This work investigates divisible load distribution using multi-installment processing on complete b -ary tree networks. Classic methods of distributing a divisible load divide the computation and communication processes into multiple time intervals in a pipelined fashion. The algorithm M ( multi-installment) herein uses multi-installment processing with pipelined communication to reduce the initial distribution time and to improve the performance. Closed-form expressions for the parallel processing time and speed-up are derived. This work reveals that the asymptotic speed-up of the proposed algorithm is bβ +1 where β is the computation-to-communication ratio of a node in the system. Algorithm M outperforms the classic algorithm in all cases. The algorithm S (start-up cost) that is developed herein includes the computation and communication start-up costs. Finally, two algorithms M and S are combined to form algorithm MS with even better performance than each.
|Number of pages||13|
|Journal||IEEE Transactions on Parallel and Distributed Systems|
|Publication status||Published - 2015 Oct 1|
All Science Journal Classification (ASJC) codes
- Signal Processing
- Hardware and Architecture
- Computational Theory and Mathematics