Cost-Performance Analysis: A Comparative Study of CPU-Based Serverless and GPU-Based Training Architectures

Abstract: The field of distributed ML faces increasing demands for scalable and cost-effective training solutions, particularly in the context of large, complex models. Serverless computing has emerged as a promising paradigm to address these challenges by offering dynamic scalability and resource-efficient execution. Building upon our previous work, which introduced the Serverless Peer Integrated for Robust Training (SPIRT) architecture, this paper presents a comparative analysis of several serverless distributed ML architectures. We examine SPIRT alongside established architectures like ScatterReduce, AllReduce, and MLLess, focusing on key metrics such as training time efficiency, cost-effectiveness, communication overhead, and fault tolerance capabilities. Our findings reveal that SPIRT provides significant improvements in reducing training times and communication overhead through strategies such as parallel batch processing and in-database operations facilitated by RedisAI. However, traditional architectures exhibit scalability challenges and varying degrees of vulnerability to faults and adversarial attacks. The cost analysis underscores the long-term economic benefits of SPIRT despite its higher initial setup costs. This study not only highlights the strengths and limitations of current serverless ML architectures but also sets the stage for future research aimed at developing new models that combine the most effective features of existing systems.