A systematic evaluation of machine learning on serverless infrastructure

Recently, the serverless paradigm of computing has inspired research on its applicability to data-intensive tasks such as ETL, database query processing, and machine learning (ML) model training. Recent efforts have proposed multiple systems for training large-scale ML models in a distributed manner on top of serverless infrastructures (e.g., AWS Lambda). Yet, there is so far no consensus on the design space for such systems when compared with systems built on top of classical "serverful" infrastructures. Indeed, a variety of factors could impact the performance of training ML models in a distributed environment, such as the optimization algorithm used and the synchronization protocol followed by parallel executors, which must be carefully considered when designing serverless ML systems. To clarify contradictory observations from previous work, in this paper we present a systematic comparative study of serverless and serverful systems for distributed ML training. We present a design space that covers design choices made by previous systems on aspects such as optimization algorithms and synchronization protocols. We then implement a platform, LambdaML , that enables a fair comparison between serverless and serverful systems by navigating the aforementioned design space. We further improve LambdaML toward automatic support by designing a hyper-parameter tuning framework that leverages the ability of serverless infrastructure. We present empirical evaluation results using LambdaML on both single training jobs and multi-tenant workloads. Our results reveal that there is no "one size fits all" serverless solution given the current state of the art-one must choose different designs for different ML workloads. We also develop an analytic model based on the empirical observations to capture the cost/performance tradeoffs that one has to consider when deciding between serverless and serverful designs for distributed ML training.