Core behaviour

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Rebel
Posts: 5364
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Wed Jul 26, 2017 9:36 pm

Further CPU scheduling tests can be followed here.

User avatar
Rebel
Posts: 5364
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Thu Jul 27, 2017 6:45 am

Rebel wrote:Further CPU scheduling tests can be followed here.
In addition I changed the number of games to 16,000 and the TC to 40/5+0.1.

Dani (and others), how do you usually test, time control and number of games?

http://rebel13.nl/misc/scheduling.html

User avatar
cdani
Posts: 2168
Joined: Sat Jan 18, 2014 9:24 am
Location: Andorra
Contact:

Re: Core behaviour

Post by cdani » Thu Jul 27, 2017 10:31 am

Rebel wrote: Dani (and others), how do you usually test, time control and number of games?
My usual testing is at 15+0.04 for stc sometimes less, and ltc ranges from 30+0.04 to 80+0.05.

The number of games varies depending on how well is going, so similar to using SPRT but as the test runs on 5 computers at the same time, I caculate when to stop roughly by eye. Sometimes I just resume the test if I'm not convinced. Also when I'm trying to see if a patch scales well by time, I tend to do more games to be more sure that is not just luck. Most tests are between 15000-35000 games.

User avatar
Rebel
Posts: 5364
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Fri Jul 28, 2017 6:45 am

cdani wrote: I suppose that any high performance software will profit at least a little from the cpu scheduling, isn't it?
I have measured it for ProDeo, 5% faster in nodes.

Code: Select all

Affinity     16:05:57 (119.360M nodes) NPS = 2.059K
Normal       16:09:42 (113.690M nodes) NPS = 1.954K

Depth Stats   MIDG   END0   END1   END2
Affinity     10.72  11.29  12.18  15.22
Normal       10.63  11.21  12.12  15.12
The speed gain is not my primary goal for doing this but to be complete I will test this too pitching the 64-bit versions of SF, Komodo and Andscacs against each other with and without affinity, see the web-page. Will take one week to complete the in total 9 runs.

User avatar
Rebel
Posts: 5364
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Sat Jul 29, 2017 6:23 am

An amazing result of 53.9% (+27 elo) [Andscacs+affinity] vs [Andscacs without affinity]. See: http://rebel13.nl/misc/scheduling.html

Comments c.q. issues to discuss.
1. I wonder if the setup of the match is fair in the sense the [Andscacs+affinity] version is perhaps hurting the performance of the [Andscacs without affinity] version.

2. 40/5+0.1 is ultra fast and perhaps favors the [Andscacs + affinity] looking at the depths comparison.

Anyway, reason enough to test

[SF4+affinity] vs [SF4 without affinity]
[Komodo 5 + affinity] vs [Komodo 5 without affinity]

first.

And see if they produce similar results.

Post Reply