From issue1536.  This we can do without a format change.

Whoops! I just noticed this bug was wrong assigned to Luca Capello, when 
it should have been assigned to Luca Molteni instead.  My apologies to 
both Lucas

I got curious about cache performance, so I've written bash scripts that
generate 65K bogus cache files per cache subdir (patches/ inventories/
and pristine.hashed/), non-bucketed-ly and bucketed-ly:

http://hub.darcs.net/gh/bench/browse/do_cache_bogus.sh
http://hub.darcs.net/gh/bench/browse/do_cache_bogus_bucketed.sh

I did a test of comparing cloning some repo over http ( time darcs clone
--lazy http://hub.darcs.net/darcs/darcs-wiki ) with an empty cache or
with a big bogus cache, and found no relevant difference. I guess the
speed of accessing and writing a particular file is independent from how
many files are in its directory.

Now the slow operation is to list all files in a big directory.
Currently darcs never needs to list files. But I tested as follow on my
personal laptop (bought 3 years ago but still relevant):

time (tree cache/darcs/pristine.hashed/ |wc -l ) : 17s
time (tree cache/darcs/pristine.hashed/ |wc -l ) : 8s

When doing the same with 5 times as many files (325K files), the
difference grows to 1m35s versus 32s.

(I resetted the cache between each measure with sudo sh -c 'echo 3 >
/proc/sys/vm/drop_caches' , and I'm using ext4 on both machines.)

So, I'm not expecting darcs to run faster with the switch to bucketed
cache right now, nor slower, but I bet global cache garbage collection
will be (when it's implemented). Moreover the global cache will be more
manageable by third-party programs (like ls or rm :-) ).

The following patch sent by Marcio Diaz <marcio.diaz@gmail.com> updated issue issue1624 with
status=resolved;resolvedin=2.10.0 HEAD

* resolve issue1624: bucketed cache. 
Ignore-this: 2d077f5c10156e4a00631fbc4f8c3119