Re: [Gnu-arch-users] Some issues

[Matthieu Moy]

Andrew Suffield <address@hidden> writes:

>> Do you have  a reference on this?  I've just made the test  on an ext3
>> filesystem, and reading  10000 files of 1 lines  was incredibly faster
>> than 1 file of 10000 lines.
>
> That's a fairly useless test. You're sampling buffer-cache performance
> for sequential access, not filesystem performance for random access.

But isn't that similar  to what tla has to do to  collect file ids for
your local tree?

To  fix the  ideas:  Suppose you  have  a project  which  is just  one
directory  with 100  files in  it, using  explicit IDs.  tla  add will
create 100 ID files in addition to those source files.

Now, if you  run for example "tla tree-lint", one  of the jobs carried
out is to collect the IDs, and look for duplicates. For that, tla need
to access  the 100 ID files,  and read one  line in each file.  With a
1-big-file approach,  tla would have  to read only one  100-lines long
file,  and parse  it (There's  an overhead  due to  parsing).  This is
simmilar to the little benchmark  I've just carried out on ext3 unless
I've missed something.

Off course, file moving/adding/removing would be more expensive with a
1-big-file. There's no perfect approach.

-- 
Matthieu