Improving interactive system performance using tipme

摘要: This thesis presents a new measurement methodology especially designed to improve the performance of interactive systems as perceived by user (user-perceived performance). Current and tuning techniques suffer from multitude problems when applied systems. Our reliance on these for system has caused be tuned in suboptimal manner with often failing provide predictable performance. Current concentrate improving through-put rather than latency. These also tend measure behavior under tightly-controlled situations. approach makes it highly unlikely discover infrequent that occur result unexpected interactions among several agents system, such operating kernel, application programs, various servers daemons. We have devised addresses weaknesses current techniques. is determine causes normal use plague users. We accomplish this goal using continuous monitoring postmortem analysis. Once exact cause problem determined, we verify our analysis constructing microbenchmarks recreate load condition exhibit problem. then evaluate possible remedies reproducible manner. Using devised, identified BSD/OS kernel common algorithms would otherwise had gone undetected. In one instance, processing pagefault taken X server was delayed seconds because allowed pageout daemon monopolize pre-allocated structure needed initiate page-in request. The CSCAN disk head scheduling algorithm further contributed not favoring latency-critical, requests over less timing-critical, requests. another an oversight process priority calculation newly-created, compute-bound processes starve processes.