An approach for energy efficient co-scheduling of parallel applications on multi-core platforms

The main contribution of this work is to prove that a clever resource-aware co- scheduling policy can lead to a wide variety of benefits, such as performance speed- ups, energy efficiency improvements, reduction of thermal hotspots. To do this, a policy had been designed and implemented as an extension of the multi-objective modular scheduler YaMS, from the BarbequeRTRM resource manager. This was done because the BarbequeRTRM resource manager is portable, features a ”for free” profiling support for the applications and is characterized by energy efficiency and performance improvements with respect to the standard Linux manager. We proved that, focusing not only on which are the best tasks to schedule but also on where to schedule them, further improvements in performance and energy efficiency can be achieved with respect to the current results. We validated the policy, codename CoWS (CO-scheduling WorkloadS), on two different platforms: the first one featuring a quad-core Processor (eight cores with hyper-treading), the second a NUMA device featuring four quad-core processors with distinct cache hierarchies. The validation had been based on the comparison of energy-delay product of different workload scenarios in Linux, the resources being managed by Linux standard manager, BarbequeRTRM framework, BarbequeRTRM framework with CoWS support. In addition to the above mentioned contribution, a flow spanning from application analysis and integration with CoWs to EDP evaluation of the workload has been proposed, to guide the users through development and scheduling of his applications.