Parallel Computing and Dynamic Load Balancing of ADPAC On a Heterogeneous Cluster of Unix and NT Operating Systems

Abstract

This chapter involves parallel processing of an advanced ducted propfan analysis code(ADPAC) on local area and wide area networks formed by combined Unix and Windows/NT environments. The results obtained from each workstation are compared and verified by using a computational fluid dynamics (CFD) test case. Timing evaluation and dynamic load balancing studies are conducted. From the timing results and speedups obtained, it is evident that parallel computations are less satisfactory than using either Pentium/NT or RS6K machines individually as computational platform. Although, in some cases, the NT cluster has achieved better speedup, its computational speed is about two times slower than the RS6K cluster. On the other hand, the RS6K workstation is less efficient compared to the NT workstation for increased number of blocks, mainly due to the inability of the network speed to keep up with the faster processing speed of RS6K. In heterogeneous workstation, the communication speed may also vary due to the difference in communication bandwidth for different computer architecture. As for the dynamic load balancing (DLB) test runs, the load balancing leads to an improvement in total elapsed time of ADPAC computations. In the local area network (LAN) case, balanced block distribution reflects performance and computational power of computers involved in this run. An improvement of 25% in elapsed time for the ADPAC computation was observed. In the case of the wide area network (WAN) environment, similar improvement in the ADPAC computation times was observed.