[OLPC-devel] Re: [olpc-software] Federico Mena-Quintero's recent memory work.

[Mike Hearn]

On 4/23/06, Jim Gettys <jg at laptop.org> wrote:
> It does also beg the question of whether the current Linux shared
> library scheme is a good one; some of the other ones don't have that
> overhead, as the mail thread points out.

That's not quite true. In the no-rebase case Windows PE is pretty
efficient, but, that only works because most DLLs are either:

a) Provided by Microsoft, so, their locations are known ahead of time and fixed.
b) Provided by the application itself, so, their locations can be
chosen not to conflict with the first set

The Linux model is not so organised, we just have this big pool of
libraries which come and go as they please. Pretty much every app will
have its own combination.

However when there ARE conflicts with PE, it rapidly becomes less
efficient, as the DLL has to be in-place binary edited which usually
makes the whole thing unsharable.

So it's a cost thing .... PE is more efficient in the best case, much
less efficient in the worst case, and ELF sits somewhere inbetween all
the time.

The real problems with ELF are speed and reliability related rather
than memory usage related - its global search scope not only makes it
impossible to guarantee reliablity of a deployed program but slows
things down significantly (which is why prelink was developed ....)

thanks -mike