I did some experiments in R and I think that variable sized Q blocks as in my latest document on the JIRA should be just as accurate as constant sized blocks.
There is a slight efficiency degradation with too many small blocks, but no real accuracy concern. On Tue, Oct 26, 2010 at 12:30 PM, Dmitriy Lyubimov <[email protected]>wrote: > PPS > > not sure about numerical stability with 100% streaming though. > > if that turns out to be a problem, i can always go back to multiple Q > blocks > per mapper though to improve on it. >
