How to do nothing in less time

Speaker: David Haensel (JSC)
Date: Friday, 4 December 2015, 08:30-10:00
Session: Numerics II
Talk type: Project talk (30 min)

Abstract: In today's MD simulations the scaling bottleneck is shifted more and more from computation towards communication. Especially the calculation of long range interactions in O(n) amplifies this imbalance even more. The Fast Multipole Method used in this project is one algorithm for calculating long range interactions with linear complexity. How can we reduce the total run-time further? To reduce the latency and bandwidth costs along the critical communication path we must change the communication pattern. This leads to additional replication of computation which eventually can be overlapped with communication. To reduce the visibility of such communication inside the algorithm we introduced several abstraction layers. This allows an easy exchange of underlying communication techniques without changing code at the algorithmic level. Additionally, this enables us to implement different parallelization schemes on top of this infrastructure. In this talk we show the advantages of the implemented communication concept as well as the straightforward application thereof. Finally we present the benchmark results of a latency avoiding communication scheme.