Interoperable GPU Kernels as Latency Improver for MEC

Abstract

Mixed reality (MR) applications are expected to become common when 5G goes mainstream. However, the latency requirements are challenging to meet due to the resources required by video-based remoting of graphics, that is, decoding video codecs. We propose an approach towards tackling this challenge: a client-server implementation for transacting intermediate representation (IR) between a mobile UE and a MEC server instead of video codecs and this way avoiding video decoding. We demonstrate the ability to address latency bottlenecks on edge computing workloads that transact graphics. We select SPIR-V compatible GPU kernels as the intermediate representation. Our approach requires know-how in GPU architecture and GPU domain-specific languages (DSLs), but compared to video-based edge graphics, it decreases UE device delay by sevenfold. Further, we find that due to low cold-start times on both UEs and MEC servers, application migration can happen in milliseconds. We imply that graphics-based location-aware applications, such as MR, can benefit from this kind of approach.