Congestion avoidance on manycore high performance computing systems
Proceedings of the 26th ACM international conference on Supercomputing
| Hi-index | 0.00 |
The paper addresses general many-to-many collective communications, whose scheduling may be needed when writing application-specific communication routines or communication libraries. Optimum schedules with the number of steps equal or close to theoretical lower bounds are designed with the use of evolutionary algorithms. Optimization is carried out for a given topology of a direct interconnection network; network nodes can be single or multiple processors connected to a router. Wormhole switching, full duplex links and single-port non-combining nodes are assumed. The developed scheduling could be advantageous mainly for networks on chip (NoC) and application-specific communication architectures.