http://knol.google.com/k/linux-performance-tuning-and-measurement
ld.so (Dynamic Loader)

When a dynamically linked application is executed, the Linux loader, ld.so,
runs first.  ld.so loads all the application's libraries and connects
symbols that the application uses with the functions the libraries provide.
Because different libraries were originally linked at different and possibly
overlapping places in memory, the linker needs to sort through all the
symbols and make sure that each lives at a different place in memory.
When a symbol is moved from one virtual address to another, this is called
relocation.  It takes time for the loader to do this, and it is much better
if it does not need to be done at all.  The prelink application aims to
do that by rearranging the system libraries of the entire systems so that
they do not overlap.  An application with a high number of relocations
may not have been prelinked.

The Linux loader usually runs without any intervention from the user, and
by just executing a dynamic program, it is run automatically.  Although the
execution of the loader is hidden from the user, it still takes time to
run and can potentially slow down an application's startup time.  When you
ask for loader statistics, the loader shows the amount of work it is doing
and enables you to figure out whether it is a bottleneck.

The ld command is invisibly run for every Linux application that uses
shared libraries.  By setting the appropriate environment variables,
we can ask it to dump information about its execution:

me% env LD_DEBUG=statistics LD_DEBUG_OUTPUT=/tmp/ld$$ ls
Desktop  bin           doit.taskers  partitions  src
Mailbox  cron          etc           projects    sysctl.out
Maildir  doit.dirsize  libexec       rescue      usr-local

me% cat /tmp/ld1180.1849
1849:
1849: runtime linker statistics:
1849:   total startup time in dynamic loader: 716832 clock cycles
1849:             time needed for relocation: 200664 clock cycles (27.9%)
1849:                  number of relocations: 119
1849:       number of relocations from cache: 6
1849:         number of relative relocations: 0
1849:            time needed to load objects: 348336 clock cycles (48.5%)
1849:
1849: runtime linker statistics:
1849:            final number of relocations: 161
1849: final number of relocations from cache: 6