- Performance Tuning and OpenMP
- Outline
- Motivation
- Performance Basics: Speed Up
- Performance Basics: Amdahl’s Law
- Consequence of Amdahl’s law: necessary parallelization
- Performance measurement: Profiling
- General issues: Problem size dependency of performance
- OpenMP version of Matrix Norm
- Performance on a PC (Dual Pentium II, 450 MHz)
- Performance on a Vectorcomputer (NEC SX-5Be)
- Performance on the Hitachi SR8000
- Performance comparison (2 threads with OpenMP)
- Use OpenMP only with sufficient workload: if-clause
- Performance with if-clause
- Performance with OpenMP: Avoid thread creation
- Performance with OpenMP: Avoid barriers
- Performance with OpenMP: Avoid barriers (II)
- Performance with OpenMP: load balancing
- Summary
- Exercise: Matrix Norm Calculation
- Exercise: Matrix Norm Calculation on SR8000
- Exercise: Matrix Norm Calculation on SR8000
- Exercise: Matrix Norm Calculation on SR8000
- Exercise: Matrix Norm Calculation
- Gnuplot
|