Authors: Rohan Basu Roy (University of Utah), Yankai Jiang (Northeastern University), Raghavendra Kanakagiri (IIT Tirupati), and Devesh Tiwari (Northeastern University)
Abstract: This paper introduces HolyDiver, a serverless function scheduling and execution strategy that jointly optimizes the carbon footprint, water footprint, and service time of serverless functions. HolyDiver is the first system to navigate the challenging three-way trade-off among these competing objectives by leveraging spatio-temporal variations in the environmental factors that impact the carbon and water footprint of electricity generation, datacenter operation, and cooling. To solve this optimization problem, HolyDiver demonstrates a novel application of Harris Hawks Optimization (HHO) and extends HHO with dynamic hawk count adjustment, water scarcity-aware scheduling, and peer-guided performance estimation to adapt to the rapidly changing environmental conditions and function invocation patterns inherent to serverless environments. HolyDiver is evaluated on a cluster of nodes distributed across 10 regions with production-grade function execution traces and established benchmarks. We observe that HolyDiver consistently approaches the optimal solution and can complement existing schedulers to reduce the environmental footprint of serverless computing.