Abstract :
Despite the large research efforts in the SW-DSM community, this technology has not yet been adapted widely for significant codes beyond benchmark suites. One of the reasons contributing to this is the existence of a large variety of different, incompatible systems and APIs, which severely restricts portability. This work therefore proposes a DSM framework, called HAMSTER, which allows the low complex implementation of basically arbitrary DSM APIs on top of a single core and thereby enables the easy retargeting of the system to specific application needs. This flexibility is achieved by offering a comprehensive set of shared memory services grouped into orthogonal modules. Combined, they offer a comprehensive interface capable of supporting the different requirements of the various DSM APIs.
