1SvPablo: A multi-language Architecture-Independent Performance Analysis SystemLuiz A. De Rose and Daniel A. ReedPresented by Jusub Kim11/08/05 CMSC714Goal of their work.z Development of a GUI tool kit which can…z Correlate sequential/ parallel source code with dynamic performance datafrom both software and hardware measurementsz Provide a portable, intuitive and easily used interface.2Their achievements.z SvPablo(source view pablo):z Is a graphical environment for instrumenting application source code and browsing dynamic performance data.z Is portable, language independent, and performance metric independent( enabled by using the SDDF-self describing data format ).z Supports for applications executing on both sequential and parallel systemsand exploits hardware support of performance counters.Details: How did they achieve that?z Performance Instrumentationz Hardware performance integrationz Performance Visualizationz Language and Architecture Transparency3Performance instrumentationz Supports both interactive and automatic instrumentation.z Automatic instrumentationz Support analysis of HPF code.z Why not allow interactive instrumentation?z HPF compiler does high-level optimizations and allowing users to instrument data parallel source code can potentially inhibit any of these optimizations, reducing performance.z Let HPF compiler emit instrumented code, which calls the svPablo library. It captures data for each executable line and every procedure call.Performance instrumentationz Interactive Instrumentationz Most compilers for sequential languages primarily focus on local optimizations.z Support interactive instrumentation of C and Fortran.z Restrict instrumentable constructs to outer loops and procedure calls.4Details: How did they achieve that?z Performance Instrumentationz Hardware performance integrationz Performance Visualizationz Language and Architecture TransparencyHardware performance integrationz Software instrumentaion is simply not enough! We need to consider also…z Superscalar instruction execution effectz Cache effect in hardware-managed distributed shared memory hierarchyz New microprocessors commonly provide a set of performance registers. (cycles, cache misses, floating point instructions, branch misprediction, etc.)5Hardware performance integrationz During program execution, the SvPablo library queries the user-selected counters and records them.z After program execution, the SvPablo records its statistical analyses in a set of summary files, one for each executing process and merges them computing new global statistics, which is input to the analysis graphical interface. Details: How did they achieve that?z Performance Instrumentationz Hardware performance integrationz Performance Visualizationz Language and Architecture Transparency6Performance metricsColor boxes are clickableDetails: How did they achieve that?z Performance Instrumentationz Hardware performance integrationz Performance Visualizationz Language and Architecture Transparency7Language and Architecture Transparencyz Requires a flexible specification mechanism for instrumentation points and performance metrics.z SDDF (self-describing data format) separates performance data presentation from language and architecture z A group of record descriptors and record instances.z Three groups of record descriptors: mapping, configuration, and statistics.8Language and Architecture Transparencyz SDDF main features.z Portability: portable directly across systems, in binary or ASCII formatz Generality: a variety of events or data types can be supportedz Extensibility: tool developers can easily add new metrics Language and Architecture Transparencyz Event Statisticsz Performance data is represented by a Event Mapping record and a set of Statistics recordsz Procedure Statisticsz Procedure statistics records define the performance metrics associated with all procedures: # of calls to the procedure and the exclusive duration of the procedure. etc.9Related Worksz Paradyn (U of Wisconsin)z a tool for measuring the performance of large scale parallel programsz Instrumentation and visualization are performed during run-time.z Pablo (U of Illinois)z Consists of several components for instrumenting and tracingparallel programs and for analyzing the trace file.z AIMS (NASA Ames Research)z A tool kit for parallel applications.z Support MPI and PVM.SvPablo Summary.z A graphical environment for instrumentingapplication source code and browsing dynamic performance data.z Support for not only software performance measurement but also access to hardware performancecounters.z Language and architecture transparency by representing performance data via XML-like
View Full Document
Unlocking...