Re: Debian Mirror with lzma compressed packages
Russell Coker <firstname.lastname@example.org> writes:
> On Friday 19 January 2007 02:19, Daniel Baumann <email@example.com> wrote:
>> For the KDE flavour, this takes less than two minutes for downloading
>> the packages, but about 5 minutes for unpacking them. This is done on a
>> reasonable fast i386 machine (3.2ghz, 1gb ram, two 250gb barracudas in
>> raid0). Now, if we change from gzip to a slower algorithm, the unpack
>> part will take even more time for this :/
> Last time I checked the gzip source had no assembler optimisation for systems
> other than i386. So if your 3.2GHz machine (which obviously would be a P4 at
> least not an i386) is running the AMD64 instruction set then you could
> probably improve performance by running the 32bit binary.
> A compression algorithm that is more complex (slower) than the gzip algorithm
> but which is implemented in 64bit assembler might compare well with the
> current gzip performance for modern machines.
Lets see: AMD Athlon(tm) 64 Processor 3200+
gunzip-64 < linux-source-2.6.18.tar.gz > /dev/null 2.16s user 0.03s system 98% cpu 2.211 total
gunzip-32 < linux-source-2.6.18.tar.gz > /dev/null 2.31s user 0.03s system 94% cpu 2.481 total
bunzip-64 < linux-source-2.6.18.tar.bz2 > /dev/null 14.55s user 0.14s system 98% cpu 14.900 total
bunzip-32 < linux-source-2.6.18.tar.bz2 > /dev/null 18.64s user 0.13s system 98% cpu 19.034 total
I'm told the intel em64t mode isn't as fast as amds so maybe the
results differe there. But here the extra registers gain you more than
the asm code for i386 does.