Presentation is loading. Please wait.

Presentation is loading. Please wait.

Tracing and Performance Analysis Tools for Heterogeneous Multicore System by Soon Thean Siew.

Similar presentations


Presentation on theme: "Tracing and Performance Analysis Tools for Heterogeneous Multicore System by Soon Thean Siew."— Presentation transcript:

1 Tracing and Performance Analysis Tools for Heterogeneous Multicore System
by Soon Thean Siew

2 Goals Collection profiles and traces on heterogeneous multicore platform, targets on Cell B.E architecture. Performance analysis and simulation based on the profiles and traces. Performance tuning based on the result of analysis/simulation. Performance visualization. System architecture evaluation based on trace-driven simulation. A trace-driven simulation produces a trace of memory references (for data and instructions). The trace can be used to model memory system performance & to model an instruction pipeline evaluate the performance of a specific multiprocessor with respect to a given workload, if traces concerning this workload are available

3 Trace Collecting Handler Library
Tool Overview Preprocessing Runtime Trace Collection Post-processing and Visualization Source Code Instrumented Trace API Visualization Tools Report Executable Profile Instrumentation Tool Trace Collecting Handler Library Converter Compiler Architecture Dependent Trace-driven Simulation

4 Components Trace API Trace Collecting Handler Library Converter
Serves as a protocol between runtime trace collection and post-processing Trace Collecting Handler Library A set of analysis routines which allocate collecting data into proper predefined structures Converter Transforms collecting trace into suitable format which is to be fed into visualization tools

5 Progress I Literature research on current available performance analysis tools for Cell BE: No robust tool that is capable for collecting computation and communication traces simultaneously. PDT mainly focus on SPE/communication specific problem Oprofile supports profiling on the PPU events and SPU time profiling (Fedora 7 only)

6 Progress II Literature research on solving code size problem on SPE local storage: Partition Manager from University of Delaware (overlay + software cache) Dynamic code loading

7 Progress III Set up and familiarize with instrumentation tool. CIL
TAU (tau_instrumentor + Program Database Toolkit)


Download ppt "Tracing and Performance Analysis Tools for Heterogeneous Multicore System by Soon Thean Siew."

Similar presentations


Ads by Google