Dynamic Speedup Calculation through Self-Analysis

In an multiprocessor environment, with applications running concurrently, the scheduler is responsible for optimizing the system utilization. It distributes processors among applications according to a scheduling policy. Some policies allocate processors taking into account information such as the expected speedup. This information is usually provided by the users to the scheduler as an a priori input, and it is obtained by running the applications several times with different input sets. However, the large number of executions needed to obtain an accurate information constitute the major drawback of this approach, since they may consume a lot of time. A recent work has suggested that the efficiency of the applications can be dynamically estimated. This information can be used by the scheduler, avoiding the necessity of a priori information. The goal of our work is to present a different approach to dynamically compute the speedup achieved by parallel applications in order to provide this information to the scheduler. This approach is based on the traditional speedup equation. We will show that the dynamically calculated speedup approaches the speedup calculated as the relationship between the parallel execution with one processor and the parallel execution with P processors.