Estimate Load-dependent Service Demand for Modern CPU

Estimate request service demand is vital for resource management and capacity planning in large web servicing system. Most existing works assume the service demand is load-independent. This study showed that it is not the case for modern processors with Dynamic Frequency Scaling (DFS) and Simultaneous Multi-Threading (SMT) capabilities. Through experiments in a Xeon processor under Linux, this study showed that the CPU demand can be modeled as a polynomial function of CPU utilization with degree 2 or 3. This study proposed a quadratic programming based optimization method to infer the polynomial coefficients from readily available CPU utilization and response time data. Comparing with widely used regression method, proposed optimization method can reduce the error by more than half in most cases. Proposed method is also more accurate than the existing load-dependent estimation method, particularly in workload of requests with different sizes.