Enabling GPU Acceleration for the Dandelion Serverless Executor

Abstract

With the slowing of Moore's law, modern computing platforms are becoming increasingly heterogeneous. In particular, graphics processing units (GPUs) have found application in various general-purpose fields, for instance driving the machine learning revolution of recent years. Serverless cloud computing has, however, had trouble adjusting to this reality. Most systems in use today are still largely restricted to CPU-only execution. At its core, this work is about extending Dandelion, a novel function-as-a-service (FaaS) platform, with the means to execute untrusted user code on GPUs. By making use of Dandelion's innovative programming model, we show that efficient and performant GPU-accelerated serverless computing is possible and can lead to significant benefits in key workloads.