High-performance parallel applications running on massively parallel, distributed architectures require efficient resource management and optimized implementations of communications, input/output, processor, and memory management. We propose a customizable operating system architecture that can be optimized for individual parallel applications. Our approach is to organize such an operating system as an object-oriented system. Our long-term goal is to show that carefully designed object-oriented operating systems can allow high-performance parallel applications to be ported and run efficiently and economically on distributed parallel processors built out of stock or commercially available ``off-the-shelf'' components.

The advent of killer networks high-speed, low-cost networks capable of 1 to 10 gigabits of bandwidth enable a new class of applications for clusters and workgroups (sometimes called NOW). Our objective is to explore novel networking, operating systems scheduling, input/output, and network interface architectures to enable these new high-performance distributed applications. For example, HPVMs can reduce the effort required to build efficient parallel applications on distributed resources, increase their performance, or make completely new applications possible. We are building real prototypes and have a large experimental cluster with 64 processors, multigigabyte memories, and large disk storage capacity.

Operating systems need new algorithms, services, and protocols to support processing of audio/visual streams according to quality of service (QoS) specification and provide graceful degradation in case resource availability decreases. This project concentrates on the design of a brokerage service with admission and adaptation control of soft real-time and nonreal-time tasks and its negotiation protocols between application and the OS broker. An integrated time/event scheduler will provide control for guarantees as well as notification if deadlines cannot be met. The goal is to provide guaranteed services in OS and integration of guaranteed services between the OS and the network to achieve end-to-end QoS guarantees for distributed multimedia systems.