Author: Amir Ban
Date: 04:49:57 03/14/01
Go up one level in this thread
On March 12, 2001 at 16:09:46, Robert Hyatt wrote: >On March 12, 2001 at 15:29:15, Uri Blass wrote: > >>On March 12, 2001 at 13:57:32, Robert Hyatt wrote: >> >>>On March 12, 2001 at 13:49:41, Uri Blass wrote: >>> >>>>On March 12, 2001 at 13:15:08, Tanya Deborah wrote: >>>> >>>>> >>>>> >>>>>Hi Guys! >>>>> >>>>>I bought Deep junior6 3 days ago, and now i am interested to know if Deep J6 is >>>>>stronger than Junior6. I mean, running the 2 engines in one single processor. >>>> >>>>The programmer Amir Ban did not claim that Deep Junior is stronger on a dingle >>>>processor and I know that Deep Junior is sligtly slower than Junior6 on a single >>>>processor. >>>> >>>>Uri >>> >>> >>>While I didn't write the code, "slightly slower" should mean 0.1% slower at >>>most. I don't think you can tell the difference with that small of a speed >>>degradation. >> >> >>I remember that I read that it is not the case for Deep Junior and it is about >>10% slower for one processor. >> >>Uri > > >OK. I can't imagine how a program could be written to behave like that, but >I assume it is possible since you saw it somewhere... > Deep Junior is slower by a few percent. I think about 5%. It's not significant for playing strength. The reason Deep Junior is slower is rather easy to guess: The single-processor version has a single thread and context. The multi-processor has several. Boards, piece tables, evaluation data and what not have several concurrent copies so many procedures have an extra context parameter, and often more indirecr addressing. Transposition table handling is also more complex because things are easier when positions are handled serially. This overhead is of course absolutely necessary when actually having several processors, but is dead weight in a single processor. When I wrote the Deep version I expected the overhead to be higher, and overall I think the Deep Junior code does a good job of keeping the penalty low. Having only a 0.1% overhead seems to indicate that you don't have a version optimized for single processors. Amir >In the case of Crafty, .1% is the right number. In the "main search loop" there >is a single test/branch that is executed needlessly, once per node, for the >SMP version when there is no extra processors being used.
This page took 0 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.