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
cdani
Posts: 2133
Joined: Sat Jan 18, 2014 9:24 am
Location: Andorra
Contact:

Re: Core behaviour

Post by cdani » Sat Jul 22, 2017 5:12 pm

Rebel wrote:Nobody cares?

Or do I miss something?
Thanks for your investigations!
As I don't feel that I have a problem to solve, for the moment I keep doing like always, but this is noted for if something related seems to happen. Is like what Bob said of testing with Linux, very useful in some context.

User avatar
Rebel
Posts: 5253
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Mon Jul 24, 2017 7:25 pm

cdani wrote:
Rebel wrote:Nobody cares?

Or do I miss something?
Thanks for your investigations!
As I don't feel that I have a problem to solve, for the moment I keep doing like always, but this is noted for if something related seems to happen. Is like what Bob said of testing with Linux, very useful in some context.
The developers of Linux have invested much more attention in core scheduling than Microsoft, if it's better remains the question, see the WIKI - Scheduling_(computing)

I have finished my first test with affinity and the results are dramatic. 8 cores, 8 matches, each match 'with its own core. I took a change good for +/-10 elo based on a previous testrun of 3 x 4800 games on 40/15, 40/30 and 40/60.

Code: Select all

CPU         Result    Perc   Time  ELO   LOS
Normal    2485|4800  51.8%  40/15  +12  99.9%
Affinity  2455|4800  51.1%  40/15   +8  98.0%
0.7% less, alright can happen with just 4800 games.

Code: Select all

CPU         Result    Perc   Time  ELO   LOS
Normal    2437|4800  50.8%  40/30   +5  92.2%
Affinity  2417|4800  50.4%  40/30   +3  74.7%
Another loss, still acceptable, but then the 40/60 run:

Code: Select all

CPU         Result    Perc   Time  ELO   LOS
Normal    2470|4800  51.5%  40/60  +10  99.7%
Affinity  2369|4800  49.4%  40/60   -4  11.1%
A loss of 14 elo ???

I don't understand what's going on, it looked so good in theory. Anyway, I am going to try 2 other programs in order to see if this is something universal chronic or ProDeo related.

To be continued.

Modern Times
Posts: 2554
Joined: Thu Jun 07, 2012 9:02 pm

Re: Core behaviour

Post by Modern Times » Mon Jul 24, 2017 8:18 pm

cdani wrote:
Rebel wrote:Nobody cares?

Or do I miss something?
Thanks for your investigations!
As I don't feel that I have a problem to solve, for the moment I keep doing like always, but this is noted for if something related seems to happen. Is like what Bob said of testing with Linux, very useful in some context.
Indeed, I don't feel that I have a problem to solve either. Even if core scheduling isn't quite optimal in Windows, it is better than the pain and hassle of running Linux. I like the O/S to help me in what I do, not to hinder me.
.

Opinions expressed here are my own, and not necessarily those of the CCRL Group.

bob
Posts: 20912
Joined: Mon Feb 27, 2006 6:30 pm
Location: Birmingham, AL

Re: Core behaviour

Post by bob » Tue Jul 25, 2017 5:06 am

Modern Times wrote:
cdani wrote:
Rebel wrote:Nobody cares?

Or do I miss something?
Thanks for your investigations!
As I don't feel that I have a problem to solve, for the moment I keep doing like always, but this is noted for if something related seems to happen. Is like what Bob said of testing with Linux, very useful in some context.
Indeed, I don't feel that I have a problem to solve either. Even if core scheduling isn't quite optimal in Windows, it is better than the pain and hassle of running Linux. I like the O/S to help me in what I do, not to hinder me.
How on earth does Linux "hinder" you? It prevents all the virus nonsense windows enjoys, avoids all the software bugs that windows saddles you with, etc. Installs far easier than windows nowadays also.

mar
Posts: 2122
Joined: Fri Nov 26, 2010 1:00 pm
Location: Czech Republic
Full name: Martin Sedlak

Re: Core behaviour

Post by mar » Tue Jul 25, 2017 7:43 am

bob wrote:It prevents all the virus nonsense windows enjoys, avoids all the software bugs that windows saddles you with, etc. Installs far easier than windows nowadays also.
Viruses prefer to taget certain OSes for obvious reasons, bugs happen everywhere.

Make Linux #1 desktop OS and you'll se the focus change of malware authors.

I like Linux but dislike faboys (smell complexes).

elpapa
Posts: 203
Joined: Sun Jan 18, 2009 10:27 pm
Location: Sweden
Full name: Patrik Karlsson

Re: Core behaviour

Post by elpapa » Tue Jul 25, 2017 11:55 am

bob wrote:Sorry, but not even close. And if you are talking windows 10, not at all.
What is it that makes Windows 10 worse than 7 or 8?

User avatar
Rebel
Posts: 5253
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Tue Jul 25, 2017 1:07 pm

Continuation...

I took 2 programs (SF4 and Andscacs) which as far as I know don't have special cpu scheduling code, such as Texel. I pitched SF4-64-bit against SF4-32-bit on my 8-core PC and AND-64-bit vs AND-32-bit on my other 4 core PC. First run at 40/15.

Code: Select all

Match [Andscacs 64bit vs Andscacs 32bit] [4 cores]
CPU       Games  Time   Perc   ELO  LOS   Depth Depth Diff
Normal     2400  40/15  63.9%  +99  100%  16.30 14.80 1.50
Affinity   2400  40/15  63.5%  +97  100%  16.48 15.03 1.45
Looks okay but what springs in mind is that both engines seem to profit from the affinity settings resulting in a higher average depth. I noticed the same with the three ProDeo runs. SF4 then:

Code: Select all

Match [Stochfish 4 64bit vs Stocfish 4 32bit] [8 cores]
CPU       Games  Time   Perc   ELO  LOS   Depth Depth Diff
Normal     4800  40/15  57.6%  +52  100%  17.34 16.61 0.73
Affinity   4800  40/15  58.6%  +60  100%  17.82 17.02 0.80
One full percent more but with (just) 4800 games that's not conclusive, a sign at best. But the increase of the average depth of both is remarkable, about half a ply more for the 64-bit version and +0.4 for the 32-bit version.

Now in progress both matches at 40/30, will take about 1.5 days.

User avatar
cdani
Posts: 2133
Joined: Sat Jan 18, 2014 9:24 am
Location: Andorra
Contact:

Re: Core behaviour

Post by cdani » Tue Jul 25, 2017 1:21 pm

Rebel wrote: I took 2 programs (SF4 and Andscacs) which as far as I know don't have special cpu scheduling code, such as Texel.
Thanks for the test. True, Andscacs does not use cpu scheduling code. It has an uci option WinGroupAffinity, but is disabled by default, and should be used only for more than 64 core machines that I know.

I suppose that any high performance software will profit at least a little from the cpu scheduling, isn't it?

jwes
Posts: 778
Joined: Sat Jul 01, 2006 5:11 am

Re: Core behaviour

Post by jwes » Tue Jul 25, 2017 2:16 pm

elpapa wrote:
bob wrote:Sorry, but not even close. And if you are talking windows 10, not at all.
What is it that makes Windows 10 worse than 7 or 8?
One problem is that it is difficult to stop Windows 10 from installing updates.

User avatar
Rebel
Posts: 5253
Joined: Thu Aug 18, 2011 10:04 am

Re: Core behaviour

Post by Rebel » Tue Jul 25, 2017 6:18 pm

cdani wrote: I suppose that any high performance software will profit at least a little from the cpu scheduling, isn't it?
Guess so. But what I am after is the accuracy of the matches we play. I am curious to find out.

Post Reply