Ok I may have not explained very clearly. In our case we only use
SLURM for the resource manager.
The difference here is that the SLURM version that we use has
support for ALPS. Therefore when we run our job using the mpirun
command, since we have the alps environment loaded, it's the ALPS
RAS which is selected, and the ALPS PLM as well. I think I could
even not compile the OpenMPI slurm support.
Here is what we do for example: here is my batch script (with the
patched version)
#!/bin/bash
#SBATCH --job-name=HelloOMPI
#SBATCH --nodes=2
#SBATCH --time=00:30:00
set -ex
cd /users/soumagne/gele/hello
mpirun --mca ras_base_verbose 10 --mca plm_base_verbose 10 -np 2
--bynode `pwd`/hello
And here is the output that I get:
soumagne@gele1:~/gele/hello> more slurm-165.out
+ cd /users/soumagne/gele/hello
++ pwd
+ mpirun --mca ras_base_verbose 10 --mca plm_base_verbose 10
--bynode -np 2 /use
rs/soumagne/gele/hello/hello
[gele2:15844] mca: base: components_open: Looking for plm components
[gele2:15844] mca: base: components_open: opening plm components
[gele2:15844] mca: base: components_open: found loaded component alps
[gele2:15844] mca: base: components_open: component alps has no
register functio
n
[gele2:15844] mca: base: components_open: component alps open
function successfu
l
[gele2:15844] mca: base: components_open: found loaded component slurm
[gele2:15844] mca: base: components_open: component slurm has no
register functi
on
[gele2:15844] mca: base: components_open: component slurm open
function successf
ul
[gele2:15844] mca:base:select: Auto-selecting plm components
[gele2:15844] mca:base:select:( plm) Querying component [alps]
[gele2:15844] mca:base:select:( plm) Query of component [alps] set
priority to
75
[gele2:15844] mca:base:select:( plm) Querying component [slurm]
[gele2:15844] mca:base:select:( plm) Query of component [slurm] set
priority to
75
[gele2:15844] mca:base:select:( plm) Selected component [alps]
[gele2:15844] mca: base: close: component slurm closed
[gele2:15844] mca: base: close: unloading component slurm
[gele2:15844] mca: base: components_open: Looking for ras components
[gele2:15844] mca: base: components_open: opening ras components
[gele2:15844] mca: base: components_open: found loaded component cm
[gele2:15844] mca: base: components_open: component cm has no
register function
[gele2:15844] mca: base: components_open: component cm open function
successful
[gele2:15844] mca: base: components_open: found loaded component alps
[gele2:15844] mca: base: components_open: component alps has no
register functio
n
[gele2:15844] mca: base: components_open: component alps open
function successfu
l
[gele2:15844] mca: base: components_open: found loaded component slurm
[gele2:15844] mca: base: components_open: component slurm has no
register functi
on
[gele2:15844] mca: base: components_open: component slurm open
function successf
ul
[gele2:15844] mca:base:select: Auto-selecting ras components
[gele2:15844] mca:base:select:( ras) Querying component [cm]
[gele2:15844] mca:base:select:( ras) Skipping component [cm]. Query
failed to r
eturn a module
[gele2:15844] mca:base:select:( ras) Querying component [alps]
[gele2:15844] ras:alps: available for selection
[gele2:15844] mca:base:select:( ras) Query of component [alps] set
priority to
75
[gele2:15844] mca:base:select:( ras) Querying component [slurm]
[gele2:15844] mca:base:select:( ras) Query of component [slurm] set
priority to
75
[gele2:15844] mca:base:select:( ras) Selected component [alps]
[gele2:15844] mca: base: close: unloading component cm
[gele2:15844] mca: base: close: unloading component slurm
[gele2:15844] ras:alps:allocate: Using ALPS configuration file:
"/etc/sysconfig/
alps"
[gele2:15844] ras:alps:allocate: Located ALPS scheduler file:
"/ufs/alps_shared/
appinfo"
[gele2:15844] ras:alps:orte_ras_alps_get_appinfo_attempts: 10
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: added NID 16 to list
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 16
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: added NID 20 to list
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:read_appinfo: got NID 20
[gele2:15844] ras:alps:allocate: success
I am nid00020 process 2/2
I am nid00016 process 1/2
[gele2:15844] mca: base: close: unloading component alps
[gele2:15844] mca: base: close: component alps closed
[gele2:15844] mca: base: close: unloading component alps
I think that in this case you would not break anything since it's
really a basic patch which enables you to directly do an mpirun,
without having to manually select any reservation id (assuming that
the user has the SLURM version with ALPS support which will be
available soon).
Jerome
On 07/09/2010 03:06 PM, Ralph Castain wrote:
Afraid I'm now even more confused. You use SLURM to do the
allocation, and then use ALPS to launch the job?
I'm just trying to understand because I'm the person who generally
maintains this code area. We have two frameworks involved here:
1. RAS - determines what nodes were allocated to us. There are both
slurm and alps modules here.
2. PLM - actually launches the job. There are both slurm and alps
modules here.
Up until now, we have always seen people running with either alps
or slurm, but never both together, so the module selection of these
two frameworks is identical - if you select slurm for the RAS
module, you will definitely get slurm for the launcher. Ditto for
alps. Are you sure that mpirun is actually using the modules you
think? Have you run this with -mca ras_base_verbose 10 -mca
plm_base_verbose 10 and seen what modules are being used?
In any event, this seems like a very strange combination, but I
assume you have some reason for doing this?
I'm always leery of fiddling with the SLURM modules as (a) there
aren't very many slurm users out there, (b) the primary users are
the DOE national labs themselves, using software provided by LLNL
(who controls slurm), and (c) there are major disconnects between
the various slurm releases, so we wind up breaking things for
someone rather easily.
So the more I can understand what you are doing, the easier it is
to determine how to use your patch without breaking slurm support
for others.
Thanks!
Ralph
On Jul 9, 2010, at 6:46 AM, Jerome Soumagne wrote:
Well we actually use a patched version of SLURM, 2.2.0-pre8. It is
planned to submit the modifications made internally at CSCS for
the next SLURM release in November. We implement ALPS support
based on the basic architecture of SLURM.
SLURM is only used to do the ALPS ressource allocation. We then
use mpirun based on the portals and on the alps libaries.
We don't use mca parameters to direct selection and the alps RAS
is automatically well selected.
On 07/09/2010 01:59 PM, Ralph Castain wrote:
Forgive my confusion, but could you please clarify something? You
are using ALPS as the resource manager doing the allocation, and
then using SLURM as the launcher (instead of ALPS)?
That's a combination we've never seen or heard about. I suspect
our module selection logic would be confused by such a
combination - are you using mca params to direct selection?
On Jul 9, 2010, at 4:19 AM, Jerome Soumagne wrote:
Hi,
We've recently installed OpenMPI on one of our Cray XT5
machines, here at CSCS. This machine uses SLURM for launching jobs.
Doing an salloc defines this environment variable:
BASIL_RESERVATION_ID
The reservation ID on Cray systems running
ALPS/BASIL only.
Since the alps ras module tries to find a variable called
OMPI_ALPS_RESID which is set using a script, we thought that for
SLURM systems it would be a good idea to directly integrate this
BASIL_RESERVATION_ID variable in the code, rather than using a
script. The small patch is attached.
Regards,
Jerome
--
Jérôme Soumagne
Scientific Computing Research Group
CSCS, Swiss National Supercomputing Centre
Galleria 2, Via Cantonale | Tel: +41 (0)91 610 8258
CH-6928 Manno, Switzerland | Fax: +41 (0)91 610 8282
<patch_slurm_alps.txt>_______________________________________________
devel mailing list
de...@open-mpi.org <mailto:de...@open-mpi.org>
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
--
Jérôme Soumagne
Scientific Computing Research Group
CSCS, Swiss National Supercomputing Centre
Galleria 2, Via Cantonale | Tel: +41 (0)91 610 8258
CH-6928 Manno, Switzerland | Fax: +41 (0)91 610 8282
_______________________________________________
devel mailing list
de...@open-mpi.org <mailto:de...@open-mpi.org>
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org <mailto:de...@open-mpi.org>
http://www.open-mpi.org/mailman/listinfo.cgi/devel
