The initial building (and an occational compression) performance just got a major boost. I knew there was a problem there, but I hadn't realized it was this easy to fix: http://hg.dovecot.org/dovecot-1.1/rev/4445415da4ff
My 1,4 GB test lkml mbox can be now indexed in 7 minutes, doubling the speed form earlier 14 minutes. I added some more benchmark results here: http://wiki.dovecot.org/Plugins/FTS
So the initial Squat index building is about 2-3x slower than the non-indexed search. I think that's pretty good. Also I think it means that the non-indexed search code could be optimized more. :)
Maildir test shows that it indexes about 5 MB/CPUsec. My earlier tests showed that it indexed 8-10 MB/CPUs when feeding data directly to it, so 3-5 MB/CPUs is going elsewhere, probably to message parsing and charset conversions and such..