Research On Container Lifecycle-aware Task Scheduling Mechanism In Serverless Computing

Cloud computing is increasingly popular in industry in recent years,thanks to the maturity of virtualization technology and the development of container technology.It is very common to deploy applications in a container manner in cloud environments.Serverless computing is a new cloud computing paradigm derived from container technology.Containers bring advantages such as no explicit provisioning,paying as you go,and elastic scaling for serverless computing.However,this paradigm breaks down traditional monolithic applications into stateless,fine-grained functions.When busty workloads arrive,a serverless computing platform launches a large number of containers and initializes function execution environments(known as cold starts),which incurs significant startup delay.To reduce cold starts,platforms usually suspend a container after it serves a request,and reuse this container for subsequent requests.However,this reuse strategy cannot efficiently reduce cold starts because existing scheduling strategies are agnostic of container lifecycle.CAS(Container Lifecycle-Aware Scheduling)is aware of lifecycle state changes of containers when scheduling requests.The key idea is to control distribution of requests and creation or eviction of containers according to different lifecycle phases of containers.When scheduling a request,the worker with available containers is preferred.If a new container has to be created,the worker that does not require evicting soon needed containers is preferred.When creating a new container on a worker,if an available container released by a finished request appears on the worker,this container is immediately used for request processing.The prototype of CAS is implemented on the open-source platform,Open Whisk.The experimental results show that CAS reduces 72% cold starts and therefore brings a 59% reduction at 95 th percentile latency compared with a state-of-the-art strategy using consistent hashing and power of two choices,and does not bring significant performance overhead.