fdesc allocation optimization

[Matthew Dillon]

:Alexey Dokuchaev <danfe at FreeBSD.org> writes:
:> i've been browsing some of dfbsd resources recently, and found this one
:> being pretty interesting:
:>
:> http://leaf.dragonflybsd.org/mailarchive/commits/2005-06/msg00526.html
:>
:> however, it seemingly did not get attention in our lists.  so i am
:> wondering if there are work/plans on porting hsu@'s work?  i remember
:> that at some point we adopted some openbsd-derived algorithm, but since
:> matt states that this is "far better algorithm then anything we or
:> freebsd thought up before", i figured it worth a look.
:
:Bollocks.  Our current algorithm (which I wrote) is so fast you don't
:even notice it's there.  It's actually simpler than the OpenBSD code
:from which it was inspired, and in theory it should be slower, but I
:discovered that the overhead of the "better" algorithm was so high
:that it consistently lost to the simpler one for reasonable amounts of
:file descriptors (up to about 100,000 per process).
:
:The source code for the microbenchmark I used, and selected graphs
:comparing my code to the previous implementation, are available at
:<URL:http://people.freebsd.org/~des/fdbench/>.
:
:(the strange artifacts you see on the red graphs are the result of the
:file descriptor table overrunning the CPU cache)
:
:DES
:-- 
:Dag-Erling Smørgrav - des at des.no
    
    Well, I expect that once you get past a pure linear search any algorithm
    will fast enough that it probably doesn't matter in real life.  But I 
    would not pooh-pooh Jeff's implementation of the Solaris algorithm so
    quickly.  It is the fastest, cleanest, most elegant implementation of
    a file descriptor handling algorithm that I've ever seen in my life.
    Anyone who actually reads the code will come away wondering why FreeBSD 
    is still screwing around with linear bitmaps.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>