Abstract (EN):
We present the architecture of a parallel programming system, the di_pSystem. Our target machine consists of clusters of multiprocessors interconnected with very fast networks. The system aims to provide a programming style close to the shared memory programming model. This is achieved by a software layer between the programmer and the operating system that supports the communication between individual computational agents and that dynamically balances the work-load in the system. This layer effectively hides much of the complexity of programming distributed architectures from the programmer while being competitive in performance. The low-level communication in the di_pSystem is implemented using MPI as a backbone. Initial results indicate that the system is close in performance to MPI, a fact that we attribute to its ability to dynamically balance the work-load in the computations. © Springer-Verlag Berlin Heidelberg 1999.
Language:
English
Type (Professor's evaluation):
Scientific
No. of pages:
8