Re: [OMPI users] Running application with MPI_Comm_spawn() in multithreaded environment

Wed, 1 Oct 2008 09:08:11 -0400

Okay, I believe I understand the problem. What this error is telling you is that the Torque MOM is refusing our connection request because it is already busy. So we cannot spawn another process.
If I understand your application correctly, you are spinning off multiple threads, each attempting to comm_spawn a single process - true? The problem with that design is that - since OMPI is not thread safe yet - these threads are all attempting to connect to the MOM at the same time. The MOM will only allow one connection at a time, and so at some point we are requesting a connection while already connected. 


Since we are some ways off from attaining thread safety in these  
scenarios, you really have three choices:



1. you could do this with a single comm_spawn call. Remember, you can  
provide an MPI_Info key to comm_spawn essentially telling it where to  
place the various process ranks. Unless you truly want each new  
process to be in its own comm_world, there is no real need to do this  
with 10000 individual calls to comm_spawn.



2. you could execute your own thread locking scheme in your  
application so that only one thread calls comm_spawn at a time.



3. remove the threaded launch scenario and just call comm_spawn in a  
loop.



In truth, the threaded approach to spawning all these procs isn't  
gaining you anything. Torque will only do one launch at a time anyway,  
so you will launch them serially no matter what. You may just be  
adding complexity for no real net gain.


Ralph

On Oct 1, 2008, at 1:56 AM, Roberto Fichera wrote:


Ralph Castain ha scritto:

Hi Roberto


There is something wrong with this cmd line - perhaps it wasn't  
copied

correctly?

mpirun --verbose --debug-daemons --mca obl -np 1 -wdir `pwd`
testmaster 10000 $PBS_NODEFILE

Specifically, the following is incomplete: --mca obl

I'm not sure if this is the problem or not, but I am unaware of such
an option and believe it could cause mpirun to become confused.

Ops! Sorry, I copied the wrong log, below there the right one:

[roberto@master TestOpenMPI]$ qsub -I testmaster.pbs
qsub: waiting for job 920.master.tekno-soft.it to start
qsub: job 920.master.tekno-soft.it ready

[roberto@master TestMPICH2]$ cd /data/roberto/MPI/TestOpenMPI/

[roberto@master TestOpenMPI]$ mpirun --debug-daemons --mca btl  
tcp,self

-wdir "`pwd`" -np 1 testmaster 100000 $PBS_NODEFILE
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
add_local_procs

[master.tekno-soft.it:05407] [[11340,0],0] node[0].name master  
daemon 0

arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[1].name cluster4  
daemon

INVALID arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[2].name cluster3  
daemon

INVALID arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[3].name cluster2  
daemon

INVALID arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[4].name cluster1  
daemon

INVALID arch ffc91200
Initializing MPI ...
[master.tekno-soft.it:05407] [[11340,0],0] orted_recv: received
sync+nidmap from local proc [[11340,1],0]
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
collective data cmd
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
message_local_procs
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
collective data cmd
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
message_local_procs
Loading the node's ring from file
'/var/torque/aux//920.master.tekno-soft.it'
... adding node #1 host is 'cluster4.tekno-soft.it'
... adding node #2 host is 'cluster3.tekno-soft.it'
... adding node #3 host is 'cluster2.tekno-soft.it'
... adding node #4 host is 'cluster1.tekno-soft.it'
A 4 node's ring has been made
At least one node is available, let's start to distribute 100000 job
across 4 nodes!!!
****************** Starting job #1
****************** Starting job #2
****************** Starting job #3
Setting up the host as 'cluster4.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task 'testslave' on node 'cluster4.tekno-soft.it'
Setting up the host as 'cluster3.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task 'testslave' on node 'cluster3.tekno-soft.it'
Setting up the host as 'cluster2.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task 'testslave' on node 'cluster2.tekno-soft.it'
Setting up the host as 'cluster1.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task 'testslave' on node 'cluster1.tekno-soft.it'
****************** Starting job #4

Daemon was launched on cluster3.tekno-soft.it - beginning to  
initialize
Daemon [[11340,0],1] checking in as pid 9487 on host cluster3.tekno- 
soft.it

Daemon [[11340,0],1] not using static ports
--------------------------------------------------------------------------

A daemon (pid unknown) died unexpectedly on signal 1  while  
attempting to

launch so we are aborting.

There may be more information reported by the environment (see above).


This may be because the daemon was unable to find all the needed  
shared
libraries on the remote node. You may set your LD_LIBRARY_PATH to  
have the

location of the shared libraries on the remote nodes and this will
automatically be forwarded to the remote nodes.
--------------------------------------------------------------------------

[master.tekno-soft.it:05407] [[11340,0],0] ORTE_ERROR_LOG: Resource  
busy

in file base/plm_base_receive.c at line 169

[master.tekno-soft.it:05414] [[11340,1],0] ORTE_ERROR_LOG: The  
specified

application failed to start in file dpm_orte.c at line 677
[master.tekno-soft.it:05414] *** An error occurred in MPI_Comm_spawn
[master.tekno-soft.it:05414] *** on communicator MPI_COMM_WORLD

[master.tekno-soft.it:05414] *** MPI_ERR_SPAWN: could not spawn  
processes

[master.tekno-soft.it:05414] *** MPI_ERRORS_ARE_FATAL (goodbye)


[master.tekno-soft.it:05407] [[11340,0],0] ORTE_ERROR_LOG: Resource  
busy

in file base/plm_base_receive.c at line 169
[master.tekno-soft.it:05407] [[11340,0],0] orted_cmd: received
add_local_procs

[master.tekno-soft.it:05407] [[11340,0],0] node[0].name master  
daemon 0

arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[1].name cluster4  
daemon

INVALID arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[2].name cluster3  
daemon

1 arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[3].name cluster2  
daemon

INVALID arch ffc91200

[master.tekno-soft.it:05407] [[11340,0],0] node[4].name cluster1  
daemon

INVALID arch ffc91200
[cluster3.tekno-soft.it:09487] [[11340,0],1] orted: up and running -
waiting for commands!





Ralph


On Sep 30, 2008, at 8:24 AM, Roberto Fichera wrote:


Roberto Fichera ha scritto:

Hi All on the list,


I'm trying to execute dynamic MPI applications using  
MPI_Comm_spawn().

The application I'm using for tests, basically is

composed by a master, which spawn a slave in each assigned node  
in a

multithreading fashion. The master is started with a
number of jobs to perform and a filename, containing the list of
assigned nodes. The idea is to handle all the dispatching

logic within the application, so that the master will try to take  
as

busy as possible each assigned node. Said that, for each spawned
job, the master allocate a thread for spawning and handling the
communication, than generate a random number, send it to the
slave which simply send it back to the master. Finally the slave
terminate its job and the relative node become free for a new one.
The things will continue until all the requested jobs are done.


The test program I'm using *doesn't* work flawless in mpich2  
because it

has a ~24k spawned job limitation, due to a monotonically
increasing of its internal context id which basically stops the

application due to a library internal overflow. The internal  
context

id,
allocated
for each terminated spawned job, are never recycled at moment. The
unique MPI-2 implementation, so supporting MPI_Comm_spawn(),

which was able to complete the test is currently the HP MPI. So  
now I
would start to check OpenMPI if it's suitable for our dynamic  
parallel

applications.


The test application is linked against OpenMPI v1.3a1r19645,  
running of

Fedora8 x86_64 + all updates.


My first attempt end up on the error below which I basically  
don't know

where to look further. Note that I've already checked PATHs and
LD_LIBRARY_PATH, the application is basically configured correctly
since
it uses two scripts for starting and all the paths are set there.

Basically I need to start *one* master application which will  
handle

all

the things for managing slave applications. The communication is  
*only*

master <-> slave and never collective, at moment.

The test program is available on request.

Does any one have an idea what's going on?

Thanks in advance,
Roberto Fichera.

[roberto@cluster4 TestOpenMPI]$ orterun -wdir
/data/roberto/MPI/TestOpenMPI -np
1 testmaster 10000 $PBS_NODEFILE
Initializing MPI ...
Loading the node's ring from file
'/var/torque/aux//909.master.tekno-soft.it'
... adding node #1 host is 'cluster3.tekno-soft.it'
... adding node #2 host is 'cluster2.tekno-soft.it'
... adding node #3 host is 'cluster1.tekno-soft.it'
... adding node #4 host is 'master.tekno-soft.it'
A 4 node's ring has been made
At least one node is available, let's start to distribute 10000 job
across 4
nodes!!!
****************** Starting job #1
****************** Starting job #2
****************** Starting job #3
****************** Starting job #4
Setting up the host as 'cluster3.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task './testslave.sh' on node 'cluster3.tekno-soft.it'
Setting up the host as 'cluster2.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task './testslave.sh' on node 'cluster2.tekno-soft.it'
Setting up the host as 'cluster1.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task './testslave.sh' on node 'cluster1.tekno-soft.it'
Setting up the host as 'master.tekno-soft.it'
Setting the work directory as '/data/roberto/MPI/TestOpenMPI'
Spawning a task './testslave.sh' on node 'master.tekno-soft.it'
--------------------------------------------------------------------------

A daemon (pid unknown) died unexpectedly on signal 1  while
attempting to
launch so we are aborting.


There may be more information reported by the environment (see  
above).



This may be because the daemon was unable to find all the needed  
shared

libraries on the remote node. You may set your LD_LIBRARY_PATH to
have the
location of the shared libraries on the remote nodes and this will
automatically be forwarded to the remote nodes.
--------------------------------------------------------------------------

[cluster4.tekno-soft.it:21287] [[30014,0],0] ORTE_ERROR_LOG:
Resource busy in
file base/plm_base_receive.c at line 169
[cluster4.tekno-soft.it:21287] [[30014,0],0] ORTE_ERROR_LOG:
Resource busy in
file base/plm_base_receive.c at line 169


_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
























					Reply via email to