SNOW researcher Andre Luis Gobbi Sanches will endure a scientific mission to Fraunhofer FIRST in Berlin, from July 1st till August 31th. During this mission, he will investigate alternatives to implement the MPI (Message Passing Interface) as a collection of configurable software components in the realm of Application-Oriented System Design. MPI is the major standard for process communication in clusters of workstations. Therefore it is strongly needed in SNOW.
We were able to find free implementations of MPI of excellent quality, but all of them are monolithic implementations. In SNOW, we will only need a subset of the functions specified in MPI, so there is no reason in paying the overhead of such ``fat'' implementations. At the present moment, a working version of EPOS requires a few kilobytes of memory, while a generic MPI implementation would enlarge it to more than a megabyte. This does not align with SNOW requirements of bringing its parallel execution environment close to high-performance embedded applications. With an implementation based on a collection of software components designed according to guidelines of Application-Oriented System Design, only the features really needed by applications would be present in a given instance of our MPI implementation. Such instances shall be considerably smaller than ordinary monolithic implementations.
As mentioned before, only a subset of MPI will be implemented. In an earlier version, our aim is to provide basic funcionality for SNOW, so the implementation will run only in LINUX and EPOS, using EPOS' MYRINET network abstraction as the inter-node communication substrata. Efforts will be made, however, to make the implementation portable to other networks. Indeed, the desired portability and adaptability shall be delivered by EPOS family of network adapters.
The Fraunhofer FIRST Institue, a German partner of SNOW, has large experience in Operating System and Software Engineering. Moreover, they have been using MPI in many of their projects for years. So, this mission will bring valuable know-how on the subject to LISHA, being of fundamental importance to the development of this novel MPI implementation.