ompi/ticket/967
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
To be precise, given this hypothetical launching pattern:
host1: vpids 0, 2, 4, 6
host2: vpids 1, 3, 5, 7
The local_rank for these procs would be:
host1: vpids 0->local_rank 0, v2->lr1, v4->lr2, v6->lr3
host2: vpids 1->local_rank 0, v3->lr1, v5->lr2, v7->lr3
and the number of local procs on each node would be four. If vpid=0 then does a comm_spawn of one process on host1, the values of the parent job would remain unchanged. The local_rank of the child process would be 0 and its num_local_procs would be 1 since it is in a separate jobid.
I have verified this functionality for the rsh case - need to verify that slurm and other cases also get the right values. Some consolidation of common code is probably going to occur in the SDS components to make this simpler and more maintainable in the future.
This commit was SVN r14706.
Test for system limits (where known) prior to doing things like fork and pipe since some systems aren't very nice about it when we try to exceed such limits.
This commit was SVN r14494.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
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
This change utilizes the new num_processors function. I also left the mods made to ompi_mpi_init and the bug fix for the default value of mpi_yield_when_idle. Note that the mods to mpi_init will not really take effect as the mca param will now *always* be set (either by user or odls). We will need those mods later, so no point in removing them now.
This commit was SVN r13519.
1. if the user has specified sched_yield, we simply do what we are told
2. if they didn't specify anything, try to get the number of processors on this node. Note that we already now get the number of local procs in our job that are sharing this node - that now comes in through the proc callback and is stored in the ompi_proc_t structures.
3. if we can get the number of processors, compare that to the number of local procs from my job that are sharing my node. If the number of local procs exceeds the number of processors, then set sched_yield to true. If not, then be a hog and set sched_yield to false
4. if we can't get the number of processors, default to conservative behavior and set sched_yield to true.
Note that I have not yet dealt with the need to dynamically adjust this setting as more processes are added via comm_spawn. So far, we are *only* looking within our own job. Given that we have now moved this logic to mpi_init (and away from the orteds), it isn't yet clear to me how a process will be informed about the number of procs in *other* jobs that are also sharing this node.
Something to continue to ponder.
This commit was SVN r13430.
rc (which is -1 or 4 if we hit this case) resulted in an odd error that a
signal killed the proc (instead of a startup error, as is reality).
Instead, use the W_EXITCODE macro (if available) to build up an exit
code that has an error code for exit status, but does not make it look
like the process died from a signal
This commit was SVN r12890.
I found only two places that were looking at the tokens:
1. the odls - we used the tokens to separately process the globals container data from everything else. In this case, I left the subscription that returned the globals data alone, but "stripped" the subscription that returned the launch data for the procs. These subscriptions have nothing to do with the xcast message.
2. the pml_base_modex - the callback function was getting process names from the returned tokens. Actually, this function was doing a very bad thing - it was assuming that the first token returned was *always* the process name. This is currently true, but is one of those assumptions that someone could have easily changed - and suddenly found the system inexplicably failing. I modified the function to (a) get the name sent back to us, (b) "stripped" the value structures of tokens and segment strings, and (c) correctly obtained process names from the returned values. I also reindented the heck out of the code so it was legible (at least, to my old eyes).
This commit was SVN r12813.
Obviously, people like bproc will have to get the app_num via another avenue...but that's a problem for another day. Several options are easily available.
This commit was SVN r12788.
Modify the RMAPS framework so we eliminate communicating a map to a backend node when certain attributes are set. The proxy functions are now implemented in the base, and a check made for HNP/non-HNP operation made in the map_jobs function prior to execution.
This commit was SVN r12619.
