Performance Enhancement of Serverless System

Abstract

Serverless computing has emerged as a transformative paradigm in cloud computing, offering scalability and cost efficiency by abstracting infrastructure management. This thesis presents a holistic investigation into serverless computing, focusing on three key contributions: performance and cost modeling, performance and cost optimization, and event-driven state management. In the first contribution, it addresses the limitations of existing performance and cost models by introducing an analytical model that accommodates complex application structures such as loops, self-loops, parallel paths, and cycles. This model, applied to serverless workflows, provides an accurate estimation of end-to-end response time and cost. Traditional resource allocation models are inadequate for serverless platforms, which charge based on actual resource consumption. To bridge this gap, it develops an efficient predictive cost model that considers parameters such as execution time, memory usage, and function invocation patterns. Optimizing performance and cost in serverless computing is essential for achieving an optimal balance between resource utilization and economic viability. A greedy-based optimization algorithm is proposed to determine the optimal memory configuration, achieving the best response time within budget constraints. This algorithm explores trade-offs between performance optimization and cost minimization, considering the dynamic nature of serverless environments. The cold-start problem, a significant challenge in serverless computing, is addressed in the second contribution. An integrated adaptive model is introduced that leverages the LSTM-based deep learning approach to predict future workloads, minimizing both the frequency and delay of cold starts. The proposed efficient container placement module accelerates container delivery, and a high-performance serverless containerization prototype optimizes cold-start delays by grouping similar functions. The third contribution focuses on state management, a critical aspect of serverless computing. An event-driven state management strategy is proposed that aligns state changes with triggering events, ensuring a scalable and natural approach to managing state in serverless architectures. This approach enhances reliability and efficiency, addressing the challenges of maintaining state in the transient and dynamic nature of serverless environments. This thesis advances the understanding of serverless computing through a comprehensive study of performance modeling, optimization, and state management. The proposed models and strategies contribute to overcoming existing research gaps, providing valuable insights for practitioners and researchers. The findings underscore the significance of balancing performance and cost while addressing challenges such as the cold-start problem and state management in serverless computing.