System support for end-to-end performance management

摘要: This dissertation introduces, implements, and evaluates the novel concept of "Service Paths", which are system-level abstractions that capture describe dynamic dependencies between different components a distributed enterprise application. Service paths because they natural interactions application services dynamically composed to offer some desired end user functionality. such sets run on networked machines in data centers. cross multiple levels abstraction link like web browsers with system http providing communications embedded hardware-supported encryption. created without user, application, or middleware input, but despite these facts, able application-relevant performance metrics, including end-to-end latencies for client requests contributions from application-level processes software/hardware resources protocol stacks network devices. Unique service classes requests, thus making notion request class basic quantum management support. Finally, constructed by software modules into operating systems participating machines. These measure analyze program behavior. Beyond conceiving demonstrating their utility, this thesis makes three concrete technical contributions. First, we propose set signal analysis techniques called "E2Eprof" identify taken across IT infrastructure time spent each path. It uses algorithm "pathmap" computes correlation message arrival departure timestamps at node detect among them. A second contribution is monitoring toolkit "SysProf", captures information granularity, ranging tracking systemlevel activities triggered single call, capturing client-server associated paths, characterizing server consumed clients behaviors. SysProf operates enterprise-wide it can track application- resource usage used typical infrastructures. The third publish-subscribe based delivery framework "QMON". QMON offers high predictability supports utility-aware while also differentiate monitoring, corresponding SLAs maintained applications. Service Paths E2EProf methods implementing them shown useful evaluated enterprise-class applications, include shared file service, multi-tier auctioning site an Operational Information System (OIS) our industrial partners. For automatically then find bottlenecks those paths. In addition, enforce adapting middleware's application's scheduling dispatching behavior classes. assist tuning provisioning, using when decisions adding servers, adjusting number threads changing buffer sizes, etc. demonstrate efficacy general approach runtime work first isolate problematic nodes path use additional relatively more intrusive capabilities "zoom" perform micro-level exact workloads, components, and/or causing problem. (Abstract shortened UMI.)