Exploiting Processor Heterogeneity in Interactive Services

To add processing power under power constraints, emerging heterogeneous processors include fast and slow cores on the same chip. This paper demonstrates that this heterogeneity is well suited to interactive data center workloads (e.g., web search, online gaming, and financial trading) by observing and exploiting two workload properties. (1) These workloads may trade response quality for responsiveness. (2) The request service demand is unknown and varies widely with both short and long requests. Subject to per-server power constraints, traditional homogeneous processors either include a few high-power fast cores that deliver high quality responses or many low-power slow cores that deliver high throughput, but not both.

This paper shows heterogeneous processors deliver both high quality and throughput by executing short requests on slow cores and long requests on fast cores with Fast Old and First (FOF), a new scheduling algorithm. FOF schedules new requests with unknown service demands on the fastest idle core and migrates requests from slower to faster cores. We simulate and implement FOF. In simulations modeling Microsoft’s Bing index search, FOF on heterogeneous processors improves response quality and increases throughput by up to 50% compared to homogeneous processors. We confirm simulation improvements with an implementation of an interactive finance server using Simultaneous Multithreading (SMT), configured as a dynamic heterogeneous processor. Both simulation and experimental results indicate processor heterogeneity offers a lot of potential for interactive workloads.