Author: Robert Hyatt
Date: 10:44:41 09/04/02
Go up one level in this thread
On September 04, 2002 at 13:03:23, Vincent Diepeveen wrote: >On September 04, 2002 at 12:25:09, Robert Hyatt wrote: > >the time control is 3 minutes a move at most tournaments, >it sure used to be in world champs. it was acutually much more. It was 40/2hrs. Cray Blitz averaged searching 5-6 minutes per move, with book moves and pondering saving a lot of time. If you look at the 16 processor times in the DTS article, those match exactly with the times used in the actual game that year. If you notice, I excluded "book positions". The first search was for over five minutes. The next was almost 5. Move six was searched for 7 1/2 minutes, and move 8 was searched for over 11 minutes. Move 17 was searched even longer. All of them are over 180 seconds except for one "easy or forced" move at move 18... that's why i wanted to use that as the basis for the experimental setup, because those were _real_ numbers from a real game, using the 16 cpu machine. We might have lost the game, and definitely should have drawn it, but I didn't pick the game for its chess value, I picked it because it had a wide variet of problems in it. Fail highs. Fail lows. Score drops. Score increases. Deep searches. Deep draws. You name it. It seemed like a perfect cross-section of what we would see in a tournament. I could have picked a game where we blew out out opponent. Those generally produce better speedups. I tried to pick something representative instead. > >Please also report the node counts. i have a version of >crafty here, 18.15 which is compiled with msvc 6 sp4 processor >pack (P6 optimalisations which default m$ 6 doesn't have). What node counts am I to report? I'm not going to run any tests for you. you are a grown man. Surely you can do that for yourself. > >It is interesting to realize that cray blitz had about 500k nodes >a second at that cray machine when using 16 processors. So? that was the rough speed on the C90. Is there a point in why that is "interesting"??? This version also didn't have SE. Is that interesting? 1994 was the only year we used that and it blew up badly... Are there any more things that are "interesting" you'd like to share here? > >>On September 04, 2002 at 12:11:53, Slater Wold wrote: >> >>>Bob, do you have those 20 positions from the ICCA article? >>> >>>If so could you e-mail or post them here for me? I would greatly apprecaite it. >> >> >>you mean the 24 positions from the MchessPro vs Cray Blitz game that I used >>in the JICCA article? If so, here they are... >> >>r2qkbnr/ppp2p1p/2n5/3P4/2BP1pb1/2N2p2/PPPQ2PP/R1B2RK1 b kq - ; bm a1a1 >>r2qkbnr/ppp2p1p/8/nB1P4/3P1pb1/2N2p2/PPPQ2PP/R1B2RK1 b kq - ; bm a1a1 >>a1a1 r2qkbnr/pp3p1p/2p5/nB1P4/3P1Qb1/2N2p2/PPP3PP/R1B2RK1 b kq - ; bm a1a1 >>black a1a1 >>r2qkb1r/pp3p1p/2p2n2/nB1P4/3P1Qb1/2N2p2/PPP3PP/R1B1R1K1 b kq - ; bm a1a1 >>a1a1 r2q1b1r/pp1k1p1p/2P2n2/nB6/3P1Qb1/2N2p2/PPP3PP/R1B1R1K1 b - - ; bm a1a1 >>black a1a1 >>r2q1b1r/p2k1p1p/2p2n2/nB6/3PNQb1/5p2/PPP3PP/R1B1R1K1 b - - ; bm a1a1 >>r2q1b1r/p2k1p1p/2p5/nB6/3Pn1Q1/5p2/PPP3PP/R1B1R1K1 b - - ; bm a1a1 >>r2q1b1r/p1k2p1p/2p5/nB6/3PR1Q1/5p2/PPP3PP/R1B3K1 b - - ; bm a1a1 >>r2q1b1r/p1k2p1p/8/np6/3PR3/5Q2/PPP3PP/R1B3K1 b - - ; bm a1a1 >>r4b1r/p1kq1p1p/8/np6/3P1R2/5Q2/PPP3PP/R1B3K1 b - - ; bm a1a1 >>r6r/p1kqbR1p/8/np6/3P4/5Q2/PPP3PP/R1B3K1 b - - ; bm a1a1 >>5r1r/p1kqbR1p/8/np6/3P1B2/5Q2/PPP3PP/R5K1 b - - ; bm a1a1 >>5r1r/p2qbR1p/1k6/np2B3/3P4/5Q2/PPP3PP/R5K1 b - - ; bm a1a1 >>5rr1/p2qbR1p/1k6/np2B3/3P4/2P2Q2/PP4PP/R5K1 b - - ; bm a1a1 >>5rr1/p2qbR1p/1kn5/1p2B3/3P4/2P2Q2/PP4PP/4R1K1 b - - ; bm a1a1 >>4qRr1/p3b2p/1kn5/1p2B3/3P4/2P2Q2/PP4PP/4R1K1 b - - ; bm a1a1 >>5qr1/p3b2p/1kn5/1p1QB3/3P4/2P5/PP4PP/4R1K1 b - - ; bm a1a1 >>5q2/p3b2p/1kn5/1p1QB1r1/P2P4/2P5/1P4PP/4R1K1 b - - ; bm a1a1 >>5q2/p3b2p/1kn5/3QB1r1/p1PP4/8/1P4PP/4R1K1 b - - ; bm a1a1 >>5q2/p3b2p/1k6/3QR1r1/p1PP4/8/1P4PP/6K1 b - - ; bm a1a1 >>5q2/p3b2p/1k6/4Q3/p1PP4/8/1P4PP/6K1 b - - ; bm a1a1 >>3q4/p3b2p/1k6/2P1Q3/p2P4/8/1P4PP/6K1 b - - ; bm a1a1 >>3q4/p3b2p/8/1kP5/p2P4/8/1P2Q1PP/6K1 b - - ; bm a1a1 >>3q4/p3b2p/8/2P5/pk1P4/3Q4/1P4PP/6K1 b - - ; bm a1a1 >> >>Note that these were produced by Vincent. I had to edit them as he had >>bad castling status. Note also that the "best move" is nonsense, but in >>these positions, "best move" is not important. >> >>How to run the test: >> >>1. set some search time limit with the st=n command. >> >>2. type "test filename" where the above file is used. >> >>Run this with mt=0 and mt=2. >> >>Now comes the fun. For each position in the mt=0 run, find the _last_ >>bit of output, such as a PV change or the end of iteration. Write the >>time down. Flip to the mt=2 run and for that position, find the _same_ >>depth/score/move output and write down that time. Now you can compute >>a speedup for that position. >> >>Repeat for all 24 positions and you have 24 speedups. >> >>Now you come to a decision point. If you do as I said above, you will then >>average the 24 speedups to get an overall speedup. But this is just one way >>to do it, because this counts all positions equally in the final speedup. >>But suppose some positions have a much longer "time" value? In a real game, >>won't their speedup affect the overall average "more"? So you might choose to >>go back and take the times as derived above, add them together, and _then_ >>divide the mt=2 time into the mt=0 time to compute an overall speedup. >> >>There is also another way but it is not as easy... >> >>use st=999999, sd=N to specify a specific search depth for every position to >>be searched to. Then you can use the total elapsed time for mt=0 and divide >>that by the total elapsed time for mt=2, and compute the speedup that way. >>The drawback is that if you pick too deep a search, some positions will take a >>_long_ time. If you pick it too shallow, some will finish very quickly. >> >>You _can_ get at least three different speedup values from the above, and which >>one you use is really optional...
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.