Performance of Java in Function-as-a-Service Computing

One of the newest forms of serverless computing is Function-as-a-Service (FaaS). FaaS provides a framework to execute modular pieces of code in response to events (e.g., clicking a link in a web application). The FaaS platform takes care of provisioning and managing servers, allowing the developers to focus on their business logic. Additionally, all resource management is event-driven, and developers are only charged for the execution time of their functions. Despite so many apparent benefits, there are some concerns regarding the performance of FaaS. Past work has shown that cold starts typically have a negative effect on response latency (e.g., the initialization could add more than 10× execution time to short Python FaaS functions). However, the magnitude of the slowdown is subject to varying from language to language. This paper investigates how containerization and cold starts impact the performance of Java FaaS functions, and compares with the findings from the prior Python study.We find that containerization overhead slows Java FaaS functions from native execution by 4.42× on average (geometrical mean), ranging from 1.69× up to 15.43×. Comparing with Python in warm containers, Java has more overhead on three of the functions, but faster on the other functions (up to 27.08× faster). The container initialization time for Java is consistently less than half that of Python. However, Java has the additional overhead due to Java Virtual Machine (JVM) warmup which contributes varying amount of latency to the execution depending on the Java function properties. Overall, Java has about 2.60×(2.65×) speedup across seven FaaS functions over Python in cold (warm) start scenarios, respectively.