Re: [patch,libgfortran] PR37754 [4.4 Regression] READ I/O Performance regression from 4.3 to 4.4

[Jerry DeLisle <jvdelisle at verizon dot net>]

Steve Kargl wrote:
> On Thu, Dec 25, 2008 at 09:32:57AM -0800, Jerry DeLisle wrote:
> > This patch recovers the performance from this regression by creating a 
> > stream read_char function which is simply a trimmed down version of sread 
> > (fd_read).  I was actually surprised when I saw the test results.  I 
> > suspect that the simplification allows some better optimizations.
> >
> > The patch also refactors next_char in list_read.c to eliminate goto's and 
> > inlining a small portion of the "done:" code.  The refactoring of next_char 
> > alone gains 2.8% over current trunk.  The use of the new read_char function 
> > gains significant additional performance.
> >
> > Using the countlines.f test case in the PR for comparison, average 5 runs.
> >
> > gfortran 4.3: 3.357 seconds
> >
> > gfortran 4.4 current trunk: 3.821 seconds
> >
> > gfortran 4.4 patched: 3.164 seconds
> >
> > This is a 5.7% improvement over 4.3 for this test case and 17% improvement 
> > over current trunk.
>
> Here's some numbers using /usr/bin/time on i686-*-freebsd
>
>        real   user   sys
> 4.2.5  34.33  24.91  8.82  dynamic linked
> 4.3.3  30.49  21.48  8.90  dynamic linked
> 4.4.0  29.28  20.48  8.71  static linked, w/o patch
> 4.4.0  30.51  20.83  9.32  static linked, patched
>
> This is reading a 17.2 million line file with lines ranging from
> 0 to 100 or so characters.  The numbers are the averages of 5
> consecutive runs.  The patch does appear to either help or
> hinder gfortran on FreeBSD.  I'll leave it to Janne to review
> since he's working in this area.

Steve, your tests imply we have no regression to begin with.

It appears that the results are not consistent with what I was seeing on my 
laptop.  The file I use is 8.2 mbytes with fixed length lines of about 80 
characters. I am going to do some more tests on a few more systems.

Dominiq was not seeing any significant improvement on his system.

I wonder now about the usefulness of the test case. Maybe we are just seeing 
disk latency or some similar side effect.

Jerry