Experiences with hypercube operating system instrumentation
International Journal of High Speed Computing - Special issue: session on parallel systems performance
Performance Measurement Intrusion and Perturbation Analysis
IEEE Transactions on Parallel and Distributed Systems
SPLASH: Stanford parallel applications for shared-memory
SPLASH: Stanford parallel applications for shared-memory
Performance observability
Architecture-oriented visualization
Proceedings of the 11th ACM SIGPLAN conference on Object-oriented programming, systems, languages, and applications
An Open Visual Model for Object-Oriented Operating Systems
IWOOOS '95 Proceedings of the 4th International Workshop on Object-Orientation in Operating Systems
Hi-index | 0.00 |
The nascent and rapidly evolving state of parallel systems oftenleaves parallel application developers at the mercy of inefficient,inflexible operating system software. Given the relativelyprimitive state of parallel systems software, maximizing theperformance of parallel applications not only requires judicioustuning of the application software, but occasionally, thereplacement of specific system software modules with others thatcan more readily respond to the imposed pattern of resourcedemands. To assess the feasibility of application and performancetuning via malleable system software and to understand theperformance penalties for detailed operating system performancedata capture, we describe a set of performance instrumentationtechniques for parallel, object-oriented operating systems and aset of performance experiments with Choices, an experimental,object-oriented operating system designed for use with parallelsys- tems. These performance experiments show that (a) theperformance overhead for operating system data capture is modest,(b) the penalty for malleable, object-oriented operating systems isnegligible, but (c) techniques are needed to strictly enforceadherence of implementation to design if operating system modulesare to be replaced.