The Quest for a Zero Overhead Parallel Machine

点击翻译按钮获取中文摘要

The primary goal of this project is to demonstrate the viability of cost- effective parallel computing techniques for two application domains that make computing ubiquitous. The first one is parallel databases embodying a clients/server paradigm used in data mining/data warehousing applications. The second one is the physical simulation of virtual multimedia environments that has potential use in several everyday situations. The challenge in this domain is to perform realistic physical simulation in real time of multiple actors (some real and some synthetic), each with several degrees of freedom. This research will investigate the system architecture issues in the design of parallel (cluster) servers engineered with commodity processing and network components for supporting such applications. A shared address space across the cluster is an attractive abstraction from the perspective of programming such applications. However, accessing and caching data from this shared address space can be expensive in terms of observed latencies in a cluster. Therefore, clever techniques that hide the communication overheads for accessing the shared address space are imperative to achieving good performance. The ultimate goals is to make the cluster server appear as though the applications incur zero overhead for communication. Specific research issues to be addressed in this research include (a) mechanisms for realizing shared address space across the cluster, (b) mechanisms for supporting variable granularity, caching, explicit communication, and synchronization (within the framework of shared space) tailored to the needs of these applications, (c) hardware assists necessary to make clusters appealing as cost-effective parallel machines, and (d) detailed performance evaluations of cluster servers under such multimedia and database workloads.