ceClub: A Look at Multiprocessor Real-Time Scheduling in Linux: Generality, Feasibility, and Scalability

Speaker:
Bjorn Brandenburg (Max Planck Institute for Software Systems in Kaiserslautern, Germany)
Date:
Monday, 19.5.2014, 11:30
Place:
Room 337 Taub Bld.

Given the popularity of Linux and the increasing adoption of embedded multicore platforms, Linux (and Linux-like RTOSs such as LynxOS and QNX) are increasingly used to host real-time workloads on multiprocessors. This naturally exposes new limitations, both in the actual implementation and in the supporting analytical foundations, and has thus been a driving force for a number of recent advances in multiprocessor real-time scheduling.

In this talk, I'm going to present two such results: first, I'm going to discuss how Linux's notion of processor affinities turned out to be strictly more general than many of the models typically studied in the real-time literature. And second, I'll discuss the (lack of) scalability of the Linux scheduler on large multicore platforms in terms of average- and worst-case overheads, and present an alternative, much more scalable design based on message passing.

Bio:
Bjorn Brandenburg is a tenure-track research group leader at the Max Planck Institute for Software Systems (MPI-SWS) in Kaiserslautern, Germany. He joined MPI-SWS in 2011 after obtaining his PhD from the University of North Carolina at Chapel Hill, where he worked with Jim Anderson. His research is focused on operating systems for predictable multiprocessor real-time systems and spans from the analytical foundations on the one hand to practical systems and implementation issues on the other hand, with a special focus on real-time locking and scheduling. He is the lead developer and maintainer of LITMUS RT (http://www.litmus-rt.org), a real-time extension of the Linux kernel that has been in development since 2006.

Back to the index of events