In the most common sense all processors of a shared-memory machine have direct access to the same physical memory via a global bus. This results in the (global) shared-memory programming model, which is considered superior against the message-passing programming model, because it gives the most intuitive view on how a parallel application accesses its data.
Closer examination of typical problems reveals that, for a broad class of problems, global shared-memory is actually 'oversized'. Local sharing of data is often sufficient to solve the problems.
Multiprocessors with global shared-memory all suffer from a lack of scalability. Even the use of fast special buses and caching techniques can only postpone this flaw. These systems have an upper limit on the number of processing nodes.
The MEMSY system demonstrates the approach towards a highly scalable MIMD multiprocessor architecture which utilizes memory shared between a set of adjacent nodes as communication medium. We refer to this kind of shared-memory as distributed shared-memory.
The MEMSY system shall continue the line of systems which have been built at Erlangen using distributed shared-memory (e.g. EGPA and DIRMU).
The MEMSY project is founded by the Deutsche Forschungsgemeinschaft as part of the Sonderforschungsbereich 182.