Rybka3, Arena & Hashtable szie

[Kohflote]

Hi all,

I'm running Rybka3 on my pc (2Gb RAM), single cpu, Win-XP under Arena1.99beta5, fixed search depth=16 and different hashtable size for this position:

1. e4 d6 2. d4 Nf6 3. Nc3 g6 4. Be2 Bg7 5. f4 O-O 6. Nf3 c5 7. Be3

Each time, I "clear hash" and re-boot my pc before I use different hastable size. I am surprised with the following result

(1) hashtable size = 64Mb, move play is 7...Qa5, score +0.09, time taken is 02:56

(2) 128Mb, 7...cxd5, score -0.01, time:5:52

(3) 256Mb, 7...Qb6, score 0.00, time: 03:50

(4) 384Mb, 7...cxd5, score +0.02, time:11:20 (!)

(5) 512Mb, 7...Qa5, score +0.11, time: 04:32

My questions are:
(a) why at different hashtable size, the move played is different?
(b) 64Mb and 512Mb, Rybka selects the same move but time taken is quite significantly different. Similarly for 128Mb and 384Mb.
(c) Thus, at what hashtable size would (i) give the best move (ii) take the least time to find and select the best move?

Thank you!
Kah Huat, Koh
- from Singapore

[chesstango]

Go down, and down; try 32mb, 16, etc and ull be surprised.....

[Nimzovik]

Hmmmmmmmmm......... Perplexing. The search for chess truth is indeed baffling. Although in this case I would suggest the "horizon effect" is the issue... espcially as the scores the computer gives for the moves are relatively close.....

[Kohflote]

Hi all,

So far, no positive reply to my questions. Surely someone could throw some light to my questions, right?

Thank you.

Yours sincerely,
Kah Huat, Koh
- from Singapore

[Ovyron]

As much RAM as possible as long as there's no memory swapping, it should give the best results for analysis in most positions. It's different because Rybka saves different data depending on the Hash size.

Rybka only accepts powers of 2 so 384 and 256 should be identical.

[Kohflote]

Hi Ulysses,

Thank you for your reply. But I am still puzzled:

(a) the result for 384 and 256, i.e. the move played by R3 is different.
(b) the move played for 64Mb and 512Mb is the same but time taken is significantly different.

Thus, what hash size gives the best move and (if possible) takes the quickest time?

Yours sincerely,
Kah Huat, Koh
- from Singapore

[Ovyron]

Please provide the outputs from hash 256 and hash 384, something strange may be going on. Also, make sure that you always get the same output.

[chesstango]

what Kah Huat, Koh says ......

[Kohflote]

Hi Ulysses,

The output is in my message of this thread - I guess this is what you meant. Indeed I have tried twice and get the same result before I posted this thread.

Best Regards,
Kah Huat

[Ovyron]

I need to see output like this:

CPU0: AuthenticAMD x86 Family 6 Model 10 Stepping 0 2133 MHz
Engine: Rybka 3 1-cpu 32-bit (64 MB)
by Vasik Rajlich, Larry Kaufman
2.00 0:00 -0.21 7...Ng4 (344) 11
2.00 0:00 -0.41 7...Qb6 (544) 10
3.00 0:00 -0.06 7...Qb6 (1.321) 12
3.00 0:00 -0.21 7...Ng4 (1.452) 12
4.00 0:00 -0.33 7...Ng4 (1.762) 12
5.00 0:00 -0.38 7...cxd4 8.Bxd4 Nc6 (3.267) 13
6.00 0:00 -0.20 7...cxd4 8.Nxd4 Qb6 (7.085) 15
7.00 0:00 -0.08 7...cxd4 8.Nxd4 Nc6 9.O-O Qb6 (11.256) 15
7.00 0:00 -0.40++ 7...Ng4 (12.903) 15
7.00 0:00 -0.47 7...Ng4 8.Bg1 cxd4 (14.774) 15
8.00 0:01 -0.11 7...Ng4 8.Bg1 cxd4 9.Bxd4 Nc6 10.Bxg7 Kxg7 11.Qd2 (27.207) 14
9.00 0:04 -0.31++ 7...Ng4 (62.633) 15
9.00 0:07 -0.15 7...Ng4 8.Bg1 Nc6 9.d5 Na5 10.h3 Nf6 (104.826) 15
10.00 0:11 -0.10 7...Ng4 8.Bg1 cxd4 9.Nxd4 e5 10.fxe5 dxe5 11.Nf3 Qe7 12.Nd5 Qd6 13.Qd3 Bh6 (166.648) 14
11.00 0:25 -0.03 7...Ng4 8.Bg1 cxd4 9.Nxd4 e5 10.fxe5 Bxe5 11.Bxg4 Qh4+ 12.Bf2 Qxg4 13.Qxg4 Bxg4 14.h3 Be6 15.O-O-O Nc6 16.Ndb5 Rad8 (362.951) 14

From 256MB and 384MB, to compare them.