Scheduling Containers Rather Than Functions for Function-as-a-Service
Dong Kyoung Kim, Hyun-Gul Roh · 2021
Function-as-a-Service (FaaS) is a compelling technology that allows users to run functions in an event-driven way without concerns about server management. Container-based virtualization enables functions to run in a lightweight and isolated run-time environment, but frequent function executions accompanied with container initialization (cold starts) make the platform busy and unresponsive. For performance sake, warm starts, which is to execute functions on already initialized containers, are encouraged, and thus FaaS platforms make efforts to schedule functions to warm containers.From our experience operating an on-premise FaaS platform, we found that the existing scheduler showed poor performance and unstable behavior against multi-tenant and highly concurrent workloads. This paper proposes a novel FaaS scheduling algorithm, named FPCSch, that schedules Function-Pulling-Containers instead of scheduling functions to containers. As FPCSch lets containers continuously pull functions of the same type, cold starts decrease dramatically. Our evaluations show that Apache OpenWhisk equipped with FPCSch has many desirable features for FaaS platforms; (1) quite stable throughput against the multi-tenant workloads mixed by the increasing numbers of function types, (2) growing throughput for increasing concurrency, (3) uniformly load-balancing resource-intensive workloads, and (4) nearly proportional performance for scale-out.