The problem was caused by a bad ordering between the restart of the ORTE level tcp connections (in the OOB - out-of-band communication) and the Open MPI level tcp connections (BTLs). Before this commit ORTE would shutdown and restart the OOB completely before the OMPI level restarted its tcp connections. What would happen is that a socket descriptor used by the OMPI level on checkpoint was assigned to the ORTE level on restart. But the OMPI level had no knowledge that the socket descriptor it was previously using has been recycled so it closed it on restart. This caused the ORTE level to break as the newly created socket descriptor was closed without its knowledge.
The fix is to have the OMPI level shutdown tcp connections, allow the ORTE level to restart, and then allow the OMPi level to restart its connections. This seems obvious, and I'm surprised that this bug has not cropped up sooner. I'm confident that this specific problem has been fixed with this commit.
Thanks to Eric Roman and Tamer El Sayed for their help in identifying this problem, and patience while I was fixing it.
* Add a new state {{{OPAL_CRS_RESTART_PRE}}}. This state identifies when we are on the down slope of the INC (finalize-like) which is useful when you want to close, but not reopen a component set for fear of interfering with a lower level.
* Use this new state in OMPI level coordination. Here we want to make sure to play well with both the OMPI/BTL/TCP and ORTE/OOB/TCP components.
* Update ft_event functions in PML and BML to handle the new restart state.
* Add an additional flag to the error output in OOB/TCP so we can see what the socket descriptor was on failure as this can be helpful in debugging.
This commit was SVN r18276.
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.
* 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.
the multiple threads accessing the OOB/registry asynchronously via the
callbacks. The quickest solution (but definitively not the cleanest) is
to serialize these callbacks in such a way that at any given time
only one thread can execute a callbacks.
This commit was SVN r15086.
through the win dll using multiple threads, we have to insure that
the oob callbacks happens only in a synchronous way or really bad
things happens with the current design (blocking messages from a receive
callback).
This commit was SVN r15069.
- make opal_sockaddr2str() take a sockaddr_storage instead of a sockaddr_in6
so that it works for IPv4 and IPv6 addresses, and remove a whole bunch
of #ifs in the OOOB code.
- Fix a compiler warning in the TCP BTL due to run-time determined
array size by making it a dynamicly allocated array.
- Fix the unpacking code of IPv4 addresses when using IPv6 support, so
that the address is in the correct location (instead of in an IPv6
structure, use an IPv4 structure). Refs trac:1005.
This commit was SVN r14514.
The following Trac tickets were found above:
Ticket 1005 --> https://svn.open-mpi.org/trac/ompi/ticket/1005
* Remove the connect() timeout code, as it had some nasty race conditions
when connections were established as the trigger was firing. A better
solution has been found for the cluster where this was needed, so just
removing it was easiest.
* When a fatal error (too many connection failures) occurs, set an error
on messages in the queue even if there isn't an active message. The
first message to any peer will be queued without being active (and
so will all subsequent messages until the connection is established),
and the orteds will hang until that first message completes. So if
an orted can never contact it's peer, it will never exit and just sit
waiting for that message to complete.
* Cover an interesting RST condition in the connect code. A connection
can complete the three-way handshake, the connector can even send
some data, but the server side will drop the connection because it
can't move it from the half-connected to fully-connected state because
of space shortage in the listen backlog queue. This causes a RST to
be received first time that recv() is called, which will be when waiting
for the remote side of the OOB ack. In this case, transition the
connection back into a CLOSED state and try to connect again.
* Add levels of debugging, rather than all or nothing, each building on
the previous level. 0 (default) is hard errors. 1 is connection
error debugging info. 2 is all connection info. 3 is more state
info. 4 includes all message info.
* Add some hopefully useful comments
This commit was SVN r14261.
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
Accordingly, there are new APIs to the name service to support the ability to get a job's parent, root, immediate children, and all its descendants. In addition, the terminate_job, terminate_orted, and signal_job APIs for the PLS have been modified to accept attributes that define the extent of their actions. For example, doing a "terminate_job" with an attribute of ORTE_NS_INCLUDE_DESCENDANTS will terminate the given jobid AND all jobs that descended from it.
I have tested this capability on a MacBook under rsh, Odin under SLURM, and LANL's Flash (bproc). It worked successfully on non-MPI jobs (both simple and including a spawn), and MPI jobs (again, both simple and with a spawn).
This commit was SVN r12597.
copy of the receive buffer based on the iovec struct that may have been updated
during partial reads to reflect the current offset. Need to make the copy using
the base address of the buffer.
Thanks to Sven Stork for finding this.
This should be backported to 1.0.X and 1.1.X branches.
This commit was SVN r9749.
thread, which will do progress independently of MPI. So in this case we
have to call opal_event_loop instead of opal_progress.
This commit was SVN r9551.
event library (since the event library has its own thread). So when
we are using progress threads, we really want to call opal_event_loop()
and not opal_progres().
This commit was SVN r9549.
- move files out of toplevel include/ and etc/, moving it into the
sub-projects
- rather than including config headers with <project>/include,
have them as <project>
- require all headers to be included with a project prefix, with
the exception of the config headers ({opal,orte,ompi}_config.h
mpi.h, and mpif.h)
This commit was SVN r8985.
