Computer 03

New Memory & Storage Hierarchies for HPC – Opportunities & Challenges (Panel)

Wednesday, June 2, 2010, 11:30am – 1:00pm, Hall 3

>> View detailed schedule

Chair

  • Dr. Frank Baetke, Global HPC Technology Program Manager, Hewlett-Packard SCI, Germany

Historically, early computers used a very simple memory and storage hierarchies which initially were all based on rotating devices and off-line storage. With the introduction of magnetic core devices for main memory and shortly afterwards the invention of the transistor more complex hierarchies evolved which even today can be used to describe advanced large-scale microprocessor-based systems. A popular approach defines four major storage levels:

  1. Storage internal to the processor, i. e. registers and often several levels of on-chip caches for data and instructions.
  2. Main memory including external caches and local memory structures. Reference to this level directly or indirectly is mostly via load and store instruction.
  3. On-line mass storages which until today is primarily based on rotating devices integrated with buffers and typically one cache level. Again, those devices can be local to a node, local to a system or be part of a global file system with devices physically far away from the site of system. Again there are huge latency variations and reference to those devices is mostly via read/write operations.
  4. Off-line bulk storage typically based on tape libraries or optical devices to ensure permanent data security.

In this session speakers will focus primarily on level 1) to 3) and address questions of coming memory technologies and related questions of locality, memory latencies and bandwidth. A special focus will be on storage level 3) because solid state disks (SSDs) are narrowing the classical gap between main memory and rotating devices and keep getting more and more competitive against classical disk devices. In the final panel discussion we will summarize the trends relevant to HPC and address the question what devices and what technologies and architectures will stay competitive at what levels in the years ahead.