Optimizing Memory Allocation in a Serverless Architecture through Function Scheduling

Abstract

In a serverless architecture, a function does not fully utilize the allocated memory. Such memory over-allocation increases node utilization and wastes resources, causing cold-start and latency issues. This paper presents a fine-grained scheduling approach for a serverless architecture that aims to address the issue of over-memory allocation and improve data locality. The proposed approach estimates how much memory each function uses so that similar functions can be scheduled on the same node. As a result, it makes less use of each node and keeps the state within a single node. We evaluated our approach through the existing FaaS applications and real-world data.