[r-t] Big search

[Graham John]

Don wrote:

> I'm typically seeing node rates in the 0.5 to 5 million per second range.

I designed SMC (the forerunner of SMC32) with a lead being the node (i.e.
the point at which there is potential for a branch). In developing SMC32,
Mark optimized this by collapsing leads between possible branches (based
upon the specified search criteria) into a single node. Hence Yorkshire with
calls at Middle, Wrong and Home has three nodes per course H-M, M-W, W-H.
This is what SMC32 is counting, and based on Mark's big search gives 287.9
trillion nodes divided by 1271 hours, averaging 62.9 million nodes per
second.

Finding compositions, though, slows processing down, as the program then has
to evaluate/store each composition found. The very high number of
compositions found (100,000 per second) will have had a significant slowing
effect on progress.

If SMC is used to search for bobs only extents of Bristol S Major tenors
together (yes, I know it's not possible, that's the point - to stop it
evaluating compositions found), it achieves a search speed of around 6-7
million leads per second i.e. it is the of the same order as your number,
Don. The additional optimizations in SMC32 (32-bit code, node collapsing
etc), achieve more than a ten-fold increase over SMC's performance.

Graham