Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Is more hash better? My tests say the opposite...

Author: Ross Boyd

Date: 19:41:51 08/31/03

Go up one level in this thread


On August 31, 2003 at 21:43:25, William Penn wrote:

>On August 31, 2003 at 21:18:58, Ross Boyd wrote:
>
>>On August 31, 2003 at 19:44:28, William Penn wrote:
>>
>>>The more hash I allocate, the slower the kN/s speed. Thus 4MB (the minimum) is
>>>the fastest in my tests, typically about 450kN/s. If I increase that to say
>>>256MB hash, the speed slows down to about 400kN/s. The more I increase hash, the
>>>slower the kN/s speed.
>>>
>>>The kN/s speed peaks, then eventually starts to decrease. How long this takes
>>>depends on the amount of hash. However in my tests, the long term speed
>>>advantage of bigger hash never catches up with the long term speed obtained with
>>>smaller hash. Thus I don't see any advantage whatsoever to using a hash table!
>>>The opposite seems to be true!?
>>>
>>>I'm using the Shredder7 GUI, Shredder 7.04 UCI engine, AMD XP Athlon 2400+/640MB
>>>RAM (608MB available). The GUI says the maximum I can allocate to hash is about
>>>455MB, so I'm not near the limit. Of course I'm using fairly common practical
>>>positions for these tests in Infinite Analysis mode, and the above indicated
>>>results are typical.
>>>
>>>I get very similar results running Shreddermarks with different size hash. The
>>>more hash, the lower the Shreddermark and corresponding kN/s.
>>>
>>>Now, will someone please refute this, or explain what I'm missing or
>>>overlooking? Thanks!
>>>WP
>>
>>Increasing the hash size will tend to lower the NPS in most engines.
>>
>>Its kind of hard to explain why this is so... but I'll try. When an engine gets
>>a hit in the hashtable it often cuts short the amount of exhaustive quiescence
>>searching where NPS tends to go high. Nearer the root there is generally more
>>overhead involved, with for example, more sophisticated move ordering etc...
>>whereas the move ordering at the QS tends to be cruder and hence faster.
>>
>>Anyway, its not a bad thing....
>>
>>What is more important is the total number of nodes visited to get to a certain
>>depth. You will see that increasing hash size will tend to reduce the tree...
>>and therefore (even though NPS drops slightly) the actual time taken to get to a
>>given depth is reduced (usually).
>>
>>Time how long Shredder takes to get to a given depth, and also the total nodes
>>visited, with various positions for two hash sizes.  You'll see the true benefit
>>of increasing the hash size.
>>
>>If you turn off the hash altogether you'll see the NPS increase a lot... but its
>>not going to play stronger that's for sure...
>>
>>So, NPS is not a measure of strength. Really, its only useful for comparison
>>purposes of the same engine with the same hash size on 2 different PCs.
>>
>>Hope this makes sense...
>>
>>Ross
>
>Gotcha!
>Thanks.

You're welcome.

>I needed that clarified.
>I have indeed noticed that I get about 1 ply deeper in a certain amount of time
>(say 1 hour) when using maximum hash on this box. I'll run some more tests and
>confirm it.



>
>Another important point to me is...
>
>Preferably the user should be able to access the best analysis "so far" in
>Infinite Analysis mode (or other modes), but that option apparently isn't
>available. Thus I must sit & wait & hope that it will deign to contact me with
>its internal results before I go to bed. It's unpredictable. Sometimes it takes
>a few minutes, sometimes several hours before the next analyis clip appears in
>the engine window. Is there something I'm overlooking? Is there any way to
>access the best analysis to date without stopping ongoing analysis in Infinite
>mode, or some way to make it spit out those clips more frequently? I'd like to
>see a clip about every 10-15 minutes, but at least one per hour at a minimum!?
>Thanks,
>WP

Don't know how to help you with this one as I don't own Shredder.
In Fritz 5.32 Correspondence Analysis you can adjust the branching factor to a
certain extent...

sorry.

Ross










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.