Re: [RFC] LZO de/compression support - take 4

[Daniel Hazelton <dhazelton@enter.net>]

On Monday 28 May 2007 04:37:04 Nitin Gupta wrote:
> On 5/28/07, Daniel Hazelton <dhazelton@enter.net> wrote:
> > Test code for this version (take 4) of the minimized LZO1X (from the
> > liblzo v2) is complete.
> >
> >
> > I don't see a significant slow-down comparing the complete liblzo2 to
> > this minimized code on my system (Pentium M 1.73GHz, 1GB Ram, Kubuntu
> > Feisty (stock Kubuntu kernel)). Rather, I see the opposite. This *might*
> > have been caused by the dynamic linking (or similar) so rather than rely
> > on simply doing "time xxx" I actually put checks around the calls to the
> > compress/decompress functions themselves.
> >
> > ('Tiny LZO' is what I call Nitins extremely small implementation of
> > lzo1x_[de]compress)
> >
> > Output of the provided "test" script:
> > 10 run averages:
> > 'Tiny LZO':
> >         Combined: 113.2 usec
> >         Compression: 77.4 usec
> >         Decompression: 35.8 usec
> > 'liblzo2':
> >         Combined: 140.7 usec
> >         Compression: 94 usec
> >         Decompression: 46.7 usec
> >
> > (The "Combined" average is the average time taken for a
> > compress+decompress)
> >
> > TODO:
> > -Implement userspace version of likely/unlikely
> > -Implement cpu_to_le16 so code functions on BE systems
> >
> > DRH
>
> As you mentioned in your mail, you are using lzo1x_1_11_compress()
> which is slower than what I ported (which is same as what is exported
> by miniLZO). So, can you please test with the version ported  - this
> is found in lzo/src/lzo1x_1.c (or in minilzo.c).
> Also, can you please use 'take 5' for your next testing?
>
> Thanks,
> Nitin

Will do.  (that's DBITS=15, correct?)

However, when I averaged it 100 times, lzo1x_1_11_compress() showed better 
speed than your implementation - about 1.5% faster. The *unsafe* 
decompressor, however, only shows about a 1.2% speed advantage over the safe 
decompressor.

DRH