[Beowulf] Sharing an array in an MPI program?

[Bogdan Costescu]

On Fri, 25 May 2007, Jaime Perea wrote:

> One alternative that I like and it integrates well with mpi
> is the global arrays toolkit
>
> http://www.emsl.pnl.gov/docs/global/

I disagree with the "integrates well with mpi" part of the statement. 
Global Arrays works on top of a communication layer called ARMCI which 
only uses an MPI implementation for setting up the job and tearing it 
apart (only calling MPI_Init and MPI_Finalize from what I remember), 
the communication itself is done directly via lower level protocols 
(TCP, GM, etc.) - I know because some years ago I wanted to use Global 
Arrays on a SCore cluster and discovered that I have to port ARMCI to 
PM (the low level communication protocol of SCore)... This sometimes 
creates problems due to limitations imposed by the low level protocols 
(for example, the MPI implementation would open a GM port while ARMCI 
would need to open a second one, so the per-node limit of GM ports 
would be reached much faster).

[ I don't want the above statement to sound negative towards Global 
Arrays or ARMCI, my intention was only to bring to discussion a fact 
that was missing. ]

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De