PRACE PATC Course: Node-Level Performance Engineering

 Date: Dec 3, 2013 9:00 - 17:00
Dec 4, 2013 9:00 - 17:00
Location: LRZ Building, University campus Garching, near Munich
Contents: This course teaches performance engineering approaches on the compute node level. "Performance engineering" as we define it is more than employing tools to identify hotspots and bottlenecks. It is about developing a thorough understanding of the interactions between software and hardware. This process must start at the core, socket, and node level, where the code gets executed that does the actual computational work. Once the architectural requirements of a code are understood and correlated with performance measurements, the potential benefit of optimizations can often be predicted. We introduce a "holistic" node-level performance engineering strategy, apply it to different algorithms from computational science, and also show how an awareness of the performance features of an application may lead to notable reductions in power consumption.


  • Intel and AMD x86 architectures
  • ccNUMA
  • Performance modeling & engineering approaches
  • Our Approach

Practical performance analysis

  • The LIKWID tools
  • Typical performance patterns

Microbenchmarks and the memory hierarchy                

  • Understanding the memory hierarchy
    • Data transfer between memory levels
    • Write allocate vs. NT stores
    • Modeling of cache hierarchies
    • Contention
  • NUMA effects - anisotropy and asymmetry

Typical node-level software overheads

  • Cost of synchronization
  • Work distribution

Example Problem: The 3D Jacobi solver

  • Core-level optimizations
    •  Blocking
    •  Non Temporal stores
    •  SIMD vectorization (SSE, AVX)
  • Multithreading - contention at different memory hierarchies
  • Temporal Blocking

Example Problem: The Lattice-Boltzmann Method (LBM)

  • Introduction
  • Roofline Model
  • Data layout
  • Non Temporal stores
  • Model  for in-cache data & multicore scaling
  • Sparse representation and options for propagation

Example Problem: Sparse Matrix-Vector Multiplication

  • Data layouts
  • Performance model - CPU vs. GPU
  • Bandwidth reduction

Example Problem:  A backprojection algorithm for CT reconstruction

  • The algorithm
  • Naïve analysis
  • Detailed analysis and performance model  
  • Optimizations

Energy & Parallel Scalability

  • Energy consumption of modern processors
  • The energy-to-solution metric
  • Performance engineering == power engineering
  • Case studies

Between each module, there is time for Questions and Answers!

Prerequisites Participants must have basic knowledge in programming with Fortran or C
Language: English
Teachers: Prof. Gerhard Wellen/RRZE, Dr. Georg Hager/RRZE et. al.
Registration:  via the LRZ registration form (Please choose course HNPF1W13)