StreamBox - A Lightweight GPU Sandbox for Serverless Inference Workflow

Explore a groundbreaking conference talk on StreamBox, a lightweight GPU sandbox designed for serverless inference workflows. Delve into the challenges of dynamic workloads and latency-sensitive DNN inference in serverless computing environments. Discover how StreamBox addresses the limitations of existing serverless inference systems by implementing fine-grained and auto-scaling memory management, enabling transparent and efficient intra-GPU communication across functions, and facilitating PCIe bandwidth sharing among concurrent streams. Learn about the significant improvements StreamBox offers, including up to 82% reduction in GPU memory footprint and a 6.7X increase in throughput compared to state-of-the-art systems. Gain insights into the potential impact of this innovative approach on scalable DNN inference serving and the future of serverless computing for GPU-intensive tasks.