Latency

• How long does it take to access data which is non-local?

  • It is a fact that addressing memory which is ‘far away’ in terms of interconnect hops or physical distance will take longer than addressing memory which is close

• Approaches for fighting the latency problem

  • Reduce the number of accesses to remote memory
    • Caches (shared memory architectures)
    • Use large messages (distributed memory architectures)
  • Use fast physical mechanisms for reducing latency
    • fast switches, fast interconnect
  • Design architecture to be ‘latency-tolerant’
    • Used in new multi-threaded architectures.

Previous slide

Back to first slide

View graphic version