From: Jon <jon.forums@...>
Date: 2010-03-24T02:05:13+09:00
Subject: [ruby-core:28888] Re: Indentifying key MRI-on-Windows issues

> > The snippets in
> >
> > http://redmine.ruby-lang.org/issues/show/2742
> 
> The code there runs for me on Windows 7 like this:
> 
> MRI 1.8.7:              4.2 seconds
> JRuby in 1.8 mode: 3.5 seconds
> JRuby in 1.9 mode: 3.5 seconds
> MRI 1.9.2dev:         *minutes* and minutes

Thanks to nobu and Roger for the patches for 2742. On Vista, they were able to take my micro-benchmark's "real" results from 525.16 down to 18.76 seconds on a ~48MB test file.

However, it doesn't appear root cause has been identified and there is significant room for improvement as JRuby, IronRuby, and MRI 1.8.x ran in the range of 0.14 to 0.65 seconds as the following results indicate.

For testing, I used a ~48MB test file containing 625,000 LF or CRLF delimited lines containing hyphens, created by http://gist.github.com/340719

The LF and CRLF test files were micro benchmarked using http://gist.github.com/340718

The LF read results: http://gist.github.com/341367
The CRLF read results: http://gist.github.com/341365

The LF write results: http://gist.github.com/341371
The CRLF write results: http://gist.github.com/341369

As time allows, I hope to test on Linux and with Rubinius, JRuby 1.5.0, and Mono+IronRuby as well as test http://github.com/acangiano/ruby-benchmark-suite and Rails3.

Anyone done any MRI 1.9.2dev performance testing on Windows with "real world" (or micro-benchmark) heavy IO read workloads?  Anyone done any Rails3 testing?
 
Jon