Overview of the Berkeley NOW Project David Culler, Tom Anderson, Dave Patterson Computer Science Division University of California, Berkeley Current technological trends suggest that the building block of choice for large-scale systems is the complete desktop machine with sizable memory, disk, and full-function operating system, connected by a fast network. The Berkeley NOW project seeks to understand the feasibility of this modular style of large-scale system design and the unique systems concepts that it enables. The research includes investigation of (1) low-overhead communication software and network interface hardware for "trusted-area" networks based on outgrowths of MPP single-chip switches, (2) "serverless" network file systems and software RAID, (3) global resource allocation and management, (4) efficient scheduling of parallel programs along with interactive workloads, and (5) performance analysis of such systems. If successful, such networks of workstations (NOWs) are likely to become the primary infrastructure for technical computing. Our approach is a staged development using highly leveraged commercial hardware and software components leading up to a demonstration system of 100 processors in state-of-the-art workstations. Our first prototype was a 4-node cluster of HP 9000/735s using an experimental network interface, upon which a version of Active Messages with communication latency of 16us was constructed and user-lever job control and co-scheduling of parallel jobs was demonstrated. We are currently building our second prototype, a 32-node cluster is SUN SS-10s under Solaris and connected by Myrinet switches and interface cards. Substantial portions of the Global operating system layer, GLUnix, parallel file system, XFS, and communication layer, CAL, are operational. Initial application development is underway.