Communication costs versus computation costs in parallel Gaussian elimination
Proceedings of the international workshop on Parallel algorithms & architectures
Communications of the ACM
The illusion of reality
Characterizing computer performance with a single number
Communications of the ACM
The Computer Journal
Measuring parallel processor performance
Communications of the ACM
The improbable machine
A structural analysis algorithm for massively parallel computers
Parallel supercomputing: methods, algorithms and applications
Performance evaluation of supercomputers
Performance evaluation of supercomputers
Performance Evaluation and Monitoring
ACM Computing Surveys (CSUR)
Hi-index | 0.00 |
Rigorous taxonomies are an important step in the development of science and technology, especially as an aid to paradigmatic clarity and the development of prescriptive terminology. This article explores the structure of the conceptual space of performance metrics. Some basic concepts of taxonomic methods and dimensional analyses are presented, along with templates for multi-dimensional taxonomies. A 48-way taxonomy of performance metrics is suggested, based on the four generic information-technology functions (processing, storage, communications, interface), two metric types (fundamental, derived), two metric scales (absolute, relative), and three implementation levels (system hardware, system software, applications). Metrics analogous to length, mass, and time from physics are suggested for information technologies. Innovative metrics for parallel processing are discussed, including Hockney's 's"1"/"2' (synchronization overhead) and 'f"1"/"2' (communication overhead), a serial-fraction metric, and an incremental-efficiency metric.