Author: Vincent Diepeveen
Date: 12:36:08 08/29/05
Go up one level in this thread
On August 29, 2005 at 14:41:00, Marc Bourzutschky wrote: >On August 29, 2005 at 13:46:48, Vincent Diepeveen wrote: > >>On August 29, 2005 at 13:13:33, Marc Bourzutschky wrote: >> >>>On August 29, 2005 at 10:27:17, Vincent Diepeveen wrote: >>> >>>>On August 29, 2005 at 03:15:54, Marc Bourzutschky wrote: >>>> >>>>Hi Marc, >>>> >>>>I just don't understand how you manage to generate a mate in 250 with >>>>just 500GB i/o size for krbn krb. >>>> >>>>I need like 2 TB raid5 array or something for my generator for 7 men >>>>I'm using bitmaps which are similar to mate in n, though not entirely, >>>>as i just care for WDL. >>>> >>>>that means you basically need from -mate256 to +mate255 = 9 bits per >>>>entry. >>>> >>>>Diep's entry size is probably bigger with 358G entries than what is possible >>>>if you use nalimov's scheme, i'm using my own one which simply avoids more than >>>>1 piece at the same square, no illegal reductions are getting done. >>>> >>>>Anyway. 2 files of 358G * 9 bits per entry is more like 800GB than it is >>>>close to 500GB. >>>> >>>>How do you fit this EGTB onto just 500GB disk? >>>> >>> >>>As described in the post, we split krbnkrb into two components, depending on the >>>relative color of the bishops. This reduces the problem to two independent >>>generations with each half the size. In addition, we use on-the-fly >>>compression. We used less than 250 GB during generation. >> >>Hello Marc, >> >>Thanks for your answer! >> >>So you didn't split the bishops in 4 components: >> white bishop for KRBN side and white bishop for KRB >> white bishop for KRBN side and black bishop for KRB >> black bishop for KRBN side and white bishop for KRB >> black bishop for KRBN side and black bishop for KRB >> >>of course assuming you rotate 8 fold over the king positions and >>not the bishop (as there is less bishops in that case). >> > >No, since we restrict ourselves to the usual 462 symmetry equivalent king >positions for pawnless endings. The color of any individual bishop cannot be >fixed, because some of the symmetry operations to get the kings into one of >those 462 positions would change the color of the bishop. Only the relative >color of two bishops can be fixed, since that is invariant to all the symmetry >operations. Ah yes, i had forgotten this, how dumb of me! >>how many entries does your indexing scheme use for one of those >>files KRBNKRB? > >462*(64^5)/2 In diep format it's 462 * 62 * 61 * 60 * 59 * 58 = 358752350880 Versus your format is half of 462 * 64^5 = 496068722688 Without pawns it's easy to calculate it by hand. With pawns it's very hard to calculate it by hand as i must look it up in the table first because a king on a1 gives other pawnreduction than king on b2 :) So if i understand well, uncompressed the 4 files are nearly 1T entries, when you would use 1 byte per entry. Probably you had no other option than that from speed viewpoint as otherwise you can't see so many positions per second on average. I calculated your generator gets like 30 million positions per second, when also counting the lazy evaluated ones. This is very impressive! Note in the last generator i build (i built 3 generators over the time) it is using also a different format than the diep WDL egtb's is using, because working with 64 possibilities goes a hell of a lot faster, for the same reason. It's basically cpu speed that is needed. A small disk array of 8 harddisks and a 4ware raid card already gives at raid0 like 400MB per second which would support speeds of up to several hundreds of millions of positions per second. I assume you used raid0 for both harddisks? If not your generator speed might be quite some faster when adding some array speeds. Because a single IDE can get a throughput of at most like 50MB/s, which at 1 byte per entry would be 50 million positions. Note in diep generator format i use 1 bit per position at generation time. >>how many bytes (or bits) per entry did you use at generation time? > >Only a few bitmaps for a subset of the 462 king-king positions are in memory >during generation. 1 bitmap for 19 different king-king positions, plus two >additional bitmaps for a given king position. A general 7-man endgame thus >requires 21 * 64^5 / 8 = 2.8 GB of RAM. With permutation symmetries or bishop >parity tricks this can be reduced accordingly. > >We'll give more details in a technical paper we are working on. No need, really, your posts are really appreciated on this matter. Keep doing the good work! Oh, i hardly dare to ask as someone who has everything privateware himself, but is your resulting EGTB downloadable somewhere? Vincent >-Marc
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.