Predicting Performance and Cost of Serverless Computing Functions with SAAF

Abstract

Next generation software built for the cloud recently has embraced serverless computing platforms that use temporary infrastructure to host microservices offering building blocks for resilient, loosely coupled systems that are scalable, easy to manage, and extend. Serverless architectures enable decomposing software into independent components packaged and run using isolated containers or microVMs. This decomposition approach enables application hosting using very fine-grained cloud infrastructure enabling cost savings as deployments are billed granularly for resource use. Adoption of serverless platforms promise reduced hosting costs while achieving high availability, fault tolerance, and dynamic elasticity. These benefits are offset by pricing obfuscation, as performance variance from CPU heterogeneity, multitenancy, and provisioning variation obscure the true cost of hosting applications with serverless platforms. Where determining hosting costs for traditional VM-based application deployments simply involves accounting for the number of VMs and their uptime, predicting hosting costs for serverless applications can be far more complex. To address these challenges, we introduce the Serverless Application Analytics Framework (SAAF), a tool that allows profiling FaaS workload performance, resource utilization, and infrastructure to enable accurate performance predictions. We apply Linux CPU time accounting principles and multiple regression to estimate FaaS function runtime. We predict runtime using a series of increasingly variant compute bound workloads that execute across heterogeneous CPUs, different memory settings, and to alternate FaaS platforms evaluating our approach for 77 different scenarios. We found that the mean absolute percentage error of our runtime predictions for these scenarios was just ~3.49% resulting in an average cost error of $6.46 for 1-million FaaS function workloads averaging $150.45 in price.