Ok, a bit disappointing not to see any estimates, but I had pretty low expectations in the first place.
Anyway, here are the results of benchmarking (on AMD Athlon x2 3800 - single core used):
0) no caching tricks - search pattern is expected to be evenly distributed
1) 100 mln unique strings of 20 bytes each, data size: 2 GB
2) Indexing takes 30 mins
3) Generated index size is ~810 MB.
4) Running searches for 10,000 randomly selected strings, with 100 runs (total 10 mln searches) results in a sustained performance of ~232,000 (that's thousands) searches per second.
The system supports multiple indices, so it's not like everything is done for 100 mln strings: scalability is perfect because by adding 2nd CPU you will get double the speed, same goes with extra machines which will not offer extra speed but also load balancing and redundancy.
Could you do better than that? I doubt it. Not because I think I am so ****ing amazing, but because low level algorithms used in the system are such that mathematically pretty much perfect: you can't cheat them with probabilistic modelling because searches will be evenly distributed so caching tricks are off.