Energy-Efficient Scheduling of Interactive Services on Heterogeneous Multicore Processors

A heterogeneous multicore processor has several cores that share the same instruction set architecture but run at different speeds and power consumption rates, offering both energy efficient cores and high-performance cores to applications. We show how to exploit such processors to make significant energy reduction to serve large interactive workloads such as web search by carefully scheduling requests. Scheduling is a challenging task. Intuitively, we want to run short requests on slow cores for energy efficiency and long requests on fast cores for timely responses. However, there are two key challenges: (1) request service demands are unknown; and (2) the most appropriate core to run a request may be busy.

We propose an online algorithm, Fast-Preempt-Slow (FPS), which improves response quality subject to deadline and total power constraints. We conduct a simulation study using measured workload from a large commercial web search engine as well as using a variety of synthetic workloads to assess the benefits of FPS. Our results show significant benefits, achievable under a wide variety of conditions: The throughput of a heterogeneous processor is 60% higher than that of the corresponding homogeneous processor with the same power budget; equivalently, to support a large workload as in Web search, FPS on the heterogeneous processors reduces the number of servers by approximately 40%.