Detail Improving interactive system performance using TIPME Bibliografi Author: Endo, Yasuhiro ; Seltzer, Margo (Advisor) Topik: COMPUTER SCIENCE Bahasa: (EN )    ISBN: 0-599-64045-6     Penerbit: Harvard University Press     Tahun Terbit: 2000     Jenis: Theses - Dissertation Fulltext: 9960496.pdf (0.0B; 2 download) Abstract This thesis presents a new measurement methodology especially designed to improve the performance of interactive systems as perceived by the user (user-perceived performance). Current performance measurement and tuning techniques suffer from a multitude of problems when applied to interactive systems. Our reliance on these techniques for interactive system performance tuning has caused the systems to be tuned in a suboptimal manner with systems often failing to provide predictable performance. Current performance measurement techniques concentrate on improving through-put rather than latency. These techniques also tend to measure system behavior under tightly-controlled situations. This approach makes it highly unlikely to discover infrequent performance problems that occur as a result of unexpected interactions among several agents in the system, such as the operating system kernel, application programs, and various servers and daemons. We have devised a methodology that addresses the weaknesses of current measurement techniques. Our methodology is designed to determine the causes of performance problems that occur in interactive systems under normal use and plague users. We accomplish this goal using continuous monitoring and postmortem analysis. Once the exact cause of a performance problem is determined, we verify our analysis by constructing microbenchmarks that recreate the load condition that causes the system to exhibit the problem. We then use these microbenchmarks to evaluate possible remedies in a reproducible manner. Using the methodology devised, we have identified the exact causes of several problems with BSD/OS operating system kernel and common algorithms that would otherwise had gone undetected. In one instance, processing of a pagefault taken by the X server was delayed for several seconds because the system allowed the pageout daemon to monopolize a pre-allocated structure needed to initiate a page-in request. The CSCAN disk head scheduling algorithm further contributed to this problem by not favoring latency-critical, page-in requests over less timing-critical, pageout requests. In another instance, an oversight in the process priority calculation algorithm caused newly-created, compute-bound processes to starve interactive processes. Opini Anda Klik untuk menuliskan opini Anda tentang koleksi ini! Lihat Sejarah Pengadaan  Konversi Metadata   Kembali

Copyright © 2006, 2007 Unika Atma Jaya, all rights reserved