[OLPC-devel] just verified: 2306 + olpc + linux kernel works.

David Woodhouse dwmw2 at infradead.org
Thu May 25 20:52:12 EDT 2006


On Fri, 2006-05-26 at 02:33 +0200, Carl-Daniel Hailfinger wrote:
> 
> 7571810 libgcj.so.7.0.0.gz.4kchunks
> 8043146 libgcj.so.7.0.0.bz2.4kchunks
> 6477839 libgcj.so.7.0.0.lzma.4kchunks
> 
> Ran it again with 16k sized chunks:
> 
> 6693730 libgcj.so.7.0.0.gz.16kchunks
> 6608000 libgcj.so.7.0.0.bz2.16kchunks
> 5451594 libgcj.so.7.0.0.lzma.16kchunks
> 
> Hmmm... strange results indeed. I'll leave it to you to interpret
> them.

Why so strange? They seem fairly much in line with what I expected. 

Using 16KiB chunks instead of 4KiB gives you about a 10% improvement in
gzip performance, and a better improvement in bzip2 because bzip2 really
sucks without a large window.

LZMA also seems to get a fairly significant improvement (~15%) by going
to 16KiB blocks, which isn't massively surprising. And it beats gzip by
15% or more in both cases.

If it's slow to compress but relatively quick to _decompress_, that can
still be extremely useful. One of the items on my TODO list is to have
per-file selection of compression algorithms -- we could select LZMA for
stuff like libraries and executables which are usually only written once
and read many times.

-- 
dwmw2




More information about the Devel mailing list