POSH

A more or less up-to-date version of POSH is available on GitHub. Right now you can find v0.11 there.

New! A nice autoconfautomake system has been added to POSH. Here comes v0.1.1!

It can be downloaded from here: [GZ] or [BZ2].

It requires:

a C++11 compiler
flex
perl
some parts of Boost: boost.Interprocess, Boost.Threads, Boost.System

The current and very first version of POSH is v0.1. It is very preliminary. It does not feature any nice automake/autoconf configuration and compilation system yet (but it comes with a Makefile), the algorithms used inside are quite simplistic, but it works.

It can be downloaded from here: [GZ] or [BZ2].

C. Coti: POSH: Paris OpenSHMEM: A High-Performance OpenSHMEM Implementation for Shared Memory Systems, in Proceedings of the International Conference on Computational Science (ICCS 2014), Cairns, Aus, June 2014. [PDF] [WWW]
C. Coti: POSH: Paris OpenSHMEM: A High-Performance OpenSHMEM Implementation for Shared Memory Systems, in CoRR abs/1403.7791 [cs.DC], March 2014, 13 pages [arXiv]

C. Coti: POSH: Paris OpenSHMEM, prestented at the International Conference on Computational Science (ICCS 2014), Cairns, Aus, June 2014. [PDF].

The memory model specified by OpenSHMEM and implemented by POSH consists of two parts for each processing element:

Its private memory, that cannot be accessed by other processing elements;
Its symmetric heap, which is public and can be accessd in read and write mode by other processing elements.

Global, static variables are put inside of the symmetric heap.

The fact that it public heaps are symmetric is important for memory management. As a matter of fact, memory allocations in the symmetric heap must be done in a symmetric way, in a sense that all the processing elements perform the same memory allocation. The OpenSHMEM specification states that before exiting, allocations in the symmetric heap must call a synchronization barrier in order to ensure this symmetry.

OpenSHMEM specifies one-sided peer-to-peer communications between processing elements. Basically, a process can read data from another process's symmetric heap or write data in another process's symmetric heap.

Memory is allocated in symmetric heaps using the shmalloc() routine. A family of shmem_*_put() and shmem_*_p() routines are available (one for each supported datatype) to write data at a remote address.

long source[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
static long* target = (int *) shmalloc( 10 * sizeof( long ) );

if (_my_pe() == 0) {
/* put 10 elements into target on PE 1 */
shmem_long_put(target, source, 10, 1);
}

Similarly, a family of shmem_*_get() and shmem_*_g() routines are available (one for each supported datatype) to read remote data.

short* ptr;
ptr = (short*)shmalloc( sizeof( short ) );
short toto = shmem_short_g( ptr, ( rank + 1 ) % size );