Supercomputers for Starters

• Some preliminary words
• Introduction
  • Goals
  • Why supercomputing
  • What it is not
  • A compartmentalised supercomputer
  • Overview of the notes
• Processors for supercomputers
  • Introduction
  • A very basic CPU
  • Instruction level parallelism 1: Pipelining
  • Instruction level parallelism 2: Superscalar processing
  • Data parallelism through vector computing
  • Shared memory multiprocessing
  • Distributed memory computing
  • Parallelism in a modern (super)computer
  • Node architecture
  • Fast (past) evolution
  • An important difference between a PC and a supercomputer...
  • Lessons learnt
  • Further reading
• The memory hierarchy
  • The memory performance gap
  • The memory pyramid
  • AMD Rome revisited
  • AMD Milan revisited
• Storing data on supercomputers
  • Introduction
  • Problems with a parallel disk setup
  • Parallel file systems
  • A storage revolution?
  • To remember
  • Further reading
• Putting it all together
  • Scaling
  • Dennard scaling
  • Transistor cost
  • Three keywords: Streaming, Parallelism, and Hierarchy
  • Andy and Bill's law
  • Software first, not hardware
  • Building up the supercomputer
• Middleware
  • Introduction
  • Shared memory computing
  • Vector computing
  • Distributed memory supercomputing
  • Some examples
  • Further reading
• What can we expect?
  • Speed-up
  • Illustration: Matrix-matrix multiplication
  • Speed-up and efficiency of DGEMM
  • Conclusions
• Accelerators
  • What are accelerators?
  • Offloading
  • CPUs with accelerator features
  • Accelerator programming
  • Status of GPU computing
  • Further reading
• Conclusions