Glaurung: Beta testers with quad-CPU machines wanted!

[Tord Romstad]

Hi all,

I am now almost finished with a new Glaurung version with support for more than 2 CPUs. Unfortunately I cannot test it very well myself, because I don't have any quad-CPU machine available for testing.

If somebody is willing to help me test the quad-CPU version, please drop me an e-mail or write me a personal message on this board. I can't offer a binary, you therefore need to be able to compile the source code yourself.

It would also be really cool if someone has a quad (or better!) I could use as a tournament machine during the tournament in Mainz (August 17-18).

Tord

[Tord Romstad]

After some quick tests performed by Andreas Guettinger on a quad-CPU PowerMac G5, I'm feeling slightly more confident that my multi-CPU version of Glaurung works. I've now decided to release a public beta version:

http://www.glaurungchess.com/beta/g060522-src.tar.gz

Only the source code is available at the moment. If the source code is compiled without changes, you will get an executable with a maximum of 4 threads (but the default is still 2 threads, you will have to manually change the "Number of threads" UCI parameter in order to make the program use 4 search threads). If you only have 2 CPUs, you can make a binary with a slightly smaller memory footprint by changing the constant MaxNumOfThreads (found near the top of glaurung.h) from 4 to 2. If you have a monster machine with more than 4 CPUs, you can also increase MaxNumOfThreads. I have tested with 10 threads, and the program still appears to work fine. Finally, you can also compile a single-CPU executable by setting MaxNumOfThreads, or (better) by commenting out the line with #define SMP in glaurung.h.

Please note that this is a highly experimental beta version. It should not be used in serious games and tournaments. It is possible, perhaps even likely, that it is stronger than Glaurung 1.1 on single or dual-CPU machines, but until it is more thoroughly tested I cannot recommend it for serious use. Those who run serious tournaments should keep using Glaurung 1.1 until Glaurung 1.2 is released (probably some time between early June and early August).

I would be very happy to see some more informal tests of this beta version, though, especially on quad-CPU machines. My main purpose for releasing this beta is to find and fix the most obvious bugs in my multi-threaded search before I release my next version.

Tord

[Volker Pittlik]

...Finally, you can also compile a single-CPU executable by setting MaxNumOfThreads, or (better) by commenting out the line with #define SMP in glaurung.h...

Done.

... It is possible, perhaps even likely, that it is stronger than Glaurung 1.1 on single or dual-CPU machines, but until it is more thoroughly tested I cannot recommend it for serious use.

My few data do not confirm that:

Code: Select all

Computer chess game
XP2000, 180+3, ponder off, Ubuntu, 2006.05.16 - 2006.05.22
                       Score              Glau     Ruff    Toga Jonn    Frui    Craf     The
--------------------------------------------------------------------
 1: Glaurung 1.1     14.0 / 24   XXXX 00== 0100 ==01 =10= 1111 1111
 2: Ruffian 2.0.0        3.0 /  4   11== XXXX .... .... .... .... ....
 3: Toga 1.2              3.0 /  4   1011 .... XXXX .... .... .... ....
 4: Jonny 2.83           2.0 /  4   ==10 .... .... XXXX .... .... ....
 5: Fruit 2.1               2.0 /  4   =01= .... .... .... XXXX .... ....
 6: Crafty-19.20         0.0 /  4   0000 .... .... .... .... XXXX ....
 7: The Baron 1.7.0   0.0 /  4   0000 .... .... .... .... .... XXXX


Code: Select all

Computer chess game
XP2000, 180+3, ponder off, Ubuntu, 2006.05.16 - 2006.05.22
                       Score                 Glau     Jonn Toga  The     Craf   Frui    Ruff
--------------------------------------------------------------------
 1: Glaurung 060522   9.0 / 24   XXXX 0000 0000 =0=1 0101 1=10 10=1
 2: Jonny 2.83              4.0 /  4   1111 XXXX .... .... .... .... ....
 3: Toga 1.2                 4.0 /  4   1111 .... XXXX .... .... .... ....
 4: The Baron 1.7.0     2.0 /  4   =1=0 .... .... XXXX .... .... ....
 5: Crafty-19.20           2.0 /  4   1010 .... .... .... XXXX .... ....
 6: Fruit 2.1                 1.5 /  4   0=01 .... .... .... .... XXXX ....
 7: Ruffian 2.0.0          1.5 /  4   01=0 .... .... .... .... .... XXXX


Those who run serious tournaments should keep using Glaurung 1.1 until Glaurung 1.2 is released ...

Fortunately I run not serious tournaments only

Thanks for the engine in any case.

Volker

[Tord Romstad]

... It is possible, perhaps even likely, that it is stronger than Glaurung 1.1 on single or dual-CPU machines, but until it is more thoroughly tested I cannot recommend it for serious use.

My few data do not confirm that:

Thanks for the data, but I'm still not convinced. I don't like to give precise information about my own test results, but I can say that Glaurung 060522 scored better than Glaurung 1.1 in Silver matches against Fruit 2.1 and Gambit-Fruit 1.0 Beta 4bx. It also won a self-play Silver match between 060522 and 1.1. However, the differences are small and well within the margins of error.

Thanks for the engine in any case.

And thanks to you for testing! It's a pity you don't have a quad.

Tord

[Volker Pittlik]

... but I can say that Glaurung 060522 scored better than Glaurung 1.1 in Silver matches against Fruit 2.1

Because I never used them I'll try that over night.

...It's a pity you don't have a quad...

Yes!

Volker

[Volker Pittlik]

...

Because I never used them I'll try that over night.
...

Oops no. Is there something like Epd2wb ready for use for Linux?

Volker

[Tord Romstad]

...

Because I never used them I'll try that over night.
...

Oops no. Is there something like Epd2wb ready for use for Linux?

I have no idea. What does Epd2wb do?

Tord

[Uri Blass]

...

Because I never used them I'll try that over night.
...

Oops no. Is there something like Epd2wb ready for use for Linux?

I have no idea. What does Epd2wb do?

Tord

Epd2wb is a tool that can be used to ask engines to analyze test suites and calculate the results.

It is possible to ask many engines to analyze different test suites
and save the result in different files by the right batch file that you run.

Uri

[Tord Romstad]

What does Epd2wb do?

Epd2wb is a tool that can be used to ask engines to analyze test suites and calculate the results.

It is possible to ask many engines to analyze different test suites
and save the result in different files by the right batch file that you run.

Thanks for the explanation. I don't know of any equivalent tool for Linux, but at least PolyGlot can be used to run an EPD test suite with a single UCI engine. Use a command line like this:

Code: Select all

polyglot epd-test -epd ecmgcp.epd -min-time 5 -max-time 10

The output looks like this:

Code: Select all

"ECM.946" 1    1    1 -  7   0.50    186585  +5.46 Bg4 Qf2 d4 Re1 dxe3+ Nxe3 Nd5 Bc4 Kh8 Bxd5 cxd5
"ECM.949" 1    2    2 - 10   8.99   5834307  +2.67 Nxg4 fxg4 Bxg4 Qxg4 Rh2+ Kxh2 Qf2+ Kh1 Bxg3 Qh3 Rh8 Qxh8+ Kxh8 Re2 Qxe2
"ECM.952" 1    3    3 - 11   5.64   4280776  +0.46 Nd5 cxd5 c3 Qxb2 Qxd4 Qb6 Qxb6 cxb6 exd5 Nxd5 c4 Nc3
"ECM.961" 1    4    4 -  7   0.12     79296  +1.17 Bxd5 Ra7 Be4 Nd7 Na4 Qb8 Re1 h6 f4 Rc7 Qg4 Qa7+ Kh1
"ECM.973" 1    5    5 -  9   0.77    560190  +1.87 Rxc8 Nxc8 Rd7 c3 Qc1 Rb3 Rxe7 Bxe7 Qc2 Ra3 Kf1 Ra1+ Ke2
"ECM.977" 1    6    6 -  8   1.18    857810  +1.25 Rxa1 Rxa1 Bxd4 Rxd4 Nxb3 Rd2 Nxa1 Qxa1 Be6 Rb2 hxg3+ fxg3
"ECM.978" 1    7    7 -  9   2.82   1991600  +1.50 Nxc6 bxc6 Bxd6 Bxd6 Qxc6 Qxe2 Qxd6 Qe6 Qxe6+ Bxe6 Rfe1 Ne4
"ECM.986" 1    8    8 -  9   0.35    242640  +2.32 d3 Bxd3 Qe5 Nb2 Ba3 Rab1 Nb4 Qe2 Bxb2 Qxb2 Qxb2 Rxb2 Rxd3
"ECM.987" 1    9    9 - 10   3.35   2385334  +2.96 Qxf7+ Rxf7 Ra8+ Ne8 Nh6+ Kh8 Rxf7 Rc7 Bg5 Bxg5 Rxd7 Rxd7 Rxe8+ Kg7 Ng4 Ra7 Kh2

... (big snip) ...

"ECM.1618" 1  161  179 -  6   0.07     30059  +1.17 Bxh7+ Kxh7 Rxf6 Nd5 Rf3 Nxe3 Rxe3 exd4 cxd4 b5 Qc5 Kg8
"ECM.1620" 1  162  180 -  8   0.37    237687  +2.35 Bxh7+ Kxh7 g6+ fxg6 Qh3+ Kg8 Qxe6+ Kh7 Rf3 Nf6 exf6 gxf6 Qxf6 Qg7 Rh3+ Kg8
"ECM.1621" 0  162  181 - 10   3.63   2522772  -0.25 Qb8 Qxb4 Nxg4 Qxb8 Raxb8 h3 Nf6 d3 g4 hxg4 Nxg4 b4
"ECM.1622" 0  162  182 -  4   0.01      2619  +0.57 Bd2 h6 Nf3 h5 Qh3 Ba4 Rab1 Qd7 Ng5 Rfe8 g4 Nc6 Qf3 hxg4 Qxg4
"ECM.1623" 0  162  183 -  8   0.49    303773  -0.14 Qd4 f6 Ng5 Bg7 Ne6 Qe7 Nxc7 Qxc7 Qe4 Qf7 Qf3 h5 Rh4
162/183 - 7.9 1.19 867464


The different columns in the output are: Fen ID string, solved or not (1 means solved, 0 means not solved), number of positions so far, number of correct solutions so far, search depth at which the solution was found, time to solution, nodes to solution, score, and PV.

The final line, which reads "162/183 - 7.9 1.19 867464", is a result summary. It means that 162 of 183 positions were solved, that the average depth to solution was 7.9 plies, the average time to solution was 1.19 seconds, and the average nodes to solution was 867464.

Tord