A low noise unikernel for extrem-scale systems

Abstract

We expect that the size and the complexity of future supercomputers will increase on their path to exascale systems and beyond. Therefore, system software has to adapt to the complexity of these systems to simplify the development of scalable applications. In cloud environments, the activity of a virtual machine on a neighboring core may decrease performance due to issues such as cache contamination (noise neighbor problem). In this paper, we present the unikernel operating system HermitCore coming up with predictable runtimes, which improves the scalability. It extends the multi-kernel approach with unikernel features while providing better programmability and scalability for hierarchical systems. In addition, the same binary can be used to run as unikernel within virtual machines. By using a unikernel, the memory footprint of Virtual Machines (VMs) is decreased, which reduces the pressure on the cache system and improves the overall performance. We prove the predictable runtime of the design via micro benchmarks by taking the example of HermitCore on the upcoming manycore architecture Knights Landing.