ARM has a relaxed memory model, previously specified in informal prose for ARMv7 and ARMv8. Over time, and partly due to work building formal semantics for ARM concurrency, it has become clear that some of the complexity of the model is not justified by the potential benefits. In
particular, the model was originally non-multicopy-atomic: writes could
become visible to some other threads before becoming visible to all —
but this has not been exploited in production implementations, the
corresponding potential hardware optimisations are thought to have
insufficient benefits in the ARM context, and it gives rise to subtle
complications when combined with other ARMv8 features. The ARMv8
architecture has therefore been revised: it now has a multicopy-atomic
model. It has also been simplified in other respects, including more
straightforward notions of dependency, and the architecture now includes
a formal concurrency model.

This is work presented in POPL this year. I will also present some of
the background and context on relaxed memory which is absent from the
necessarily compressed talk format of POPL.

Title: Efficient Dynamic Mapping of Parallel Applications to NUMA Architectures by Reinforcement Learning

Abstract: We present a dynamic framework for mapping threads and data of parallel applications to computational cores/memory nodes of parallel non-uniform memory architecture (NUMA) systems. We use a feedback-based mechanism where the performance of each thread is collected and used to drive the reinforcement-learning policy of assigning affinities of threads/data to CPU cores/memory nodes. The proposed framework can address different optimisation criteria, such as maximum processing speed and minimum speed variance. We demonstrate that we can achieve an improvement of 12% in execution time compared to the default Linux operating system scheduling/mapping of threads under varying availability of resources (e.g. when multiple applications are running on the same system).

Type systems are often presented in a declarative style, but with an emphasis on ensuring that there is some sort of type synthesis algorithm. Since Pierce and Turner’s “Local Type Inference” system, however, there has been a small but growing alternative: bidirectional typing, where types are synthesized for variables and elimination forms, but must always be proposed in advance for introduction forms and checked. You can still get away without any type annotations, as long as you write only normal forms. But where’s the fun in that? If you want to write terms that compute, you need to write type annotations at exactly the point where an introduction form collides with its elimination form, showing exactly the type at which computation is happening.

For type systems with some sorts of value dependency, the bidirectional approach seriously cuts down on the amount of annotation required in terms, needed only to achieve type synthesis. We have a real opportunity to reduce clutter and also to give a clearer account of the connections between types and computation.

But it doesn’t stop there. A disciplined approach to the construction of bidirectional type systems makes it easier to get their metatheory right. I’ll show this by reconstructing Martin-Löf’s 1971 type theory (the inconsistent one) in a bidirectional style and show why it has type
preservation, even without normalization.