A Serverless-Optimized Garbage Collector.

Abstract

Managed programming languages abstract away low-level memory management, enabling programmers to focus on high-level concerns. In these programming languages, Automatic Memory Management (AMM), often realized through a Garbage Collector (GC), automat ically handles memory allocation and deallocation. While AMM mitigates memory-related errors, its overhead can impact application performance. Modern GCs employ techniques such as concurrency, generational memory management, and adaptive algorithms to mini mize this performance impact. This dissertation focus on GC strategies specifically tailored for Function-as-a-Service (FaaS) applications. FaaS workloads exhibit a distinct memory usage pattern, characterized by ephemeral and persistent objects. Unlike traditional, general purpose GC approaches, in this work we propose and evaluate a novel GC algorithm, Serverless-Optimized Garbage Collector (SOGC). SOGC takes advantage of FaaS unique characteristics to achieve signifi cant efficiency gains. A typical memory usage cycle in a FaaS function involves an initialization phase, dur ing which data intended for the function’s entire lifetime is allocated, followed by an event handling phase, characterized by ephemeral data that is used to process the event and then promptly discarded. SOGC addresses this pattern by organizing memory into a layout that includes a persistent space for long-lived data and a separate handler space for each event. This efficient memory organization allows for rapid reclamation of unused data, minimizing garbage collector-related interruptions during the execution of business logic. To evaluate SOGC, we employ an analytical model, enabling a direct comparison with classic GC algorithms. Through this model, we assess various scenarios, demonstrating that SOGChas the potential to outperform existing solutions under certain conditions.