Toward Sustainable Serverless Computing

Although serverless computing generally involves executing short-lived “functions,” the increasing migration to this paradigm requires careful consideration of energy and power requirements. is also viewed as an economically-driven computational approach, often influenced by cost computation, users are charged for per-subsecond use resources rather than coarse-grained charging that common with virtual machines containers. To ensure startup times functions do not discourage their use, resource providers need keep these hot, passing in synthetic data. We describe real consumption characteristics serverless, based on execution traces reported literature, potential strategies (some adopted from existing VM container-based approaches) can be used reduce overheads execution. Our analysis is, purposefully, biased toward machine learning workloads because: (1) increasingly being widely across different applications; (2) implement algorithms range complexity long-running (deep learning) versus short-running (inference only), enabling us consider a variety possible behaviors. The general findings easily translatable other domains.