Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Beowulf hot spots shown pictorially

Author: Sune Fischer

Date: 03:03:53 06/21/02

Go up one level in this thread


On June 21, 2002 at 05:55:59, Sven Reichard wrote:

>On June 21, 2002 at 05:12:28, Sune Fischer wrote:
>
>>On June 21, 2002 at 04:01:41, Sven Reichard wrote:
>>
>>>Maybe I'm missing something here, but if you furthermore agree to use a hash
>>>table of size 2^i, i <= 32, you can replace all mods, divs and muls by a 32 bit
>>>'and'. The performance hit of the slightly smaller table (less than factor 2)
>>>should be outweighed by the faster access.
>>>
>>>just my 2 bits
>>
>>How do you know that?
>>If you use power 2 size then you get pretty big jumps; 64,128,256,512.
>>So you really have only 4 sizes of the hash. I think it is better to use 200 MB
>>with modulo than 128 MB with AND (there are not divs or muls), the save is not
>>_that_ much compared to what you save when getting those extra hashhits.
>>Of course I don't _know_, I'm just guessing :)
>>
>>Some claim the distribution of the keys are better if the number of entries is
>>prime, though I'm not convinced of that I can't rule it out.
>>
>>just my two øre.
>>-S.
>
>Sune,
>
>my comment was based on two pieces of information:
>
>1) It was said above that the modulus calculation was a _hot spot_, so I assume
>it is taking considerable time. Of course, like everybody else, I couldn't
>download the picture, so I don't know for sure how hot this spot really is.
>
>2) Some fairly old research papers state that the effect of doubling the table
>size is about 3% for small sizes, and about 1% for large sizes. This might be
>outdated (I don't have the exact reference, but it is from the early 90's), but
>I made a similar observation in my program. I don't rule out that this is due to
>the fact that my tables aren't too clever.
>
>Since I don't think that 1% is a considerable amount, this leads me to the
>conclusion that it might be beneficial to restrict the size to powers of two.
>
>My 5 agurot back to you. :)
>
>Sven.

Sven,

I know about the numbers.
I can't confirm if they hold water, I think it also depends on the duration of
the search, the replacement scheme and other variables.

I can tell you that the modulo operation on a integer is _faster_ than if you do
a manual implementation of x%n as x-n*(int)(x/n), I tested that long ago.
So I do not believe the modulo operation is a real hotspot, well maybe on 64 bit
numbers, but you do not have to use those once you realize that the number of
entries is less than 2^32.

-S.



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.