Bastian Blank dixit:
>I think it may be a good idea to only use -3 for this arches. This
I've also seen good results for -2e. It may well be worth the
overhead of -e, while reducing from -3 to -2 so it can be de-
compressed even quicker and with half the RAM use. Question,
is the increasement in time worth it?
The following tests have been run on an otherwise almost idle
single-core Celeron 2G4Hz (the ones with almost no L2 cache,
yeah… but that’s a bare iron I’ve got) on the data.tar member
of linux-image-3.2.0-rc4-486_3.2~rc4-1~experimental.1_i386.deb
(cache flushed, then preloaded with data.tar, result to tmpfs):
method\result time size relative
none 0 79554560 100%
gzip -n9 63.6s 29985449 37.69%
xz -2 64.4s 24441920 30.72%
xz -3 100s 24253792 30.48%
xz -2e 188s 22287644 28.02%
IMHO, the difference from xz -2 to xz -3 is not worth it, at
least for Debian packages’ data.tar member (which contains a
fair amount of already gzip’d data such as manual pages and
kernel images in addition to uncompressed data).
The difference from xz -2 to xz -2e weighs heavy on time,
almost tripling it but saving another 2 MiB (2 percentage-
points) is interesting.
My proposal therefore: for “regular” packages that wish to
migrate from gzip to xz, use xz -2 which is approximately
as fast. For packages with highly compressible data, use
xz -2e instead (and consider e.g. xz -4e on high-power
architectures such as amd64 – the power of the build dæmons
should be the one that counts, as decompressing is invariant
on the -e flag). Do not use xz -6.
Suggestions, thus, taking into account both the recent change
of libc packages to xz and waldi’s wish to change linux-2.6
to xz:
"fast": alpha *-amd64 ia64 s390 s390x
"maybe": hppa *-i386 ia64 powerpc powerpcspe ppc64 sparc sparc64
"slow": arm armel armhf avr32 m68k mips mipsel sh4
where fast maybe slow
not xz-using pkgs <--- gzip -n9 --->
most xz-using pkgs -2[e] xz -2 xz -2
linux-image-*, locales* xz -4e xz -3e xz -2e
libc6-dbg xz -6e xz -4e xz -2e
As for the distinction between fast/medium/slow, that's up to
the relative speeds of the architectures. I did guesswork. Note
that I put i386 (all three of them) as maybe, not fast, but if
all build dæmons are fast enough, sure.
Not sure how many cases you want to have. Fast machines could
use xz -2e as baseline.
Rule of thumb for the "linux-image-*, locales*" category:
large packages (several MiB of data) or highly compressible.
The below-most category is intended for debug packages and
stuff like that.
bye,
//mirabilos
--
“Having a smoking section in a restaurant is like having
a peeing section in a swimming pool.”
-- Edward Burr
--
To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
