Hardware support for fine-grained event-driven computation in Anton 2

Exploiting parallelism to accelerate a computation typically involves dividing it into many small tasks that can be assigned to different processing elements. An efficient execution schedule for these tasks can be difficult or impossible to determine in advance, however, if there is uncertainty as to when each task's input data will be available. Ideally, each task would run in direct response to the arrival of its input data, thus allowing the computation to proceed in a fine-grained event-driven manner. Realizing this ideal is difficult in practice, and typically requires sacrificing flexibility for performance. In Anton 2, a massively parallel special-purpose supercomputer for molecular dynamics simulations, we addressed this challenge by including a hardware block, called the dispatch unit, that provides flexible and efficient support for fine-grained event-driven computation. Its novel features include a many-to-many mapping from input data to a set of synchronization counters, and the ability to prioritize tasks based on their type. To solve the additional problem of using a fixed set of synchronization counters to track input data for a potentially large number of tasks, we created a software library that allows programmers to treat Anton 2 as an idealized machine with infinitely many synchronization counters. The dispatch unit, together with this library, made it possible to simplify our molecular dynamics software by expressing it as a collection of independent tasks, and the resulting fine-grained execution schedule improved overall performance by up to 16% relative to a coarse-grained schedule for precisely the same computation.