The data diffusion space for parallel computing in clusters

Abstract

The data diffusion space (DOS) is an all-software shared address space for parallel computing on distributed memory platforms. It is an extra address space to that of each process running a parallel application under the SPMD (Single Program Multiple Data) model. The size of DDS can be up to 2 64 bytes, either on 32- or on 64-bit architectures. Data laid on DDS diffuses, or migrates and replicates, in the memory of each processor using the data. This data is used through an interface similar to that used to access data in files. We have implemented DDS for PC clusters with Linux. However, being all-software, DDS should require little change to make it immediately usable in other distributed memory platforms and operating systems. We present experimental results on the performance of two applications both under DDS and under MPI (Message Passing Interface). DDS tends to perform better in larger processor counts, and is simpler to use than MPI for both in-core and out-of-core computation. © Springer-Verlag Berlin Heidelberg 2005.