[ORC-639] Improve zstd compression performance

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 1.6.3, 1.7.0
• Fix Version/s: 1.6.4, 1.7.0
• Component/s: C++, compression
• Labels: performance
• Flags: Patch
• External issue URL: https://github.com/apache/orc/pull/511
• Assignee: Ion Gaztañaga
• Reporter: Ion Gaztañaga
• Created: 06/Jun/20 21:56
• Updated: 15/Sep/20 00:01
• Resolved: 11/Aug/20 04:54

Description

In Compression.cc, classes ZSTDCompressionStream and ZSTDDecompressionStream employ Zstd's "Simple API", which initializes the compression/decompression context each time. This context initialization includes time to initialize all compression/decompression tables.

 

Using the "Explicit context" API improves the performance significantly as compression/decompression tables are constructed once and reused each time "decompress" or "doBlockCompression" is called.

We've noticed more than 15% time improvement in some applications with this change, so it seems that any ORC user using Zstd can benefit from this change.

Issue Links

• links to: GitHub Pull Request #511
• links to: GitHub Pull Request #537