nOS: A nano-sized distributed operating system for many-core embedded systems

Simon J. Hollis, Edward Ma, Radu Marculescu

Abstract: We introduce nOS, a “nano-sized” fully distributed operating system aimed at large-scale, many-core embedded systems. nOS enables dynamic runtime optimisation of energy and execution time through lightweight and scalable distributed protocols. nOS implements new dynamic resource optimisation algorithms, and provides an intuitive and easy-to-use programmer API that supports runtime task energy optimisation through dynamic frequency scaling, transparent task communication tracking, and automatic task mapping. Critically, nOS has a completely distributed implementation, providing excellent scalability. Contrary to other approaches, the dynamic runtime optimisations require no a priori knowledge of workload or communication patterns. By generating runtime measurements of thread performance, core load, and process communication, we show that nOS can deliver improvements that would not be possible with only static analysis. Using a many-core system called Swallow, we show a <;3kB fullstack implementation of nOS together with application, OS and hardware. Using two applications with different communication patterns, we illustrate the power and flexibility of our approach, as well as various tradeoffs in energy and performance from making better mapping choices than would be available offline.