DEMOC: decompression memcpy operations for CODAG

Gacek, Andrew

Permalink

https://hdl.handle.net/2142/117685 Copy

Description

Title

DEMOC: decompression memcpy operations for CODAG

Author(s)

Gacek, Andrew

Issue Date

2022-12-08

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

Hwu, Wen-mei

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Gpu
• Codag
• Accelerator
• Compression
• Decompression

Language

eng

Abstract

Big-data applications utilizing increasingly large datasets require fast decompression schemes to process these growing stores of data. State-of-the-art big-data applications are accelerated on GPUs to leverage high compute throughput. GPU decompression frameworks such as RAPIDs and CODAG enable developers to create general purpose data pipelines to process compressed data. These frameworks provide high throughput for GPU decompression but do not leverage specific properties of the dataset when writing the output. To address this shortcoming, we introduce DEMOC, a set of writing functions for the CODAG decompression framework designed to allow developers to easily tune the decompression scheme's writing behavior. DEMOC provides seven memcpy functions tested against a variety of unit tests and within the CODAG framework using the Deflate decompression scheme. On highly compressed datasets, DEMOC functions can increase throughput by up to 1.4x, while an N-Byte Hybrid memcpy function obtains six percent increased throughput across all the test datasets and decreases the amount of memory requests by 25 percent. Other DEMOC memcpy functions are shown to achieve up to 2x throughput on large memcpy calls over the traditional 1-Byte function using the DEMOC standalone decompression writing unit test environment. To provide insight on the different memcpy functions' performance across different decompression schemes, this work goes into further detail on the resource requirements of each function and characteristics of each dataset.

Graduation Semester

2022-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/117685

Copyright and License Information

Copyright 2022 Andrew Gacek

Owning Collections