> In any case I browsed the code a bit more, it'd be easy to test out > the MMap'ing because the FST simply reads directly from a byte[], > which could be converted to a ByteBuffer (that's MMap'd). Nice!
The FST is basically a byte[], but like Robert said -- the layout of transitions and states will require nearly random access patterns all around it. There are some optimizations one can make to lay out initial nodes close to each other, for example (and I've done it as part of another project) -- these do optimize CPU cache utilization and I would assume MMAP'ped buffers as well -- but once you go past page fault zone the performance will be terrible. And I assume the degradation will not be smooth, but very sharp. This said, I'd be interested in looking at the benchmarks if you want to invest some time and do it ;) Dawid --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
