Reducing cold start delay in serverless computing using lightweight virtual machines

Abstract

In recent years, serverless computing has gained considerable attention in academic, professional, and business circles. Unique features such as code development flexibility and the cost-efficient pay-as-you-go pricing model have led to predictions of widespread adoption of serverless services. Major players in the cloud computing sector, including industry giants like Amazon, Google, and Microsoft, have made significant advancements in the field of serverless services. However, cloud computing faces complex challenges, with two prominent ones being the latency caused by cold start instances and security vulnerabilities associated with container escapes. These challenges undermine the smooth execution of isolated functions, a concern amplified by technologies like Google gVisor and Kata Containers. While the integration of tools like lightweight virtual machines has alleviated concerns about container escape vulnerabilities, the primary issue remains the increased delay experienced during cold start instances in the execution of serverless functions. The purpose of this research is to propose an architecture that reduces cold start delay overhead by utilizing lightweight virtual machines within a commercial architecture, thereby achieving a setup that closely resembles real-world scenarios. This research employs supervised learning methodologies to predict function invocations by leveraging the execution patterns of other program functions. The goal is to proactively mitigate cold start scenarios by invoking the target function before actual user initiation, effectively transitioning from cold starts to warm starts. In this study, we compared our approach with two fixed and variable window strategies. Commercial platforms like Knative, OpenFaaS, and OpenWhisk typically employ a fixed 15-minute window during cold starts. In contrast to these platforms, our approach demonstrated a significant reduction in cold start incidents. Specifically, when calling a function 200 times with 5, 10, and 20 invocations within one hour, our approach achieved reductions in cold starts by 83.33%, 92.13%, and 90.90%, respectively. Compared to the variable window approach, which adjusts the window based on cold start values, our proposed approach was able to prevent 82.92%, 91.66%, and 90.56% of cold starts for the same scenario. These results highlight the effectiveness of our approach in significantly reducing cold starts, thereby enhancing the performance and responsiveness of serverless functions. Our method outperformed both fixed and variable window strategies, making it a valuable contribution to the field of serverless computing. Additionally, the implementation of pre-invocation strategies to convert cold starts into warm starts results in a substantial reduction in the execution time of functions within lightweight virtual machines.