The Case for NOW
We propose to build hardware and software to enable a network of workstations (NOW) to act as a single large-scale computer. Because of volume production, commercial workstations today offer much better price/performance than the individual nodes of MPPs; in addition, switch-based networks such as ATM will provide cheap, high-bandwidth communication. This price/performance advantage is increased if the NOW can be used for both the tasks traditionally run on workstations and large programs. We hope to demonstrate a practical 100 processor system in the next few years that delivers at the same time (1) better cost-performance for parallel applications than a massively parallel processing architecture (MPP) and (2) better performance for sequential applications than an individual workstation (by using more of the resources of the network). If projects like NOW are successful, they have the potential to redefine the high-end of the computing industry.
To realize the potential of NOWs, we need to move two MPP technologies into the workstation community: low latency networking and global system software that treats a collection of processors, memory, and disks as if they were a single machine. Our approach is to leverage off-the-shelf technology as much as possible -- workstation hardware, standard workstation operating systems on each node, and local area network ATM switches. To this, we will add communications protocol software and a global system layer that together provide low overhead communication, a single view of operating system services across the cluster, parallel file I/O, and robustness to individual node failures. We will demonstrate our results by using our system for the everyday computing needs, both sequential and parallel.