Characterizing applications by integrating andimprovingtools for data locality analysis and programperformance
Institution: | The Ohio State University |
---|---|
Department: | |
Year: | 2017 |
Keywords: | Computer Science; Computer Engineering; data locality analysis, convex partitioning, HPCToolkit,CDAG, dynamic dependence graphs, DDG, data movement costs, diskbased cache, program performance, profiling, directed acyclic graphrepresentation |
Posted: | 02/01/2018 |
Record ID: | 2205747 |
Full text PDF: | http://rave.ohiolink.edu/etdc/view?acc_num=osu1492741656429829 |
Data locality is a critical factor which affects theexecution time of applicationstoday. With major advances being madein reducing the computation timeof processors, data movement costshave increasingly become a bottleneck inrunning time and energyefficiency of current applications. They help in gainingusefulinsights into programs behaviour for a given execution. The workin thisthesis extends an existing dynamic analysis framework whichhas been used todevelop dynamic analysis tools e.g. a tool toidentify vectorization potential ofexisting programs, a toolresponsible for characterizing and assessing the inherentdatalocality properties of a given computation. This existingframeworkis based on construction and analysis of the dynamicdependence graph for agiven execution.The framework is not wellintegrated with existing tools that analyze andreport programperformance. This makes the data locality analysissomewhatinaccessible for popular use. In this thesis, we try tobridge that gap by integratingthe tools for the different analysisand reporting the result in a coherentway. We also report theimprovements to the tools in considerationduring this endeavorincluding enabling scalable analysis of large dynamicdependencegraphs. Finally, we use the work to characterize certainwell knownbenchmarks.Advisors/Committee Members: Sadayappan, Ponnuswamy (Advisor).