Heat-aware Load Balancing - Is it a Thing?

Abstract

Dynamic frequency scaling, also known by the name of its most common implementation form Intel “Turbo Boost”, has been around for over ten years. While it provides a short time boost to a CPU’s clock rate, it has no permanent influence on it. Existing work either tried to characterize the boost’s behavior or explicitly disabled the boost not to influence their performance models. We present heat-aware load balancing. This approach allows migrating a service between servers in a matter that keeps the boosted state active as long as possible. We introduce a prototype implementation that shows the feasibility of our approach in a simulated environment.