Abstract
New architectures of multi- and many-core chips in large-scale supercomputers pose several challenges to designers of operating systems and runtime environments. The International Workshop on Runtime and Operating Systems for Supercomputers (ROSS) provides a forum for researchers to exchange ideas and discuss research questions that are relevant to upcoming supercomputers.
Operating systems and runtime environments on supercomputers have similar goals: both seek to provide an environment for executing applications in a scalable and high-performing way. Achieving this goal often requires minimizing the layers of indirection between the application and the architecture. Ron Brightwell argued in his 2011 keynote titled ‘Why nobody should care about operating systems for exascale’ that the operating system should be replaced with a new, specialized runtime system for the next generation of supercomputers.
This special issue presents the best papers of the ROSS’11 workshop in the critical areas of systems research on supercomputers. The contributions range from systems that emulate large-scale architectures and enable researchers to test their ideas in realistic settings to a complete, virtualized kernel for the Blue Gene/P (BG/P) supercomputer.
The first paper, by Bridges et al., describes a testbed for experimentation with operating system and runtime stacks on simulated nodes of a large-scale system. The authors show how to emulate the target system on a smaller machine using virtualization, simulation, time dilation, and slack simulation to provide an accurate estimate. Stoess et al. develop a virtual machine monitor that enables the execution of multiple virtualized applications on a BG/P system. This enables the usage of a full-featured operating system stack on the BG/P supercomputer. Jones addresses the important issue of operating system noise. The author presents a new kernel scheduling algorithm that implements co-scheduling principles for Linux; he then demonstrates impressive benefits of this approach on a large-scale Cray XT-5. Olivier et al. tackle the multi- and many-core challenge and propose a hierarchical scheduler for OpenMP tasks that matches the NUMA hierarchy of current architectures.
We anticipate that many of the principles and techniques presented in this special issue will play an important role in the design of the next successful parallel computing architecture.