Locality-Driven Scheduling of Tasks for Data-Dependent Multithreading

The amount of data movement in an application has a direct impact on both its execution time and power consumption. One way to reduce this, is the implementation of locality-aware scheduling algorithms to maximize the reuse of data during the assignment of work to hardware threads. Locality-Driven Code Scheduling (LDCS), an example of such algorithms, groups the tasks that process a common data block as phases of a single coarse-grain construct named super-task, with each phase being fired according to dataflow semantics. LDCS reduces the number of long latency operations by executing all the phases of a super-task with the same hardware thread and by reading and writing the data block from and to main memory only with the first and last phases of the super-task, while the others rely on the presence of the block in the upper levels of the memory hierarchy. This paper analyzes the impact that LDCS can have on the execution time and the power consumption of an application, and presents experimental results performed on two systems: one with a software-managed memory hierarchy and another one with hardware data caches, showing that LDCS can improve the power efficiency of an application up to 72% and by 28% on average, respectively, for weak scaling.