Charm++ parallel programming system

["Van Der Wijngaart, Rob F" <rob.f.van.der.wijngaart AT intel.com>]

We love nosiness!! Very helpful, indeed, Dave. Moving +ppn after the app name does the trick, though charmrun still complains when I say

[rfvander@bar1 jacobi2d]$ ./charmrun +p1 ./jacobi2d 20 +ppn 4 +setcpuaffinity +pemap 0-3   

p = 1 should be a multiple of ppn = 4

So it  appears in +pN, N should be the total number of threads, i.e. number of nodes * number of threads per node. Is that intended?  I would expect N simply to the number of nodes. Thanks.

 

Rob

 

 

From: David Kunzman [mailto:dmkunzman AT gmail.com]
Sent: Friday, September 12, 2014 12:17 PM
To: Van Der Wijngaart, Rob F
Cc: Phil Miller; charm AT cs.illinois.edu
Subject: Re: [ppl] [charm] Problem building charm++ on Intel platform

 

Yeah, I'm nosey. :) I thought it might be helpful to know someone local who could help.

 

Try putting the "+ppn" after the app name, so it's a parameter to the executable. As I recall, it's processed by the runtime on each node individually. I seem to remember that charmrun does this for you (binary versions), but as Phil pointed out, when you use MPI things are just passed off from charmrun to mpirun. If you do that, your command should be portable across builds. If I remembered that correctly, perhaps it would be a good idea if the MPI build could be modified to recognize and move it as well (so parameters work the same way for all builds; and perhaps drop ++verbose if it is present).

 

This is all based off my memory, which is at least several years old now... So I might have some details wrong.

 

Thanks,

Dave K

 

On Sep 12, 2014, at 11:47 AM, "Van Der Wijngaart, Rob F" <rob.f.van.der.wijngaart AT intel.com> wrote:

Thanks, Phil! In the meantime Dave Kunzman (at Intel now, and collocated with me in Santa Clara!) started butting in on a separate thread. I have now built in this way to enable multi-threading:

./build charm++ mpi-linux-x86_64 ifort mpicxx smp -j8 -DMPICH_IGNORE_CXX_SEEK -DMPICH_SKIP_MPICXX

Now funny things start to happen, even before I specify your additional flags:

[rfvander@bar1 jacobi2d]$ ./charmrun +p1 +ppn4 ./jacobi2d 200                                                                                                                

p = 1 should be a multiple of ppn = 4

[rfvander@bar1 jacobi2d]$ ./charmrun +p4 +ppn4 ./jacobi2d 200                                                                                                                

 

Running on 1 processors:  +ppn4 ./jacobi2d 200

charmrun>  /usr/bin/setarch x86_64 -R  mpirun -np 1  +ppn4 ./jacobi2d 200

[proxy:0:0@bar1] HYDU_create_process (../../utils/launch/launch.c:590): execvp error on file +ppn4 (No such file or directory)

 

The first reported error is funny. It looks like a bug. The second apparently comes about because +ppn4 is passed to mpirun verbatim, which thus chokes (using “+ppn 4” gives the same result). If I change it to “–ppn 4”, mpirun is happy, but charm++ doesn’t realize I am asking for multiple threads. I guess you cannot have multiple threads and run them, too.

 

Rob

 

From: unmobile AT gmail.com [mailto:unmobile AT gmail.com] On Behalf Of Phil Miller
Sent: Friday, September 12, 2014 10:54 AM
To: Van Der Wijngaart, Rob F
Cc: charm AT cs.illinois.edu
Subject: Re: [charm] Problem building charm++ on Intel platform

 

Ooh, I see. I forgot you were on an MPI build of Charm++, rather than the IP-based network layer (net-* or netlrts-*) we normally use in development. The charmrun utility is different for each network layer, and I think only the net-* version of charmrun supports that option. On other layers, since it's not responsible for process launching, it has nothing to report.

The other flags I mentioned should be more helpful, though.

 

 

_______________________________________________
charm mailing list
charm AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/charm

_______________________________________________
ppl mailing list
ppl AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/ppl