Data Scheduling on Processor-In-Memory Arrays Based on Data Placement and Data Movement

In the study of PetaFlop project, Processor-In-Memory array was proposed to be a target architecture in achieving 1015 oating point operations per second computing performance. However, one of the major obstacles to achieve the fast computing was interprocessor communications, which lengthen the total execution time of an application. A good data scheduling, consisting of nding initial data placement and data movement during the run-time, can give a signi cant reduction in the total communication cost and the execution time of the application. In this paper, we propose e cient algorithms for the data scheduling problem. Experimental results show the e ectiveness of the proposed approaches. Compared with default data distribution methods such as row-wise or column-wise distributions, the average improvement for the tested benchmarks can be up to 30%. This work was partially supported by NSF MIP 95-01006 and NSF ACS 96-12028. 1