Investigation of Communication Overhead of SoC Lookaside Accelerators

Lookaside accelerators (LAAs) enable opportunities to asynchronously offload computationally intensive tasks from the general-purpose processors to the accelerator in order to improve the overall performance of a given system. However, the communication overhead involved in moving data between the LAA and the CPU can be significant and can impact anticipated performance, diminishing the benefits of a LAA. This paper examines the communication overhead between LAAs and CPUs, focusing on the communication setup and data transmission components. We develop a mathematical model to analyse and capture the problem and investigate the conditions under which hardware accelerators provide effective performance. To verify and validate the proposed model, we carried out studies using an ARM-based platform. Experimental results reveal that the length of data send to the LAA significantly impacts the overall offload benefit of the LAA when compared with software-only solutions. Communication overhead, specifically communication setup latency, has a diminishing impact on smaller data lengths. The presented mathematical model and experimental results clearly demonstrate that the targeted data length for offload is a critical parameter when designing LAAs.