[mpich-discuss] [mpich2-dev] ask for help in mpich2

[James Dinan]

Hi Zony,

To reinforce what Jeff said, I also would strongly encourage you to 
explore a solution on top of MPI before modifying the MPI 
implementation.  There may be a solution that works within the space of 
the MPI standard; this is desirable since it will be portable and easier 
to maintain.

For example, one could imagine a client-server model where a group of 
MPI processes remain active on server nodes and receive requests to 
perform calculations from serial client applications.  Such requests 
could be managed using sockets if the desire is to not use MPI at all in 
the serial client code.  These workers could be launched as a part of 
running the serial program, or they could be persistent.  If the latter, 
an additional server process that interfaces between clients and workers 
could partition resources and schedule computations to serve multiple 
serial client codes.

MPI also provides the capability for dynamic connections that would 
allow you to connect a serial MPI client program to an already running 
group of MPI worker processes.  Forking processes from the sequential 
program using something like MPI_Comm_spawn() could also be an option, 
but, a word of caution on dynamic processes, this routine is not well 
supported across all MPI implementations.  An additional option that you 
might find interesting is Adaptive MPI 
(http://charm.cs.uiuc.edu/research/ampi), which is developed by the 
Charm++ group.

Best,
  ~Jim.


On 11/28/11 9:23 AM, Jeff Hammond wrote:
> Hi Zony,
>
> (mpich-discuss at mcs.anl.gov <mailto:mpich-discuss at mcs.anl.gov> is the
> appropriate list for such messages.)
>
> I strongly recommend that you work with your colleagues to learn to use
> MPI the normal way.  Hacking MPICH2 to do non-standard-compliant things
> will only lead to bigger problems for them later.
>
> MPI is already an incredibly rich programming model and it is more than
> sufficient for the vast majority of parallel applications.  I do not
> know of a single application that cannot be implemented using MPI and
> the effort required is usually rather modest.  Many application
> developers find that they need to learn only a few (~10) MPI functions
> in order to parallelize their code, even for the largest supercomputers
> in the world.
>
> There are many books on parallel programming with MPI.  One you might
> consider is by Peter Pacheco.  See also "Using MPI-1", which is written
> by some of the developers of MPICH2.
>
> Best,
>
> Jeff
>
> On Mon, Nov 28, 2011 at 9:13 AM, sun big man <332344727 at qq.com
> <mailto:332344727 at qq.com>> wrote:
>
>     hi professioners:
>          Many of my colleages complait that the parallel program is so
>       difficult to write.so now i have an idea to make some changes on
>     mpich.
>          generally,i prepare to integrate all the parallel computation
>     our lab may use into libraries.so others will just call the parallel
>     function in the libraries in traditional serial program way.i mean i
>     could fork several processes in one node and then migrate them to
>     others in the libraries. but all the work will be thransparent to
>     the ordinary programers.I know it will broke some items of the mpi
>     standard,but what we care is the easiness in programming.
>         when i just begin my job,i find it's hard for me to clarify the
>     functions of mpich implementation.so i preciate if you coule send me
>     some details in mpich's implementation ,such as the latest design
>     document,the architecture of mpich2, the realtionship of the
>     functions,the original intentions .etc.
>          i also preciate for your other ideas on my job.
>         i hope i have make myself understood for my poor english.
>         thank you so much
>     Zony from China
>     2011.11.18
>
>
>
>
> --
> Jeff Hammond
> Argonne Leadership Computing Facility
> University of Chicago Computation Institute
> jhammond at alcf.anl.gov <mailto:jhammond at alcf.anl.gov> / (630) 252-5381
> http://www.linkedin.com/in/jeffhammond
> https://wiki-old.alcf.anl.gov/index.php/User:Jhammond
>
>
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss