Biblio

Energy efficient High-Performance Computing (HPC) is becoming increasingly important. Recent ventures into this space have introduced an unlikely candidate to achieve exascale scientific computing hardware with a small energy footprint. ARM processors and embedded GPU accelerators originally developed for energy efficiency in mobile devices, where battery life is critical, are being repurposed and deployed in the next generation of supercomputers. Unfortunately, the performance of executing scientific workloads on many of these devices is largely unknown, yet the bulk of computation required in high-performance supercomputers is scientific. We present an analysis of one such scientific code, in the form of Gaussian Elimination, and evaluate both execution time and energy used on a range of embedded accelerator SoCs. These include three ARM CPUs and two mobile GPUs. Understanding how these low power devices perform on scientific workloads will be critical in the selection of appropriate hardware for these supercomputers, for how can we estimate the performance of tens of thousands of these chips if the performance of one is largely unknown?

In the last decade, the number of available cores increased and heterogeneity grew. In this work, we ask the question whether the design of the current operating systems (OSes) is still appropriate if these trends continue and lead to abundantly available but heterogeneous cores, or whether it forces a fundamental rethinking of how systems are designed. We argue that: 1. hiding heterogeneity behind a common hardware interface unifies, to a large extent, the control and coordination of cores and accelerators in the OS, 2. isolating at the network-on-chip rather than with processor features (like privileged mode, memory management unit, ...), allows running untrusted code on arbitrary cores, and 3. providing OS services via protocols over the network-on-chip, instead of via system calls, makes them accessible to arbitrary types of cores as well. In summary, this turns accelerators into first-class citizens and enables a single and convenient programming environment for all cores without the need to trust any application. In this paper, we introduce network-on-chip-level isolation, present the design of our microkernel-based OS, M3, and the common hardware interface, and evaluate the performance of our prototype in comparison to Linux. A bit surprising, without using accelerators, M3 outperforms Linux in some application-level benchmarks by more than a factor of five.