MPI with Globus WS-GRAM adaptor

Added by Scardaci Diego almost 13 years ago

Dear All,
we are trying to submit an MPI jobs using the WS-GRAM adaptor but we are experiencing some problem.

In particular, specifying the following parameters:
- SPMDVariation = MPI
- TOTALCPUCOUNT = 2
- NUMBEROFPROCESSES = 2

we got this kind of result:
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in

instead submitting the MPI job using the GLOBUS interface we got:
I am processor: 2 at xn07.ctsf.cdacb.in
I am processor: 1 at xn06.ctsf.cdacb.in
I am processor: 3 at xn06.ctsf.cdacb.in
I am processor: 0 at xn07.ctsf.cdacb.in

moreover the parameter PROCESSESPERHOST seems not supported.

Could you, please help us?

Thanks in advance,
Diego Scardaci

Replies (4)

RE: MPI with Globus WS-GRAM adaptor - Added by Schwarz Lionel almost 13 years ago

Hi Diego,
Are you able to compare the JDL sent by the Globus interface and the one sent by the JSAGA adaptor? This could be useful to debug.

Looking at the code, the parameter PROCESSESPERHOST seems to be supported. Together with the NUMBEROFPROCESSES, it generates something like:
<hostCount>NUMBEROFPROCESSES/PROCESSESPERHOST</hostCount>
in the JDL

There might be a bug, could you please open a ticket for this?
Thanks

Lionel

RE: MPI with Globus WS-GRAM adaptor - Added by Scardaci Diego almost 13 years ago

Hi Lionel,
in the globus interface we have two parameters:
'hostCount': 'hostCount' is the number of machines used.
'count': the number of MPI parallel thread

Then the SAGA mapping should be:
hostCount=TOTALCPUNUMBER=NUMBEROFPROCESSES/PROCESSESPERHOST
count=NUMBEROFPROCESSES

according to the adaptor implementation, which parameters I have to set?

Tomorrow I'll verify if I can get the code generated by the adaptor starting from the jsaga mpi sample.

I'm going to open a bug about that...

Cheers,
Diego

RE: MPI with Globus WS-GRAM adaptor - Added by Schwarz Lionel almost 13 years ago

Diego Scardaci wrote:

Hi Lionel,
in the globus interface we have two parameters:
'hostCount': 'hostCount' is the number of machines used.
'count': the number of MPI parallel thread

Then the SAGA mapping should be:
hostCount=TOTALCPUNUMBER=NUMBEROFPROCESSES/PROCESSESPERHOST
count=NUMBEROFPROCESSES

Indeed, that is what the adaptor does except that the TOTALCPUNUMBER seems not supported.
Try with NUMBEROFPROCESSES and PROCESSESPERHOST only.

Lionel

RE: MPI with Globus WS-GRAM adaptor - Added by Scardaci Diego almost 13 years ago

Hi Lionel,
I executed the following test:

1) NUMBEROFPROCESSES = 2 and PROCESSESPERHOST = 1

I got:
I am processor: 0 at xn06.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in

2) NUMBEROFPROCESSES = 4 and PROCESSESPERHOST = 2

I got:
I am processor: 0 at xn06.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in

3) NUMBEROFPROCESSES = 6 and PROCESSESPERHOST = 2

I got:
I am processor: 0 at xn06.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in
I am processor: 0 at xn03.ctsf.cdacb.in

very strange results...

Cheers,
Diego

(1-4/4)

Project

General

Profile

JSAGA