This talk starts with a survey of 5 hardware technologies: disks, networks, memory, processors, and systems. In each case we review the history, state-of-the-art, performance models, trends, limitations and innovations. A few common themes emerge:

  • performance is characterized by a per access piece (latency) and a per byte piece (bandwidth);
  • expect continued rapid improvements in capacity, bandwidth, and cost;
  • expect very slow improvement in latency;
  • expect continued increase in intelligence of chips throughout the system, due to Moore's Law;
  • expect limitations in delivered bandwidth and latency due to inefficiencies of buses and other standard computer interfaces;
  • Unfortunately, despite these rapid advances, systems are not keeping up with users demands for decision support processing and storage.

    The last part of the talk looks at these trends and suggests an area of investigation of the future: processors, memory, and networks are distributed with the disks rather than centralized as in traditional machines. We conclude with back-of-the-envelope estimated of the cost and performance of "Intelligent Disks" (IDISKs) of the year 2000 compared to more conventional clusters of PC based machines.

    A paper on this topic is available.

    To follow the IRAM Project , see /http://iram.cs.berkeley.edu/.