Cluster chess : not only Grid but also Achilles

[Marc Lacrosse]

Djordje Vidanovic once told us about a cluster chess machine developped at Nis university in Serbia but there was not much echo here at that time.

This "engine" has recently beaten a 2600+ GM in a four slow games match by 3.5-0.5.

There is only very little information on this at Achilles home page (http://chess.elfak.ni.ac.yu/). Here is almost all what is told in english :

Code: Select all

System Achilles:
Parallel chess system, FreeStyle version (3.1). Based on Axon 3 engine.

Hardware:
12 AMD X64 1.8Ghz cluster. 
Interface AMD Dual Core 2.2Ghz machine, secondary synchronization/server machine with AMD Dual Core 2.6GhZ (provided by AMD Adriatic).

Software:
Achilles (c) by Vladan Vuckovic.
Axon (c) by Vladan Vuckovic.
UDP cluster chess OS (c) by Vladan Vuckovic.
Internet Position Stream Generator (c) by Vladan Vuckovic.

Opening books by Djordje Vidanovic.

Does anyone know anything more ?

It looks quite interesting IMHO.

Marc

[M ANSARI]

I always thought that chess is a perfect application for multi computer setup. For example it is a well known fact that engines will perform worse if they have to load EGTB's and search an endgame for a tablebase win ... ofcourse this is offset by the fact that EGTB's will win or avoid loss in many games ... thus it evens out with a slight favour to having the EGTB's. Now if you could have another system passively analyzing the output of the first system ... you could turn off EGTB's for the first system and thus have it play stronger. If an EGTB hit occurs then the passive system would reach a certain predetermined threshold and become active and relay the winning or drawing line. This is a simple 2 computer cluster that will work and play better for any chess engine (in theory).

Now with big clusters ... things get a little more complicated. I think in the Achilles and Gridchess ... they are active clusters which share memory and simply try to brute force the search deeper by spreading out the search tasks. I am not sure that is the best way as you can get a lot of latencies and somehow you might reach a setup where you need to throw a massive amount of hardware for minute gain.

I personally believe that a successful cluster should operate in a passive fashion in service of a master computer. Different settings of engine parameters on the many passive computers would study a different set of possible lines. This in fact would mimick how a Freestyle chess player checks moves out if he has 2 or 3 computers at his disposal.

[matejst]

I am going to Nis next week. I will try to see for myself, and to send you more informations. I will send you an email in french, if you wish.

Kind regards.

BS

[gerold]

I am going to Nis next week. I will try to see for myself, and to send you more informations. I will send you an email in french, if you wish.

Kind regards.

BS

Would be interested in this also. I request you post your infor. on
the forum if you get more info.

Thanks.

Gerold.