System Load Characterization Using Low-Level Performance Measurements

The performance of a software system directly influences customer satisfaction. Self-adaptiveness can contribute to this customer satisfaction by (1) taking appropriate measures when the performance becomes critical, e.g., the system load is too high, or (2) scheduling intensive tasks when the load is low. We investigate how self-adaptive systems can use low-level system measurements to characterize the load on a system. Our approach uses a combination of statistics and association rule learning to perform the characterization. We evaluate our approach using two case studies: a large-scale industrial system and a widely used synthetic benchmark (RUBiS). From our case studies follows that our approach is capable of closely characterizing the load on a system and that it is successful in detecting performance anomalies as well.