Author: Vincent Diepeveen
Date: 15:42:08 04/30/01
Go up one level in this thread
On April 29, 2001 at 12:48:09, Bruce Moreland wrote: >On April 29, 2001 at 04:01:32, Uri Blass wrote: > >>On April 29, 2001 at 03:39:53, Bruce Moreland wrote: >> >>>It seems clear why it wasn't invited to participate: >>> >>>1) The organizer is going to use a multiprocessor machine. >>>2) Fritz and Junior run on a multiprocessor machine. >>>3) Tiger is known to not run on such a machine. >>>4) Tiger is very strong, but if it is stronger than Junior or Fritz, it's >>>probably not stronger by much. >>>5) A multiprocessor machine should produce a significant performance boost. >>>6) It is hoped that the event will produce an "accurate" winner. >>> >>>If you allow these points, you can make a case that Tiger on a single processor >>>can't be stronger than Junior or Fritz on a multi. >> >>I understood that tiger can use more than one processor. > >Did Enrique know that? > >>I guess that Tiger could earn less than Fritz or Junior because of the fact that >>Christophe and Ed had not enough time to optimize tiger for more than one >>processor but it still can earn something from more than one processor. > >If he has a multiprocessor version, yes. > >bruce > >>I am interesting to know how much tiger earns from more than one processor. >> >>If there is a parallel version it is easy to compare both versions at least in >>test positions and to give an estimate how much is it faster on 2 processors. >> >>I guess that for most programs like Junior or Fritz the number is something like >>1.7 times faster. >> >>Uri Tiger speedup i estimate would be closer to 3.0 as 1.7 if using a tough parallel algorithm. With the french algorithm it would be easily 1.7 for tiger. Reason is that tiger has a small branching factor and a lot of pruning. If you forward prune a lot the parallel win is way bigger. If you extend lot of nonsense the parallel win is smaller too. For example if i remove all 'dangerous extensions' from DIEP then even in blitz i get a 1.8 speedup. Right now i get in blitz like 1.4 or something. Not *near* 1.7 or 1.8. At 3 minutes a move i get between 1.8 and 2.0 with all dangerous extensions though. My parallellism improves my b.f. simply. So if i would run single cpu 2 processes at the same time then also b.f. gets faster. However, initially search speed is a few hundreds of nodes a second, so the overhead is a lot, because before it starts splitting at cool positions the hashtables need to get filled bigtime and it takes time to get back to the single cpu speed :) Also not to be forgotten for tiger is that its speed is not so fast as from Fritz and junior. This means you have less overhead so getting a good speedup is way easier then. Same is valid for DIEP. it's dead slow, only 40k nps at a 1Ghz machine, so getting a good speedup is way easier as i lose less to overhead. Under linux my overhead is about 3%, note it was way more with previous compiler version 2.95.2...
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.