ACL 2 for Parallel Systems Software : A Progress Report

A significant development in high-performance computing has occurred in recent years with the proliferation of “Beowulf” clusters [6]. Beowulf clusters are parallel computers assembled from commodity-priced personal computers and networks. The explosive growth of the personal computer marketplace, together with rapid technological advances in the hardware sold there, has driven the price/performance ratio so low for Beowulf clusters that they have become competitive with traditional tightly integrated (and thus expensive) supercomputers. The current bottleneck is systems software. Small clusters are already working well in a large variety of situations, as well as large systems devoted to single applications. To truly compete with highly parallel supercomputer systems, however, Beowulfs need scalable parallel libraries, system management tools, job schedulers, process managers, and other systems software to support a mix of users and applications on systems consisting of hundreds of network-connected computers. Traditional systems software either is not scalable or is tied to specific vendor systems. The Mathematics and Computer Science Division at Argonne National Laboratory conducts research and prototype software development in parallel systems software. The MPICH implementation [2] of the MPI standard for message passing is already widely used in the Beowulf community. A new project is MPD (for Multi-Purpose Daemon), whose purpose is to explore issues in scalable process management for parallel jobs on Beowulf clusters. Process management includes process startup, job control, management of standard I/O (especially for interactive jobs), security, reliable process