Thermal-aware Scheduling in OpenMP

Abstract: Multi-core computer systems have widely been accepted as the future standard in computer architecture. Many simple processing cores bundled together on a single die have been shown more successful in terms of power consumption and execution performance compared to previous large and heavily pipelined uni-core systems. And as the chip dimensions decrease, temperature effects starts to become prominent. Elevated temperature gradients and hotspot on processors put an upper limit on both the execution performance and the life-time of the devices, leading to failures, slowdown and eventually malfunction. Most work involved in software-based temperature management in multi-core systems have been in the kernel-space, hidden from the user. This work hopes to change this, and show that it is possible to make user level schedulers account for temperature changes occurring in the system. OpenMP is the current standard in parallel programming and was used to implement a scheduling policy that uses hardware feedback to effectively try to eliminate elevated temperatures on the chip. The system itself was simulated and modeled using well-established simulators and models. The results were promising, showing a decrease in time spent above the critical temperature with up to 140 times in some benchmarks and a decreased power consumption in all the benchmarks as compared to the Cilk and Breadth-first scheduler. This alone should encourage to more research in this area, and hopefully give rise to a future standard of user level temperature control in parallel based schedulers.