Concurrent Dynamic Memory Coalescing on GoblinCore-64 Architecture

The majority of modern microprocessors are architected to utilize multi-level data caches as a primary optimization to reduce the latency and increase the perceived bandwidth from an application. The spatial and temporal locality provided by data caches work well in conjunction with applications that access memory in a linear fashion. However, applications that exhibit random or non-deterministic memory access patterns often induce a significant number of data cache misses, thus reducing the natural performance benefit from the data cache. In response to the performance penalties inherently present with non-deterministic applications, we have constructed a unique memory hierarchy within the GoblinCore-64 (GC64) architecture explicitly designed to exploit memory performance from irregular memory access patterns. The GC64 architecture combines a RISC-V-based core coupled with latency-hiding architectural features to a memory hierarchy with Hybrid Memory Cube (HMC) devices. In order to cope with the inherent non-determinism of applications and to exploit the packetized interface presented by the HMC device, we develop a methodology and associated implementation of a dynamic memory coalescing unit for the GC64 memory hierarchy that permits us to statistically sample memory requests from non-deterministic applications and coalesce them into the largest possible HMC payload requests. In this work, we present two parallel methodologies and associated implementations for coalescing non-deterministic memory requests into the largest potential HMC request by constructing a binary tree representation of the live memory requests from disparate cores. We present the coalesced HMC memory request results from applications that exhibit linear and non-linear memory request patterns compiled for a RISC-V core in contrast with a traditional memory hierarchy.