Efficient WebAssembly tasks orchestration on the edge

Sammanfattning

This thesis investigates methods for improving the application level orchestra- tion of WebAssembly (WASM) tasks deployed on edge devices and quantifies the resulting performance gains. We show that simple application level orchestration policies can yield measurable performance improvements on memory constrained edge devices, without modifying the OS or runtime. When an orchestrator is deployed at the edge, its efficiency directly impacts both the number of tasks a device can process and the waiting time experienced by users. However, optimizing an edge orchestrator is challenging due to strict memory constraints, which limit task concurrency as well as the range of tasks that can be executed simultaneously. WASM is an emerging execution model that addresses security concerns associated with deploying third party tasks on edge devices while also simplifying the development of edge targeted applications. In particular, WASM enables portable execution across diverse CPU and GPU architectures without requiring detailed knowledge of the underlying hardware. It also supports language interoperability, allowing tasks to be implemented in multiple programming languages. When ahead-of-time (AOT) compilation is used, WebAssembly performance approaches native execution speeds. In this thesis, several techniques aimed at improving WASM task orchestration are proposed and evaluated. Specifically, we investigate the following enhancements: Concurrent execution of one I/O-bound task and one CPU-bound task per core (Two tasks orchestrator). Prioritization of tasks with shorter execution times (Shortest tasks first orchestrator). Prioritization of tasks based on their memory footprint, where larger tasks are executed first provided that their predicted requirements fit within the available system memory (Memory Buckets orchestrator). The proposed enhancements are evaluated on a Raspberry Pi 5 using heterogeneous WASM workloads, with the Wasmtime runtime employed for executing WASM modules. The experiments are conducted under artificially constrained memory environments to reflect realistic edge deployment conditions. Performance is compared against a FIFO-based orchestrator as a baseline, using throughput, total execution time, and average task completion time as evaluation metrics. The results demonstrate moderate improvements of up to 7% in average completion time for the shortest tasks first enhancement,and improvements of up to 2% in total execution time and throughput for thememory buckets enhancement.