fc_nblast can be very slow

[jefferis]

primarily due to reverse scores. But for some reason it seems to me that this can be much worse than using bigmat,

I just had an example with rstudio where it took a minute to do 1 x 3677 normalised scores. My suspicion is that it was actually some unrelated problem e.g. GUI update that was interleaved. These kind of error messages are typical when changing location with an X session in progress:

Tue Mar 18 14:32:29 mac1041-14.lmb.internal rsession[355] <Error>: kCGErrorIllegalArgument: CGSGetDisplayBounds (display 3f003d)

However, there is a still a more general speed issue to be tracked down since I have seen a behaviour where attaching an ff matrix to the global environment can bring access speed benefits.

[jefferis]

#13 and #14 might help a bit but still can't figure out origin of this occasional extreme slow behaviour. Triggers include

• rstudio
• accessing the diagonal or rows
• as opposed to a whole square matrix of the same size

[jefferis]

I have not had any recurrence of this since going back to bigmemory style on disk matrices. However there is still a problem because the released version of bigmemory does not run on Windows – although in theory Windows support was added back in Sep 2014.