Re: [lc-devel] [RFC] LZO de/compression support - take 6

[Nitin G. <nit...@gm...>]

On 5/30/07, Daniel Hazelton <dha...@en...> wrote:
> I just noticed a bug in my testbed/benchmarking code. It's fixed, but I
> decided to compare version 6 of the code against the *unsafe* decompressor
> again. The results of the three runs I've put it through after changing it to
> compare against the unsafe decompressor were startling. `Tiny's` compressor
> is still faster - I've seen it be rated up to 3% faster. The decompressor,
> OTOH, when compared to the unsafe version (which is the comparison that
> started me on this binge of hacking), is more than 7% worse. About 11% slower
> on the original test against a C source file, and about 6% slower for random
> data.

Unsafe vs safe is within 10%. Its okay.

> However, looking at the numbers involved, I can't see a reason to keep
> the unsafe version around - the percentages look worse than they are - from 1
> to 3 microseconds.

Not just numbers. Most of applications cannot afford to use unsafe
versions anyway (like fs people).

(well, the compressed-cache people might want those extra
> usecs - but the difference will never be noticeable anywhere outside the
> kernel)
>
> DRH
>

compressed cache people require every single percent of that
performance. For now, ccaching is not ready for mainline (many things
need to be done). So, till then I will keep off the unsafe version. If
ever compressed caching is on its way to mainline _then_ I will try
and add back the unsafe version. But I see no other project that
really cares about unsafe version so it's okay to keep it off.


- Nitin