Serverless Graph Analytics on Multi-Instance GPU

Serverless computing is a promising paradigm because it provides scalability and cost-efficiency to cloud users, while reducing the burden of infrastructure management. Most mature serverless computing solutions target CPUs, with fewer solutions targeting GPUs and typically being focused on machine learning workloads. In this paper, we explore the effectiveness of serverless computing for graph analytics workloads running on GPUs, particularly multi-instance GPU. We leverage the Knative serverless computing framework and integrate Tigr as the main GPU graph analytics engine. We optimize the placement of the graph datasets used by the serverless functions, keeping many warm in the CPU main memory and a few warm in the GPU memory based on their frequency of use. We evaluate the latency and throughput of the serverless functions, showing how the best choice of MIG configuration is impacted by the choice of function, dataset, and workload heterogeneity.