Supporting concurrent memory access in TCF processor architectures

The Thick Control Flow (TCF) model simplifies parallel programming by bundling computations with the same control flow into single flows of variable thickness, and has the prospect of alleviating redundant usage of software and hardware resources. While architectures that can support the TCF model have been proposed, current proposals cannot support concurrent memory accesses that can both simplify programming and speed up many parallel algorithms by a logarithmic factor. In this paper, we extend current TCF architectures to efficiently support concurrent read as well as write memory accesses. The solution is based on bounded size step-caches, and exploit the two-part, hybrid, frontend-backend structure of current TCF processors, and synchronization properties of the TCF model itself. According to our simulation-based evaluation, a concurrent memory access TCF processor with B backends can execute algorithms with substantial concurrent memory accesses up to B times faster than a baseline TCF processor not supporting concurrent memory access. The hardware overhead of the solution is estimated to be modest. We include parallel program code to illustrate the gains by supporting concurrent memory accesses.