Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] [OMPI svn] svn:open-mpi r25476
From: Ralph Castain (rhc_at_[hidden])
Date: 2011-11-18 10:49:54


On Nov 18, 2011, at 3:28 AM, George Bosilca wrote:

> I fixed some of the issues (nearly infinite loops on the orted due to non initialized values), but not the one in rmaps. Clearly, the code in ramps_base_ranking.c (the while loop starting with "while (cnt < jdata->num_procs))" reach an infinite loop as soon as no node->procs exists, as there is no way to increase the cnt (this is the case on the original launch).

That's a condition which should never be reached, but just to be safe, I have added a "bozo check" that will cause the routine to error out with a message if that situation occurs. I have tried everything - hostfile, dash-host, bizarre combinations of hosts, etc. - and been unable to replicate your described problem.

[rhc_at_odin ~/svn-trunk]$ cat hostfile
odin090 slots=8
odin091 slots=8
[rhc_at_odin ~/svn-trunk]$ mpirun -n 2 -mca ras ^slurm -hostfile hostfile -bynode -mca ras_base_verbose 10 hostname
[odin.cs.indiana.edu:10887] mca: base: components_open: Looking for ras components
[odin.cs.indiana.edu:10887] mca: base: components_open: opening ras components
[odin.cs.indiana.edu:10887] mca: base: components_open: found loaded component simulator
[odin.cs.indiana.edu:10887] mca: base: components_open: component simulator has no register function
[odin.cs.indiana.edu:10887] mca: base: components_open: component simulator open function successful
[odin.cs.indiana.edu:10887] mca: base: components_open: found loaded component loadleveler
[odin.cs.indiana.edu:10887] mca: base: components_open: component loadleveler has no register function
[odin.cs.indiana.edu:10887] mca: base: components_open: component loadleveler open function successful
[odin.cs.indiana.edu:10887] mca:base:select: Auto-selecting ras components
[odin.cs.indiana.edu:10887] mca:base:select:( ras) Querying component [simulator]
[odin.cs.indiana.edu:10887] mca:base:select:( ras) Skipping component [simulator]. Query failed to return a module
[odin.cs.indiana.edu:10887] mca:base:select:( ras) Querying component [loadleveler]
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:loadleveler: NOT available for selection
[odin.cs.indiana.edu:10887] mca:base:select:( ras) Skipping component [loadleveler]. Query failed to return a module
[odin.cs.indiana.edu:10887] mca:base:select:( ras) No component selected!
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:allocate
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:allocate nothing found in module - proceeding to hostfile
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:allocate checking hostfile hostfile
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:node_insert inserting 2 nodes
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:node_insert node odin090
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:node_insert node odin091
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:allocate
[odin.cs.indiana.edu:10887] [[24876,0],0] ras:base:allocate allocation already read
odin091.cs.indiana.edu
odin090.cs.indiana.edu

It makes me wonder if you are picking up a stale library somewhere?

> Moreover, reading the comments in this file, FREAK me out, as apparently each mapper is allowed to have a different behavior…

Not sure why that would "freak" you out as this has always been the case. Since the project started, we have allowed the user to separately specify mapping, ranking, and binding algorithms so they can reach a wide range of placement patterns. We have also allowed individual mappers to either use the base functions, or to completely ignore them - e.g., the rank-file mapper always ignored the base and just did things its own way.

None of that is new or changed. The only thing that changed is that we extended the range of resource types over which the user can exercise control. Instead of just slots and nodes, it now includes cores, hwthreads, and other strange beasts. So the number of possible combinations is much, much greater than before.

Don't blame me for the added complexity. I argued against some of this as I'm convinced it will rarely, if ever, be used - but I'm told this is what the user community wants, and so that is what I created.

What I did, as I stated, was to ensure that the previous minimal option behavior remains the same - i.e., default behavior and simple options like -bynode result in the same patterns as before. So only gluttons for punishment get exposed to the added complexity.

>
> george.
>
>
>
> On Nov 17, 2011, at 19:57 , Ralph Castain wrote:
>
>> I can't get it to fail, even with hostfile arguments. I'll try again in the morning.
>>
>> On Nov 17, 2011, at 8:49 PM, George Bosilca wrote:
>>
>>> Maybe the issue is generated by how the hostile is specified. I used orte_default_hostfile= in my mca-params.conf.
>>>
>>> george.
>>>
>>> On Nov 17, 2011, at 19:17 , Ralph Castain wrote:
>>>
>>>> I'm still building on odin, but will check there again to see if I can replicate - perhaps something didn't get committed cleanly.
>>>>
>>>>>
>>>>> george.
>>>>>
>>>>> On Nov 17, 2011, at 19:06 , Ralph Castain wrote:
>>>>>
>>>>>> Hmmm...well, things seem to work just fine for me:
>>>>>>
>>>>>> [rhc_at_odin ~/ompi-hwloc]$ mpirun -np 2 -bynode -mca plm rsh hostname
>>>>>> odin090.cs.indiana.edu
>>>>>> odin091.cs.indiana.edu
>>>>>>
>>>>>> [rhc_at_odin mpi]$ mpirun -np 2 -bynode -mca plm rsh ./hello_nodename
>>>>>> Hello, World, I am 1 of 2 on host odin091.cs.indiana.edu from app number 0 universe size 8
>>>>>> Hello, World, I am 0 of 2 on host odin090.cs.indiana.edu from app number 0 universe size 8
>>>>>>
>>>>>>
>>>>>> I'll get a fresh checkout and see if I can replicate from that...
>>>>>>
>>>>>> On Nov 17, 2011, at 7:42 PM, George Bosilca wrote:
>>>>>>
>>>>>>> I guess I reach one of these corner-cases that didn't got tested. I can't start any apps (not even a hostname) after this commit using the rsh PLM (as soon as I add a hostile). The mpirun is blocked in an infinite loop (after it spawned the daemons) in orte_rmaps_base_compute_vpids. Attaching with gdb indicates that cnt is never incremented, thus the mpirun is stuck forever in the while loop at line 397.
>>>>>>>
>>>>>>> I used "mpirun -np 2 --bynode ./tp_lb_ub_ng" to start my application, and I have a machine file containing two nodes:
>>>>>>>
>>>>>>> node01 slots=8
>>>>>>> node02 slots=8
>>>>>>>
>>>>>>> In addition CTRL+C seems to be broken …
>>>>>>>
>>>>>>> george.
>>>>>>>
>>>>>>> Begin forwarded message:
>>>>>>>
>>>>>>>> Author: rhc
>>>>>>>> Date: 2011-11-14 22:40:11 EST (Mon, 14 Nov 2011)
>>>>>>>> New Revision: 25476
>>>>>>>> URL: https://svn.open-mpi.org/trac/ompi/changeset/25476
>>>>>>>>
>>>>>>>> Log:
>>>>>>>> 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.
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> devel mailing list
>>>>>>> devel_at_[hidden]
>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> devel mailing list
>>>>>> devel_at_[hidden]
>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> devel mailing list
>>>>> devel_at_[hidden]
>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>
>>>>
>>>
>>>
>>> _______________________________________________
>>> devel mailing list
>>> devel_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>
>>
>> _______________________________________________
>> devel mailing list
>> devel_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel