Razor,
you mentioned xtool does auto detection on the first stream. Does it iterate through all 81 levels until match is found, or only few known to be used most often?
As for my idea,
It don't need to go through all 81 every time, only until match is found and then next stream can be tested with that level found in previous stream first. This means if streams are not mixed there will be no slowdown because first try is a match already. Also your own test implementation on first stream seem fast, why not use same technique, whatever is it doing?
I have never noticed ztool/xtool being slow at beginning, it seems to start immediately. Would there really be such big difference if it had to try 10 streams instead of 1, especially with optimization I mentioned above(starting with previously found level)? Perhaps up to few seconds of delay should not be much problem in scope of thousand streams and GB's of data?
As for MT, that's not a big deal we don't want to test whole data only first couple of streams(again assuming it should not take long).
You mentioned using stats, that sounds actually very similar to what I am talking about. In fact, if I understand correctly you may have already implemented what I am talking about here
. But interestingly, you mentioned collecting stats from threads, that pretty much equals to cmp levels. So you in fact are collecting from multiple streams, not just first?