Ways to Improve the Program Performance

Outline

• Compiler Optimization
• Library Optimization
• Instruction Set Optimization
• Nodes Optimization
• Thread Affinity

• Tool: Vtune

Compiler Optimization

Common Compiler Flags

• Optimize for maximum speed

• -O1
  • optimizations that increase size for a small benefit
• -O2(default)
• -O3
  • enable more aggressive optimizations that may not improve performance on some programs
• -Ofast
  • might not be safe for all programs

Library Optimization

eg. MPI & Numerical Libraries

MPI

• Message Passing Interface
• Communication protocol for parallel computers
• Implementations
  • Intel MPI
  • OpenMPI
  • MPICH

Numerical Libraries

• FFTW
  • for computing discrete Fourier transforms
• MKL
  • a library of optimized math routines 
  • linear algebra, fast fourier transforms, vector math...
• Example
  • DGMX_FFT_LIBRARY=xxx
  • Use FFTW3, MKL libraries for FFT support

Instruction Set Optimization

eg. SIMD

SIMD

• Single Instruction Multiple Data
• Same instruction is applied to many data streams
  • eg. add 64 numbers by sending 64 data streams to 64 ALUs to form 64 sums within a single clock cycle
• Instruction Sets
  • SSE
  • AVX256
  • AVX512

Nodes Optimization

The more, the better?

Scaling is limited

• Just like team work

Thread Affinity

Thread Affinity

• assign specific threads to a particular processor/core

Tool: Vtune

Vtune

Examples