Re: reiserfs empirical study (very long)

[Andrew Sharp <andy@netfall.com>]

Anton Blanchard wrote:
> 
> Me wrote:
> > The big endian patches change the code to use little endian ordering
> > for all on-disk structures.  IMO this is a mistake, and certainly
> > costs a dear performance penalty, because on big endian processors,
> > this method requires converting endianness both ways (reading and
> > writing) for all meta data.
> 
> Sane architectures (sparc64, ppc) have load/store with byte swap
> instructions and if reiserfs is using them you shouldnt see a
> performance penalty.
> 
> cpu_to_le* etc make use of them.

It does use them, but are these functions inlines with just one
instruction?  If not, then there is a penalty.  These handy
instructions just make that penalty a little less painful.

People have already mentioned enough cases here to convince me that
they do need to be moveable.  The one truly convincing case is
wanting to have reiserfs be the root file system, but the hardware
only allows one disk at a time.  One could be a hardass and say
yeah, just do it on a similar machine, but my coffee is kicking in
so I don't need to be a hardass anymore today.

I guess this is just another case where we suffer a small bite
because we're different from x86.  Sometime you bite the bear,
however.  Someone said that XFS is stored bigendian.  Which will be
great when it works on PowerPC/SPARC ~:^)

a