Re: [RFC] [-mm] Remove 'unsafe' LZO decompressor

["Nitin Gupta" <nitingupta910@gmail.com>]

Hi Richard,

Thanks for these perf. figures!

On 5/25/07, Richard Purdie <richard@openedhand.com> wrote:
> On Thu, 2007-05-24 at 11:50 -0700, Andrew Morton wrote:
> > On Thu, 24 May 2007 18:15:17 +0100
> > Michael-Luke Jones <mlj28@cam.ac.uk> wrote:
> >
> > > Attached is a patch which may be desirable for -mm. It applies
> > > directly to 2.6.22-rc2-mm1.
> > >
> > > The patch removes the 'unsafe' LZO decompression function, lowering
> > > the size of the minilzo.c file by nearly 500 out of an original 1727
> > > lines. It also removes references to the 'unsafe' decompression
> > > function in the public LZO header and the EXPORT_SYMBOL_GPL declaration.
> [...]
> > > Comments / disagreement all welcome :)
> >
> > This is obviously a highly desirable thing to do for a number of reasons.
> > But have we quantified the performance difference?
>
> Ok, I've done some tests:
>
> 1. Comparing the safe and unsafe functions
>
> For my minilzo kernel patch, the safe version showed a 7.2% performance
> hit. For Nitin's patch, it showed a 3.2% performance hit (but see
> below).

> Could be a lot worse and I don't object to the removal of the unsafe
> version.

Ok. Let's drop unsafe version. This will also do away will that Makefile debris.

> 2. Comparing Nitin's code with my minilzo based kernel patch.
>
> My kernel patch is about 2.25 times faster at decompression (16725Kb/ms
> vs 7399Kb/ms) and fractionally faster at compression (1434kb/ms vs
> 1490kb/ms). As things stand the performance of Nitin's patch is
> therefore unacceptable as that is a significant decompression
> performance loss.

I suspect this is mainly due to replacement of COPY4() and open coded
byte-by-byte copying with memcpy() calls. I will rollback these
particular changes, remove unsafe version and will see again. I have
retained all other code as-is.

I will also try adding benchmarking code to my (GPL) 'compress-test'
module (same which I sent to Bret). Though it's pretty basic module
but should be sufficient to give required perf. figures.

> These tests are on 32 bit Intel and in userspace but I've found them to
> be a pretty good indicator of what happens in the real world and on
> other architectures.

> For simplicity I made these tests with some existing code I had around
> but its licence is such I can't share it, much to my frustration.


Cheers,
Nitin
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/