Loading...
Optimal ATAPE task scheduling on reconfigurable and partitionable hierarchical hypercube networks
Author(s)
Phisutthangkoon, Nuntipat
Werapun, Jeeraporn
Date Issued
July 1, 2022
Type
Article
Abstract
The hierarchical hypercube (HHC) network (N=2n, n=2m+m, m≥2) is one of scalable networks for the large-scale parallel-computing systems since its implementation cost is lower than the popular hypercube (HC) network. However, the complicated HHC-routing may conflict (due to the reduction-of-edges in the hierarchical construction) and cannot be reconfigured simply for k > 2m nodes-per-task. Thus, a crucial research to fulfill the HHC-network is the conflict-free routing since the conflict is not handled easily during runtime. Recently, the HHC-conflict was solved by the shortest-path routing and the special HHC-partitioning, called the grouping-of-cross dual-cube (GCD) partitioning, for k=2m+1 nodes-per-task. In this study, we propose the reconfigurable-and-partitionable HHC (RP-HHC) network for 2m+2 ≤ k ≤ N/2M-1 nodes-per-task (M=2m-1) by the grouping-of-cross sub-system (GCS) combining (the optimal reconfiguring). Moreover, we present the efficient task-scheduling and processor-allocation, incorporated with the reconfigured binary-tree. Since the first-fit policy is fast but returns low-performance while the best-fit policy yields high-performance but is time-consuming, we propose the efficient best-fit driven-policy to work as fast as the first-fit policy while maintaining the high-performance. The correctness of new reconfiguring and time-complexity of task-allocation were analyzed. In experiments, all-to-all personalized exchange (ATAPE) communication was assessed to confirm no-conflict on the RP-HHC systems.
Citation
Parallel Computing, 111, 2022