Author: Dan Newman
Date: 16:21:27 07/20/98
Go up one level in this thread
On July 20, 1998 at 17:02:51, Roberto Waldteufel wrote: > >On July 20, 1998 at 15:20:24, Dan Newman wrote: > >>On July 20, 1998 at 14:36:12, Bruce Moreland wrote: >> >>> >>>On July 20, 1998 at 14:30:47, Dan Newman wrote: >>> >>>>One of the biggest effects I see is that I get a much larger node rate >>>>the longer I run. If I run a test for 10 s I might get 170 knps. If >>>>I run for 40 s 200 knps. It does level off of course. I'm using a >>>>P6/200, so I've attributed this effect to improving branch prediction >>>>as the code runs. Any ideas? >>> >>>Maybe it has something to do with the deeper search having characterstics that >>>differ from the shallower search. >>> >>>bruce >> >>I bet that's it or at least a lot of it. The major difference is that >>the deeper plies will have fewer pieces on average. So move generation >>will be a little cheaper, a smaller portion of the moves generated >>will be wasted, and so forth. >> >>-Dan. > >Have you considered that the deeper iterations contain the most transpositions, >so you will process more nodes quickly by hash table look-up? I think this must >surely account for at least some of the difference. > >Best wishes, >Roberto I hadn't thought about that either--but I've seen the effect even without a transposition table. Actually, I'm beginning to return to my original guess--I've noticed this effect when benchmarking the move generator in isolation too, just generating the same set of moves over and over. The longer the run, the more move generations per second. I imagine this could be due to the branch prediction hardware settling into an improved (but probably not optimal) pattern. Or, it could be the OS doing a lot of stuff immediately after loading the program in--disk accesses or whatever. Just one more in a series of quirky, unfathomable computer behaviors. -Dan.
This page took 0.04 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.