Extending the message passing interface (MPI)

Performance enhancing factors for manycore architectures: State-of-the-art

“…By using cooperative operations it can combine communication and synchronization. Features like interrupt receive calls, remote memory access extensions and active message interface were extended to basic MPI specifications as listed in [9] [10]. On a heterogeneous environment, overlapping communication and computation between tasks is difficult.…”

Managing Multiple Communication Methods in High-Performance Networked Computing Systems

“…The concurrent execution of multiple lightweight threads within a single process is a useful technique for masking variable latencies, exploiting multiprocessors, and providing concurrent access to shared resources. Various approaches to the integration of multithreading into a message-passing framework have been proposed [2,12,18,37,39,13,25,42]. The Nexus implementation of MPI supports a particularly simple and elegant model that does not require that explicit thread identifiers be exported from MPI processes.…”

“…In particular, it is not possible to define a collective operation that involves more than one thread per process. This functionality requires extensions to the MPI model [18,26,37].…”

Improving load balancing in an MPI environment with resource management

“…The environmental management function supported by the standard provides facilities that a programmer needs to write correct and robust programs and this is clearly important for the construction of highlyportable message-passing programs. However, the original specification does not provide any support for dynamic load balancing [2], largely because it was considered more important to release a core standard quickly rather than an enhanced standard which would inevitably be delayed by a couple of years.…”