Optimization opportunities for various heterogeneous pipelines

Patel, Krut Sachindev

Permalink

https://hdl.handle.net/2142/129340 Copy

Description

Title

Optimization opportunities for various heterogeneous pipelines

Author(s)

Patel, Krut Sachindev

Issue Date

2025-05-08

Director of Research (if dissertation) or Advisor (if thesis)

Mendis, Charith

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Data Processing
• Machine Learning
• Heterogeneous Systems

Abstract

Data preprocessing is an important step in machine learning workloads, and has started to take up increasingly higher share of the total run times. Frameworks such as tf.data and DataJuicer have gained popularity because they provide simple abstractions to define and potentially parallelize the preprocessing pipelines. More recent works have explored further optimization opportunities, focusing on offloading the computation to separate devices, and reordering the operators to reduce data transfer costs. However, they require manual input from the users for determining the data dependencies between various operators. This thesis focuses on studying and uncovering optimization opportunities for preprocessing pipelines. Guided by detailed performance profiling, we investigate the exact conditions under which reordering can provide benefits. We also analyze the impact of jointly optimizing reordering and device placement of the operators for better performance. Additionally, we explore the possibility of using automated search of the possible optimizations by using the ML model itself as a search metric. Moreover, we investigate a data pipeline from a state of the art multimodal model and detail the novel features of its performance characteristics. Finally, we conclude with a description of the open problems that need to be tackled for building data preprocessing systems for modern machine learning workloads.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129340

Copyright and License Information

Copyright 2025 Krut Patel

Owning Collections