Hybrid of Job Sequencing and DVFS for Peak Temperature Reduction with Nondeterministic Applications

In this paper we propose a hybrid approach, which improves system thermal behavior with nondeterministic applications based on job sequencing as well as dynamic voltage and frequency selection (DVFS). The hybrid approach consists of two steps: 1) job sequencing and 2) lowering processor´s speed by using dynamic voltage and frequency selection. The first step constructs a job sequence based on the average execution time of each job; and the constructed job sequence interleaves executions of hot jobs and cool jobs. While a hot job pushes up the temperature of a chip, a cool job cools the chip down. Accordingly the job sequencing step reduces the peak temperature of a chip. The second step further improves system-wide thermal behavior by dynamic voltage and frequency selection, which slows down the execution of hot jobs by utilizing job slacks. Hot jobs are the rooted cause for deteriorating the thermal behavior of a processor, so we only slow down the execution of hot jobs and job slacks are allocated among hot jobs. Hot jobs whose execution time deviates more above their average execution time get more slack; vice versa. Experimental results show that the proposed hybrid approach improves the system thermal behavior for various nondeterministic workloads by reducing peak temperature as well as decreasing thermal variations.