Serverless computing has provided a new paradigm of cloud service that frees developers from time-intensive server management while providing finer-grained billing. However, the stateless nature of existing serverless offerings largely limits the application capability with communication to slow storage (e.g. AWS S3). Ephemeral storage addresses this issue in existing FaaS platforms, by providing the transferring of states by invoking functions asynchronously. It also adds more types of serverless compositions to embrace a larger extent of serverless applications like real-time and data-intensive pipelines.

In this project, we will design and develop a Latency-aware Ephemeral Storage Scheduler (LESS) scheduling ephemeral storage jobs to enforce desired tail latency service level objects (SLO) required by serverless applications.

We propose two novel scheduling techniques. First, SLO aware Min-Cost-Max-Flow (SAMCMF) enables the scheduling of shared storage resources with high utilization while enforcing SLO guarantees. Second, we introduce lambda-runtime-prediction, a technique to optimize storage allocation for dynamic short-lived serverless application requests. In combination, this enables LESS to improve the job placement without violating SLO requirements by dynamically allocating and deallocating disaggregated storage resources to minimize the total cost of ownership (TCO) of the storage system. Our evaluation shows that LESS can reduce tail latency violations by 12%~40% and improve the TCO by 23%~34%.