Abstract: Our work explores the practical relevance of Randomized Shared Memory (RSM), a theoretical concept that has been proven to enable an (asymptotically) optimally efficient implementation of scalable and universal shared memory in a distributed-memory parallel system. RSM (address hashing) pseudo-randomly distributes global memory addresses throughout the nodes' local memories. High memory access latencies are masked through massive parallelism. This paper introduces the basic principles and properties of RSM and analyzes its practical efficiency in terms of constant factors through simulation studies, assuming a state-of-the-art parallel architecture. Bottlenecks in the architecture are pointed out, and improvements are being made and their effects assessed quantitatively. The results show that RSM efficiency is encouragingly high, even in a non-optimized architecture. We propose architectural features to support RSM and conclude that RSM may indeed be a feasible shared-memory implementation in future massively parallel computers.

Abstract: This paper addresses the problem of providing high-performance disk I/O in massively parallel computers. Resolving the fundamental I/O bottleneck in parallel architectures involves both hardware and software issues. We review previous work on disk arrays and I/O architectures aimed at providing highly parallel disk I/O subsystems. We then focus on the requirements and design of parallel file systems (PFSs) which are responsible to make the parallelism offered by the hardware and a declustered file organization available to application programs. We present the design strategy and key concepts of a general-purpose file system for a parallel computer with scalable distributed shared memory. The principal objectives of the PFS are to fully exploit the parallelism inherent among and within file accesses, and to provide scalable I/O performance. The machine model underlying the design is described, with and emphasis on the innovative architectural features supporting scalability of the shared memory. Starting from a classification of various scenarios of concurrent I/O requests, the features of the PFS design essential for achieving the goals are described and justified. It is argued that the inter- and intra-request parallelism of the I/O load can indeed be effectively exploited and supported by the parallel system resources. Scalability of I/O performance and of the PFS software can be ensured by avoiding serial bottlenecks through the use of the powerful architectural features.