Energy Discounted Computing on Multicore Smartphones

Multicore processors are not energy proportional: the first running CPU core that activates shared resources incurs much higher power cost than each additional core does. On the other hand, typical smartphone applications exhibit little parallelism and therefore when one core is activated by an interactive application, computing resources at other cores are available at a deep energy discount. By non-work-conserving scheduling, we exploit energy-discounted co-run opportunities to process best-effort smartphone tasks that involve no direct user interaction (e.g., data compression / encryption for cloud backup, background sensing, and offline bytecode compilation). We show that, for optimal co-run energy discount, the best-effort processing must not elevate the overall system power state (specifically, no reduction of the multicore CPU idle state, no increase of the core frequency, and no impact on the system suspension period). In addition, we use available ARM performance counters to identify co-run resource contention on the multicore processor and throttle best-effort task when it interferes with interactivity. Experimental results on a multicore smartphone show that we can reach up to 63% energy discount in the best-effort task processing with little performance impact on the interactive applications.