Previous | Next --- Slide 38 of 40
Back to Lecture Thumbnails

How have supercomputers responded to the power wall? Several factors off the top of my head:

  • It seems like supercomputers can vary power among individual units with much higher variance because of custom cooling. Could this be used to dynamically alter clock frequency to combat dynamic load imbalance?
  • Communication latency and bandwidth are huge problems for supercomputer workloads and also huge power hogs. What are the tradeoffs between latency, bandwidth, and power for interconnects?
  • Are there quantitative ways to schedule supercomputing programs to minimize power usage?
  • Does VLIW make sense in a supercomputing context? If so and not in use, why not? There are definite power advantages to having simpler hardware.

Somewhat off topic, but as supercomputers increase in number of nodes, and failures become more common, has the SC community started to adopt distributed systems techniques (specifically fault-tolerance)?