Aggregation to reduce latency

Aggregation to reduce latency

• Collect many small messages into a single large message
  • Example: sending each component separately vs. sending them together; packing cost (usually) outweighed by latency.2(s+r) vs (s + 2r + 2c)
  • Note that cache effects suggest keeping components together
• Use of collective when many copies bcast/gather

Previous slide

Next slide

Back to the first slide

View Graphic Version

I have a paper on optimizing particle methods that evaluates having components together or in separate arrays; together is better for caches