Serverless Data Analytics

Identifier:  TFM:2452

Handle: https://hdl.handle.net/20.500.11797/TFM2452

Authors:  Predoaia, Ionut

Abstract:
Serverless computing has shown vast potential for data analytics applications, especially for embarrassingly parallel workloads. Nevertheless, little consideration has been given in the literature to porting stateful applications requiring shared state to serverless. This work closes this gap by exploring hardware resource disaggregation in serverless computing, with the aim of porting to serverless stateful machine learning algorithms, i.e., k-means clustering and logistic regression. A set of guidelines, challenges and limitations of porting stateful machine learning algorithms to serverless are presented in this work. Running stateful applications on serverless architectures inherently induces overheads, as serverless functions are not directly network-addressable, hence one must rely on a remote storage service for storing the shared state. In this work, the feasibility of solving stateful machine learning algorithms is evaluated, and furthermore, optimization techniques are proposed to enhance their feasibility. The performance, scalability and overheads of the stateful machine learning algorithms running with serverless architectures have been evaluated. The serverless implementation of the k-means algorithm has achieved an 87-fold speedup compared to a sequential implementation of the algorithm. In terms of scalability, the serverless implementation has achieved a scale-up factor of 0.91 with 100 concurrent serverless functions. To raise the feasibility of the serverless implementation, intra-function parallelism has been employed as an optimization technique to parallelize the execution of the serverless functions, achieving up to 68% improved performances.