The interaction of parallel and sequential workloads on a network of workstations
Proceedings of the 1995 ACM SIGMETRICS joint international conference on Measurement and modeling of computer systems
Effective distributed scheduling of parallel workloads
Proceedings of the 1996 ACM SIGMETRICS international conference on Measurement and modeling of computer systems
The utility of exploiting idle workstations for parallel computation
SIGMETRICS '97 Proceedings of the 1997 ACM SIGMETRICS international conference on Measurement and modeling of computer systems
Automatic node selection for high performance applications on networks
Proceedings of the seventh ACM SIGPLAN symposium on Principles and practice of parallel programming
Linger Longer: fine-grain cycle stealing for networks of workstations
SC '98 Proceedings of the 1998 ACM/IEEE conference on Supercomputing
Predicting the CPU Availability of Time-Shared Unix Systems on the Computational Grid
HPDC '99 Proceedings of the 8th IEEE International Symposium on High Performance Distributed Computing
An Evaluation of Linear Models for Host Load Prediction
HPDC '99 Proceedings of the 8th IEEE International Symposium on High Performance Distributed Computing
Buffered Coscheduling: A New Methodology for Multitasking Parallel Jobs on Distributed Systems
IPDPS '00 Proceedings of the 14th International Symposium on Parallel and Distributed Processing
Elections in a Distributed Computing System
IEEE Transactions on Computers
| Hi-index | 0.00 |
In this paper we study the use of idle cycles in a network of desktop workstations under unfavourable conditions: we aim to use idle cycles to improve the responsiveness of interactive applications through parallelism. Unlike much prior work in the area, our focus is on response time, not throughput, and short jobs - of the order of a few seconds. We therefore assume a high level of primary activity by the desktop workstations' users, and aim to keep interference with their work within reasonable limits. We present a fault-tolerant, low-administration service for identifying idle machines, which can usually assign a group of processors to a task in less than 200ms. Unusually, the system has no job queue: each job is started immediately with the resources which are predicted to be available. Using trace-driven simulation we study allocation policy for a stream of parallel jobs. Results show that even under heavy load it is possible to accommodate multiple concurrent guest jobs and obtain good speedup with very small disruption of host applications.