> The way I see it, increasing FD_SETSIZE would have a very small space
> penalty, a small time penalty for fd-set ops (copying 32 bytes rather
> than 8), but the user/kernel interface costs are small-to-zero unless
> one explicitly ups the per-process descriptors limit: since
> sys_select() clips at fd_nfiles.

very large descriptor sets have severe scaling problems.

If you have a large, sparse set of descriptors, use poll(2).	

If you have a large, dense set, you have an interesting interface
design problem to make this scale well.  see solaris /dev/poll for one
approach.  I suspect that there's lot of room for more innovation
here.

					- Bill