Suche


Architectures of Supercomputers

Modulbeschreibung

In this lecture you will get to know the internal workings of the world's fastest supercomputers. In the last years, supercomputers have become an integral tool of research in virtually all fields of natural sciences. They enable scientists to carry out simulations with unprecedented accuracy and speed, thereby reducing the need for costly experiments. We will discuss the design principles used to build modern day processors and networks as well as how these are applied in recent high performance computing systems.

The exercises are an opportunity to apply the lecture's topics on real HPC systems in our parallel computing laboratory (e.g. IBM BladeCenters, NVIDIA Tesla C2050 GPGPUs, AMD Magny-Cours, Intel Nehalem-EX). Topics range from writing assembly and SSE code to measuring and exploiting cache effects on multi-cores CPUs.

This lecture is a good combination for the HPC computing group's lecture Programming Techniques for Supercomputers (PTfS).

Lecture Topics:

  • basic computer and processor architecture
  • advanced processor design paradigms
    (e.g. pipelining, super-scalar designs, out-of-order execution)
  • accelerators
    (e.g. gpgpus, but also cell, scc)
  • performance models to predict application speedup on parallel systems
  • supercomputers detailed
    • recent Top 500 #1 systems
      (e.g. Earth Simulator, IBM Roadrunner, IBM Blue Gene, Tianhe-1A)
    • local systems
      (e.g. RRZE's lima, SuperMUC)
    • upcoming systems
      (e.g. IBM Blue Waters)

Exercises Topics:

  • assembly programming
  • loop vectorization via SSE
  • cache blocking
  • multi-threading via OpenMP
  • performance measurement and visualization

Lecturers

Prof. Dr.-Ing. Dietmar Fey, Dipl.-Inf. Andreas Schäfer

Details

Lecture+Exercises, 2+2 SWS, Schein, ECTS-Studium, ECTS-Credits: 5.0
nur Fachstudium

Studienfächer / Studienrichtungen:

WPF CE-MA 1-3
WPF CE-MA-INF 1-3
WPF CE-BA-TW 3-6
WPF INF-DH-RA 5-9
WPF INF-BA-V-RA ab 5

Hazards Nehalem Microarch Memory Bandwidth Roofline Cell Chip NVIDIA Fermi Blue Waters Drawer Jaguar Installation Jaguar Cooling