Persistent Memory-Aware Scheduling for Serverless Workloads

In the last few years, persistent memory (PM) has become widely commercially available. PM’s direct load/store interface allows fine-grained storage access that fully bypasses all software I/O overheads. However, existing PM modules exhibit performance anomalies, particularly when they are concurrently accessed by many threads. PM is poised to change how all applications access storage, and in this paper we look at how PM might be best managed and used in serverless applications. We make the case that serverless presents a special opportunity for PM because serverless platforms can use the structure of serverless applications to gain more insight into their I/O behavior and because serverless platforms include specialized schedulers that can exploit the combined knowledge of application and PM behavior for managing concurrent workloads.