CPA-CPL: Feedback-Directed Resource Management in Virtual Private Machines

点击翻译按钮获取中文摘要

In modern computers with multiple CPUs per chip, multiple levels ofcomputer memory are shared among the CPUs. This sharing may causevarying response times for computer applications, potentiallyaffecting their usability. For instance, a video streamingapplication that has unpredictable response times due to interferencefrom other applications may be essentially unusable. Thus, thecomputer system must attempt to allocate shared memory resourcesequitably to increase response consistency. Virtual Private Machines(VPMs), in which shared computer resources are partitioned to givethe illusion that a program is running on a single physical CPUrunning in isolation, are one method to provide responseconsistency. This research involves the development of softwaremodeling techniques to predict the memory requirements of applicationsrun on modern computers and using those predictions to guide memoryresource allocation to provide consistent, predictable performance.The focus of this research consists of the development ofreuse-distance-based memory locality analysis for multi-threadedapplications to predict the resource requirements of thoseapplications. This research will apply the new memory locality modelsto manage the cache and bandwidth requirements for multi-threadedapplications run on a VPM within a multi-programmedenvironment. Specifically, this research will examine reuse distancefor Partitioned Global Address Space (PGAS) applications written inUnified Parallel C (UPC). The research will investigate the effects ofthe number of threads used per application on per-thread data size andthe effect on reuse distance of shared data caused by threadinteraction. The result will be novel solutions to memory localityanalysis for multi-threaded applications while allowing the compilerto predict and specify the resources needed to give multi-threadedapplications targeted performance.