Performance of communication
Latency
- How long a minimum communication takes
- Latency in bytes: latency time x bandwidth this gives the capacity penalty for each operation
- Methods for minimizing the effects of latency:
- Caching reduces latency (hits)
- Prefetching hides latency by the distance between the prefetch and the use
- Concurrency tolerates latency by giving something else to do while waiting
Bandwidth
- How communication time grows with data size
- “Bisection bandwidth”
- Methods for minimizing the cost of bandwidth:
- Caching saves bandwidth (hits)
- More generally, locality saves bandwidth when cost increases with distance
- 'Combining networks' save bandwidth to a hot spot by combining operations into one