Charm++ parallel programming system

["Galvez Garcia, Juan Jose" <jjgalvez AT illinois.edu>]

Hi Phil,

 

If you have or know of Python applications using mpi4py we would be interested in knowing about these, not just for a potential AMPI4py, but also maybe to port to CharmPy, and for benchmarking purposes.

 

An alternative, as you know, to benefit from Charm++ with Python, is to write applications using Charmpy. A lot has changed since last SC, and we are moving towards a new release soon. We have basic documentation at http://charmpy.readthedocs.io. Most Charm++ features are supported now, including load balancing. The API has been simplified further, and performance has improved (using cython and other techniques). More recently, we added a lot of support for “futures” using threaded entry methods, which can simplify code a lot. Entry method calls, including broadcasts, can return futures on an as-needed basis. Futures can be passed around, and be used as reduction targets (details are in recent commits: https://github.com/UIUC-PPL/charmpy/commits/master ). I also tested charmpy on Windows (next release will have binaries, so applications can be written and executed on Windows without a compiler).

 

-Juan

 

From: Eric Bohm
Sent: Wednesday, June 20, 2018 4:56 PM
To: Miller, Philip B
Cc: charm
Subject: Re: [charm] Experiences with MPI-4-Py built against AMPI?

 

It is in asana.  Until you asked about it, no one else had expressed any interest in MPI4Py on AMPI, and there has been some skepticism regarding whether there are enough users of MPI4Py to warrant the effort.  I can copy our findings over to a redmine issue if you like.

On Jun 20, 2018, at 4:11 PM, Phil Miller <mille121 AT illinois.edu> wrote:

Are those in the public Redmine instance, or somewhere private?

On Wed, Jun 20, 2018 at 3:19 PM, Eric Bohm <ebohm AT illinois.edu> wrote

I had an intern working on that over the spring at Charmworks.  We have a list of issues documented from that in our todo list for AMPI feature support.

> On Jun 20, 2018, at 3:06 PM, Phil Miller <mille121 AT illinois.edu> wrote:
>
> Has anyone experimented with building the Python mpi4py library on AMPI and running Python MPI applications on it? If so, how did it work out? Were there any particular challenges to work through?