Paul McKenney, IBM Linux Technology Center

In the Communications of the ACM article entitled "Why On-Chip Cache Coherence Is Here to Stay", Martin et al. argue that on-chip cache coherence can scale for the foreseeable future, with only modest impact on performance, even for very large systems. Unfortunately, a number of the authors' arguments apply only to non-real-time systems. For example, the authors in some cases discuss amortization of overheads across operations, which addresses non-real-time scalability concerns, but fails to address the worst-case overheads that real-time workloads are sensitive to.

One possible position is that real-time response is important only on relatively small systems. Although this position might seem reasonable in light of the fact that until quite recently, real-time systems were almost always uniprocessors, recent Linux-kernel bug reports complaining of 200-microsecond latency spikes on systems with 4096 CPUs present a counterexample, albeit a surprising one. In light of this counterexample, it is reasonable to take another look at the arguments presented in "Why On-Chip Cache Coherence Is Here to Stay" with an eye to how they apply to real-time systems. This talk will take up this task, delineating what is required for highly scalable real-time systems.