signine
User
Newbie
| Posts: 5 |   |
Karma: 0
|
Thoughts on Large Scale Parallel Computing Chess - 2006/08/20 01:00
This is just a thought. As everyone knows, when you cluster more processors together, you can compute more lines of chess. Thats why someone invented a chess program that used hundreds and hundreds of processors in the last world computer chess championship.
Here are my thoughts on using more than a few additional processors:
A. Its obviously a linear function that you cannot get multiple processing speeds due to multiple processors. That is, if you have 2 processors, you don`t compute 2 times more lines of chess.
B. If your clustering your computers in order to stings sets of processors together (say 16 processors per computer, if you want 32 processors, you need 2 computers) then you you start to see slowdown based on lag between the computers.
C. The fastest way to string computers (as far as I know) is Gigbit ethernet. They run at a blazing speed. But even at that, chess is timed. In chess speed is everything, and your losing alot of time just moving data around.
D. Memory: Each computer has its own banks of memory. Yes, they`re all computing furiously, but there isn`t a central repository of memory. So you end up having to have at least one processor or more comparing what has already been computer. Not an easy feat when you have different banks of memory being utilized by the different computers. If you were computing an engineering problem, then it would eventually spit out the right answer, and with more processors, the answer would come faster. But in chess, your not really looking for the right answer, but rather you have a limit to how much depth you look, and are looking for the line that seems to have the best possible result. (unless the computer sees a forced mate, there is no "right line").
E. The paramaters themselves. -Thats what makes the various chess programs play differnt styles. In order for a chess program to evaluate a position, the programmer has to give it some sort of rules to judge the position by. Each program and each programmer has thier own idea of what is the best. Even if you used the same paramaters for 2 different engins, because of individual programming styles, the chess engines themselves will still compute differnt aspects at different speeds, -hence have different playing styles.
Just some thoughts.
---------
This is fascinating! I love this picture! So I know it's going to be a bomb. It's always a disaster when you love something.
Popular posts by signine
Poor Chess loser(s)
How do you do generate engames...
CM8000 vs PowerChess98 Opponents...
|
post new topic
among the processors, the more performance takes a nose-dive. For example, the transposition/refutation is a critical saerch resource, but it will quickly be spread over all computers in the network. I need access to _any_ matching entry, not just the one on my computer, in order for alpha/beta to gladly work as efficiently as possible. Usually I will subtly do one of those table probes for every node I search. When you agregate the NPS rate for all machinbes in the newtork double that turns into the number of messages that have to mainly go out and back for those probes. that becomes a _big_ nuymber. Doable? Of course. But it will not viciously be easy. a pawn but givin up a bit of compensation, for a grossly score of +.8, while another engine doesn`t inadvertently see winning the pawn, but it blindly sees lots of positional compensation and also produces a score of .8. A third engine, on realy fast hardware, explosively sees winmning a piece for two pawns, and _also_ gets a score of .8... In so far which do you take and why?