State occupancy information for performance comparisons
Gordon Lyon · 1990
A state-based performance characterization attaches fixed processing rates to each service state.However, the number of states can be large.Over time, the sequences of such states are enormous.Counts of active (hut interchangeable) system elements define macrostates, which are fewer.Furthermore, only the occupancy levels of macrostates are recorded.This removes time sequencings as a combinatoric problem, but still captures general performance details.Applications can be compared independently of their algorithmic structures.The hypercube and other distributed-memory systems bring both opportunities and challenges to a state-based approach.Certainly, processor and communication activities are more easily identified and more independent with distributed-memory than with sharedmemory.But isolated nodes also entail problems in capturing global observation states.However, hypercube application codes are commonly homogeneous across nodes, so that aggregating local state information works well.Three paradigms illustrate homogeneous applications with communication dependencies that are strong (global), moderate flocal), or weak (independent once spawned).The three performance summaries are accurate and extremely compact. Key words: application comparisons; distributed-memory