1
1

208 Коммитов

Автор SHA1 Сообщение Дата
Ralph Castain
2dd2694f25 Fix comm_spawn in oversubscribed conditions. IF oversubscription is allowed, let nodes flow into the mapper even if they are oversubscribed, constrained by the slots_max absolute ceiling. Cleanup error messages when comm_spawn fails so it correctly and succintly reports the ereror.
This commit was SVN r25659.
2011-12-15 18:04:48 +00:00
Ralph Castain
e683b2f9c7 Minor touchup - reset the pointer to the end of the list each time to ensure we get the nodes in correct daemon order
This commit was SVN r25651.
2011-12-14 22:16:52 +00:00
Ralph Castain
f531b09a8d Correctly handle -host and -hostfile options. Ensure the initial vm launch constrains itself to the union of specified hosts if those options are given. Get oversubscribe set correctly for that case.
This commit was SVN r25648.
2011-12-14 20:01:15 +00:00
Nathan Hjelm
be11acf727 bug fix. don't add node to allocated_nodes twice
This commit was SVN r25619.
2011-12-12 19:14:41 +00:00
Ralph Castain
15facc4ba6 Fix comm_spawn yet again...add another test
This commit was SVN r25579.
2011-12-06 20:15:40 +00:00
Ralph Castain
90b7f2a7bf The rest of the multi app_context fix. Remove the restriction on number of app_contexts that can have zero np specified as multiple mappers now support that use-case. Update the ranking algorithms to respect and track bookmarks. Ensure we properly set the oversubscribed flag on a per-node basis.
This commit was SVN r25578.
2011-12-06 17:28:29 +00:00
Ralph Castain
d9c7764e9b Remove some debug
This commit was SVN r25575.
2011-12-05 22:04:50 +00:00
Ralph Castain
df2f594aa8 Some cleanup associated with multiple app_contexts. Ensure nodes only get entered once into the map. Correctly handle bookmarks. Cleanup tracking of slots_inuse and correct detection of oversubscription.
Still need to resolve the ranking issue so it starts at the bookmark, but that will come next.

This commit was SVN r25574.
2011-12-05 22:01:08 +00:00
Ralph Castain
07655e2945 Handle the case where the allocator "fibs" to us about the node names. In some cases (ahem...you know who you are!), the allocator will tell us a node number (e.g., "16"). However, the daemon will return a node name (e.g., "nid0016") - leaving us not recognizing its location.
So provide a new parameter (can't have too many!) that handles this situation by stripping the prefix from the returned node name. Also do a little cleanup to ensure we cleanly exit from errors, without generating too many annoying messages.

This commit was SVN r25562.
2011-12-02 14:10:08 +00:00
Ralph Castain
c56acf60ca Although we never really thought about it, we made an unconscious assumption in the mapper system - we assumed that the daemons would be placed on nodes in the order that the nodes appear in the allocation. In other words, we assumed that the launch environment would map processes in node order.
Turns out, this isn't necessarily true. The Cray, for example, launches processes in a toroidal pattern, thus causing the daemons to wind up somewhere other than what we thought. Other environments (e.g., slurm) are also capable of such behavior, depending upon the default mapping algorithm they are told to use.

Resolve this problem by making the daemon-to-node assignment in the affected environments when the daemon calls back and tells us what node it is on. Order the nodes in the mapping list so they are in daemon-vpid order as opposed to the order in which they show in the allocation. For environments that don't exhibit this mapping behavior (e.g., rsh), this won't have any impact.

Also, clean up the vm launch procedure a little bit so it more closely aligns with the state machine implementation that is coming, and remove some lingering "slave" code.

This commit was SVN r25551.
2011-11-30 19:58:24 +00:00
Ralph Castain
866edf6a89 Now that George has found his problem, we no longer need the bozo check. Interesting how these platform-specific issues surface...
This commit was SVN r25493.
2011-11-18 17:43:14 +00:00
Ralph Castain
1e5e9bde77 Add protection against a bozo case where we could end up in an infinite loop while calculating ranks
This commit was SVN r25491.
2011-11-18 15:35:55 +00:00
George Bosilca
61f273b987 Do not tolerate uninitialized variables.
This commit was SVN r25489.
2011-11-18 10:19:24 +00:00
Ralph Castain
6310361532 At long last, the fabled revision to the affinity system has arrived. A more detailed explanation of how this all works will be presented here:
https://svn.open-mpi.org/trac/ompi/wiki/ProcessPlacement

The wiki page is incomplete at the moment, but I hope to complete it over the next few days. I will provide updates on the devel list. As the wiki page states, the default and most commonly used options remain unchanged (except as noted below). New, esoteric and complex options have been added, but unless you are a true masochist, you are unlikely to use many of them beyond perhaps an initial curiosity-motivated experimentation.

In a nutshell, this commit revamps the map/rank/bind procedure to take into account topology info on the compute nodes. I have, for the most part, preserved the default behaviors, with three notable exceptions:

1. I have at long last bowed my head in submission to the system admin's of managed clusters. For years, they have complained about our default of allowing users to oversubscribe nodes - i.e., to run more processes on a node than allocated slots. Accordingly, I have modified the default behavior: if you are running off of hostfile/dash-host allocated nodes, then the default is to allow oversubscription. If you are running off of RM-allocated nodes, then the default is to NOT allow oversubscription. Flags to override these behaviors are provided, so this only affects the default behavior.

2. both cpus/rank and stride have been removed. The latter was demanded by those who didn't understand the purpose behind it - and I agreed as the users who requested it are no longer using it. The former was removed temporarily pending implementation.

3. vm launch is now the sole method for starting OMPI. It was just too darned hard to maintain multiple launch procedures - maybe someday, provided someone can demonstrate a reason to do so.

As Jeff stated, it is impossible to fully test a change of this size. I have tested it on Linux and Mac, covering all the default and simple options, singletons, and comm_spawn. That said, I'm sure others will find problems, so I'll be watching MTT results until this stabilizes.

This commit was SVN r25476.
2011-11-15 03:40:11 +00:00
Ralph Castain
fcee46b063 Add an option for printing a diffable process map for testing mappers
This commit was SVN r25428.
2011-11-03 14:22:07 +00:00
Ralph Castain
648c85b41b Add a simple pattern mapper as an example of how to use the topology info to create desired mappings. Let the user specify a pattern based on resource types, and map that pattern across all available nodes as resources permit.
Don't automatically display the topology for each node when --display-devel-map is set as it can overwhelm the reader. Use a separate flag --display-topo to get it.

This commit was SVN r25396.
2011-10-29 15:12:45 +00:00
Ralph Castain
2958f3de34 Add some clarifying comments and a small efficiency improvement
This commit was SVN r25322.
2011-10-18 18:30:43 +00:00
Ralph Castain
ae8e556d14 Okay, once again let's fix the vpid calculator. Identified problem with prior commit (some rmaps components already place their procs in the jdata->procs array, and others don't), so account for those variations.
This commit was SVN r25315.
2011-10-18 15:50:11 +00:00
George Bosilca
f28890fbb7 Revert r25302 as it break the --bynode option.
This commit was SVN r25311.

The following SVN revision numbers were found above:
  r25302 --> open-mpi/ompi@d7a8553179
2011-10-18 02:48:17 +00:00
Ralph Castain
d7a8553179 Fix the mapping algo for computing vpids - it was borked for bynode operations when using nperxxx directives
This commit was SVN r25302.
2011-10-17 19:49:04 +00:00
Ralph Castain
92c7372e20 Per the RFC from Jeff, move hwloc from opal/mca/common to its own static framework ala libevent. Have ORTE daemons collect the topology info at startup and, if --enable-hwloc-xml is set, send that info back to the HNP for later use. The HNP only retains unique topology "templates" to reduce memory footprint. Have the daemon include the local topology info in the nidmap buffer sent to each app so the apps don't all hammer the local system to discover it for themselves.
Remove the sysinfo framework as hwloc replaces that functionality.

This commit was SVN r25124.
2011-09-11 19:02:24 +00:00
Wesley Bland
4e7ff0bd5e By popular demand the epoch code is now disabled by default.
To enable the epochs and the resilient orte code, use the configure flag:

--enable-resilient-orte

This will define both:

ORTE_ENABLE_EPOCH
ORTE_RESIL_ORTE

This commit was SVN r25093.
2011-08-26 22:16:14 +00:00
Wesley Bland
09274cd047 Make sure that the epoch is initialized everywhere so we don't get weird output
during valgrind. This shouldn't have caused any problems with any actual
execution. Just extra warnings in valgrind.

This commit was SVN r25015.
2011-08-08 15:11:55 +00:00
Wesley Bland
e1ba09ad51 Add a resilience to ORTE. Allows the runtime to continue after a process (or
ORTED) failure. Note that more work will be necessary to allow the MPI layer to
take advantage of this.

Per RFC:
http://www.open-mpi.org/community/lists/devel/2011/06/9299.php

This commit was SVN r24815.
2011-06-23 20:38:02 +00:00
Ralph Castain
7f2d2e3de7 Track the app_context rank - will equal overall rank for single app_context jobs
This commit was SVN r24778.
2011-06-16 20:31:30 +00:00
Ralph Castain
e039c7b7ea Avoid crashing when debugging rmaps and a non-string resource constraint is given
This commit was SVN r24770.
2011-06-10 16:27:30 +00:00
Ralph Castain
bd8d9a943a Add diagnostics
This commit was SVN r24748.
2011-06-05 19:17:56 +00:00
Ralph Castain
8f401a0563 Enable the ability to constrain applications to hosts on the basis of resources.
This commit was SVN r24736.
2011-05-28 22:18:19 +00:00
Ralph Castain
dc6f616599 Enable VM launch.
For some time, ORTE has had the ability to launch daemons on all nodes prior to launching an application. It has largely been used outside of the OMPI community, and so was never explicitly turned "on" inside OMPI releases. Nevertheless, the code has been there.

Allowing VM launches does not require ANY changes to existing PLM components. All that was required was to have orterun launch the daemons as a separate call to orte_plm.spawn -prior- to launching the applications. The rest of the VM support code resides in the rmaps framework:

(a) a check when asked to map a job to see if it is the daemon job, and

(b) a separate "setup_virtual_machine" mapper in the rmaps base that creates the required map so the PLM's will do the right thing.

In order to support those users who have no RM allocation but like to give the allocation in the form of a -host or -hostfile argument to their application, there is a little more code in orterun and the setup_virtual_machine mapper to capture information passed in that manner.

This has been tested with rsh and slurm environments, and, since there is nothing environment-specific in the implementation, should work in others as well - but needs to be proven.

This commit was SVN r24524.
2011-03-12 22:50:53 +00:00
Ralph Castain
1297acde13 George raised some valid concerns about the extensibility of the revised rmaps framework. Address those by:
1. removing the enum of mapper values

2. change the req_mapper and last_mapper fields to char* so they can hold the component name instead of a mapper flag

3. revise the selection logic in the mapper components to reflect the change. Components now look for their name in the req_mapper field, or to see if other criteria (e.g., npernode) are set that mandate their doing the mapping

Several MCA params resided in the rmaps base for historical reasons - they have been in the base since at least the original 1.2 release (and perhaps earlier). However, George correctly pointed out that they really should reside in their respective components. Accordingly, move them to the components, but register synonyms to the old names to avoid breaking backward compatibility.

These revisions retain the current functionality of allowing comm_spawn'd jobs to use different mappers than the original job, and for the errmgr to utilize the resilient mapper to recover processes regardless of how they were originally mapped.

Given the large number of possible combinations, I am sure that someone will find a corner-case combination of values and selection criteria that cause either no mapper to be selected, or one other than the intended to be used. No one can test all the ways people will use this system, so I expect debugging to continue for awhile.

The ability of comm_spawn'd jobs to exploit this functionality relies on changes to the orte_dpm component - this will be committed separately.

This commit was SVN r24520.
2011-03-12 05:30:09 +00:00
Ralph Castain
3b4421d8e3 Separately track requested and last-used mapper so we don't lose that info
This commit was SVN r24502.
2011-03-09 18:51:36 +00:00
George Bosilca
9bbe00bdc3 Set the return code from the processes upstream.
This commit was SVN r24483.
2011-03-03 00:02:21 +00:00
George Bosilca
c6a5f9706a Thomas's patch: Assume we won't fail unless notified by a child.
This commit was SVN r24482.
2011-03-02 23:50:01 +00:00
Josh Hursey
62bba1bf12 Name the enum so that it represents as an actual symbol in gdb, instead of just a number.
This commit was SVN r24472.
2011-03-01 21:00:03 +00:00
Ralph Castain
65ba6af44d Cleanup our handling of VMs to ensure daemons don't get mapped when operating with a VM.
Have each mapper flag it did the map so we can see who did it later.

Ensure procs are flagged as "ready to launch".

This commit was SVN r24406.
2011-02-16 23:01:57 +00:00
Ralph Castain
9b38525d1e Remove unused include files
This commit was SVN r24394.
2011-02-16 00:32:47 +00:00
Ralph Castain
5120e6aec3 Redefine the rmaps framework to allow multiple mapper modules to be active at the same time. This allows users to map the primary job one way, and map any comm_spawn'd job in a different way. Modules are given the opportunity to map a job in priority order, with the round-robin mapper having the highest default priority. Priority of each module can be defined using mca param.
When called, each mapper checks to see if it can map the job. If npernode is provided, for example, then the loadbalance mapper accepts the assignment and performs the operation - all mappers before it will "pass" as they can't map npernode requests.

Also remove the stale and never completed topo mapper.

This commit was SVN r24393.
2011-02-15 23:24:31 +00:00
Ralph Castain
30c37ea536 Ensure that the oversubscribed condition of nodes is accurately reported by the mapper, and that the results are communicated and used by the backend orteds when setting sched_yield on local procs. Restores prior behavior that was somehow lost along the way.
Includes a patch from Damien Guinier to fix vpid assignments when cpus-per-task is specified.

This commit was SVN r24126.
2010-12-01 12:51:39 +00:00
Ralph Castain
ad5eaee4c6 Protect against NULL and provide additional resource check/error report
This commit was SVN r23432.
2010-07-19 18:33:32 +00:00
Ralph Castain
510ade9503 Do not use nodes that are flagged as down or do-not-use for this map. Modify error output to reflect possible reasons no nodes would be available
This commit was SVN r23333.
2010-07-01 19:39:31 +00:00
Abhishek Kulkarni
afbe3e99c6 * Wrap all the direct error-code checks of the form (OMPI_ERR_* == ret) with
(OMPI_ERR_* = OPAL_SOS_GET_ERR_CODE(ret)), since the return value could be a
 SOS-encoded error. The OPAL_SOS_GET_ERR_CODE() takes in a SOS error and returns
 back the native error code.

* Since OPAL_SUCCESS is preserved by SOS, also change all calls of the form
  (OPAL_ERROR == ret) to (OPAL_SUCCESS != ret). We thus avoid having to
  decode 'ret' to get the native error code.

This commit was SVN r23162.
2010-05-17 23:08:56 +00:00
Ralph Castain
871f445848 Ignore nodes that are "down" when generating maps
This commit was SVN r23119.
2010-05-12 18:08:40 +00:00
Ralph Castain
8da781af84 Continue developing support for distributed virtual machines - minor changes to ensure correct jobid gets used and that dvm's can communicate with tools
This commit was SVN r22958.
2010-04-12 22:33:09 +00:00
Ralph Castain
d3ed4e68b7 Utilize a non-used mapping policy bit to define a policy that uses only existing alive daemons to support virtual machines and restarting processes on already-active nodes
This commit was SVN r22951.
2010-04-10 05:02:47 +00:00
Ralph Castain
6b43b76f9d Some updates required for generating a LAM-style virtual machine. Retain the local node if requested. Properly setup the daemon job map for a VM launch.
This commit was SVN r22928.
2010-04-03 00:03:01 +00:00
Josh Hursey
e4f2d03d28 ErrMgr Framework redesign to better support fault tolerance development activities.
Explained in more detail in the following RFC:
  http://www.open-mpi.org/community/lists/devel/2010/03/7589.php

This commit was SVN r22872.
2010-03-23 21:28:02 +00:00
Ralph Castain
7ebf72b4aa Trivial cleanup
This commit was SVN r22813.
2010-03-10 18:24:38 +00:00
Ralph Castain
4355134991 Let the vm launcher specify the mapping policy
This commit was SVN r22797.
2010-03-08 19:13:21 +00:00
Ralph Castain
69fe5ca69b Correctly compute bynode mapping, even in the presence of a $#$%#@^$ rankfile
This commit was SVN r22748.
2010-03-02 05:21:42 +00:00
Ralph Castain
5514d9c673 Fix the stupid rankfile mapper again, hopefully not breaking everything else to accommodate it. Looks like the round-robin mappers still work, at least...
This commit was SVN r22746.
2010-03-01 20:40:47 +00:00