PoCL-R: An Open Standard Based Offloading Layer for Heterogeneous Multi-Access Edge Computing with Server Side Scalability

TL;DR

This paper introduces PoCL-R, a scalable offloading runtime for heterogeneous MEC environments that leverages open standards, minimizes latency, and significantly improves AR rendering performance and energy efficiency.

Contribution

It presents a novel open standard-based runtime for offloading compute tasks across multiple servers in MEC, handling intermittent connections and optimizing data transfer latency.

Findings

Achieves command latency of 60 microseconds in synthetic benchmarks.

Provides up to 19x improvement in AR frame rate.

Demonstrates scalable performance with 80% efficiency in fluid dynamics simulation.

Abstract

We propose a novel computing runtime that exposes remote compute devices via the cross-vendor open heterogeneous computing standard OpenCL and can execute compute tasks on the MEC cluster side across multiple servers in a scalable manner. Intermittent UE connection loss is handled gracefully even if the device's IP address changes on the way. Network-induced latency is minimized by transferring data and signaling command completions between remote devices in a peer-to-peer fashion directly to the target server with a streamlined TCP-based protocol that yields a command latency of only 60 microseconds on top of network round-trip latency in synthetic benchmarks. The runtime can utilize RDMA to speed up inter-server data transfers by an additional 60% compared to the TCP-based solution. The benefits of the proposed runtime in MEC applications are demonstrated with a smartphone-based augmented reality rendering case study. Measurements show up to 19x improvements to frame rate and 17x improvements to local energy consumption when using the proposed runtime to offload AR rendering from a smartphone. Scalability to multiple GPU servers in real-world applications is shown in a computational fluid dynamics simulation, which scales with the number of servers at roughly 80% efficiency which is comparable to an MPI port of the same simulation.