System designers have a responsibility to help the rest of the
system's community by providing a useable system.  I'll make the
following observations regarding shared libraries as if you and I were
debating putting shared libraries into a system.  I have worked both
sides of this issue, so I have an experienced opinion, not just apriori
idea of good system hygiene.

First, let me deal with your two `advantages.'

> #1 Maintenance - If you have 50 programs that depend on one library
> and you have a fix for the library how many things do you want to
> "remember to build"?  (though people are throwing up straw man
> arguments for this too.  I suspect the worst case scenario is not
> always the common case though.)

This has an ugly other side.  If I have 50 programs that depend on one
library and fix a bug in the library, I now have 50 live programs in
production that haven't been tested.  I don't know if my change has
broken something else and won't know until I re-test all 50 programs.
I now have to do all that testing at once.  With static libraries I
have the luxury of going thru the 50 programs one at a time and
relinking them with the new library or reverting to an old library if
there's a problem.
> 
> #2 Supposed physical memory savings - libSystem on Mac OS X only
> exists in memory 1 time for all the programs that use it... sorta. 
> Read Only pages are shared, writable pages are COW and yes, this adds
> a good deal of complexity to the VM of the OS to have this.

Before X windows, 10% of programs were from the library.  With X
windows that number ballooned to 90%, so there is an apparent (see
below) space savings with X windows.  This is because when to touch
anything you get everything.  That, I would argue, is bad design.  (I
added shared library to a custom embedded Unix I did 15 years ago
because the application was also mis-designed.) The X library is
really a sub-system that should be somewhere other than the user's
program.


Also on the down side, shared libraries make it hard to distrubute
binaries without also sending out the library.  The binary, or even
the files in the same directory, are no longer all that is required to
run the program.  You also have to have the correct version of the
library.  Since you get different binaries that need different
versions of the library, you now wind up with three or six or more
slight variations of the same library, all get loaded into core.  Now,
where's my space saving?  Remove one and now you have a bunch of
programs that don't work, but you don't know it yet.

To load a binary that is linked to a library takes longer to load.


I have seen code sharing done right, such as Oberon, but every version
of shared libraries in Unix I've seen cost more than they were worth.
Including my own.

Hope the above arguments at least seem rational.  They are off the top
of my head.  I'm sure others on 9fans have other, better data points
that do the engineering calculation on the cost/benefits of shared
libraries.


 bc

Reply via email to