[HADOOP-1926] Design/implement a set of compression benchmarks for the map-reduce framework

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.15.0
• Assignee: Arun Murthy
• Reporter: Arun Murthy
• Created: 20/Sep/07 11:51
• Updated: 08/Jul/09 16:52
• Resolved: 02/Oct/07 22:16

Description

It would be nice to benchmark various compression codecs for use in the hadoop (existing codecs like zlib, lzo and in-future bzip2 etc.) and run these along with our nightlies or weeklies.

Here are some steps:
a) Fix HADOOP-1851 ( Map output compression codec cannot be set independently of job output compression codec)
b) Implement a random-text-writer along the lines of examples/randomwriter to generate large amounts of synthetic textual data for use in sort. One way to do this is to pick a word randomly from /usr/share/dict/words till we get enough bytes per map. To be safe, we could store an array of Strings of a snap-shot of the words in examples/RandomTextWriter.java.
c) Take a dump of wikipedia (http://download.wikimedia.org/enwiki/) and/or the ebooks from Project Gutenberg (http://www.gutenberg.org/MIRRORS.ALL) and use them as non-synthetic data to run sort/wordcount against.

For both b) and c) we should setup nightly/weekly benchmark runs with different codecs for reduce-outputs and map-outputs (shuffle) and track each.

Thoughts?

Issue Links

• is blocked by: HADOOP-1851 Map output compression codec cannot be set independently of job output compression codec