Chess Forum

massimilianogoi wrote:About:
viewtopic.php?p=2633#p2633

A very good forum rule idea!

In this way,
Many of the members/readers may gain more and more..

Btw, actually what I can add as important details,
Which will be great to be mentioned as well e.g:

- Used Chess GUI ?
- Ponder ON or OFF ?
- Opening Book/Suite ?
- Hyper-Threading ON or OFF ?
- Number of Concurrent Games ?

Sure, there are many other important settings, such as
Large Pages are Enabled or Disabled, because these
Settings may play as serious influence over results too!
And will be not bad to be informed e.g AVX2 or BMI2 etc.

And why Hyper Threading ON or OFF, plus Concurrent Games etc
Can play as very important role, influence over played results ?
E.g recently I ran some tests on my older tour machine:
2x E5-2686 36 Cores/72 Threads, where the tests were with
Hyper Threading ON, 1 Thread per engine, with Concurrent: 50
That means the matches were with close to 70% CPU USAGE
And I've got strange NNUE results than expected...at least
I can say about 10% lower draw values (via Balsa openings)
So as conclusion (why 10% lower draws..) due to simply lower
Processor speed..in other words, on this machine, when HT ON:
It is safer be played with up to 36 Concurrent Games, 1 Thread...
Going higher Is not recommended...actually normal and we should
Not surprised, since Hyper-Threading Threads are not real CORES...
So..all testers have to be careful if running under similar cond. !

And about my current tour machine 2x EPYC 7B12 128 Cores/256 Threads
For better performance, it's safer Hyper Threading (SMT) to be Disabled
That's why almost all SCCT's produced games are played with HT OFF,
Otherwise via HT ON+ many Concurrent Games: 128 games simultaneously
Then some of the Top engines seemed to be not so stable...time forfeits etc.
Sure not very frequent..but sometimes happens with some of the engines...

One thing more,
Why opening book/suite is very important to be mentioned too ?
As we know..depending on which book..the results may be different...
Usually with weak suites are required thousands of games (per player)
Just in case not bad if min 5.000 + games even 10.000+ games better...
Otherwise the ERROR margin is expecting to be very high... in short,
It is something like...your machine has 2 GB Ram, but you prefer to
Run/test the engines 1 GB hash (per engine) etc.. do you know what will
Happen?..just in case with close in strength engines...then simply will be
As lottery.. lottery.. to see regarding which one is strongest...

Ok... hope helps ...and that's all for now )


Best,
Sedat

Good ideas. Thanks for the contribution.

But be aware that not all the engines have options related to large pages.

About that lower draw ratio with 50 processes on a 36 hyperthreaded core, I'm pretty sure that the most of them are engine "errors" (jumps from ~|0.xy| to |1.ab|,
where:

x <= 2
y,a,b = from 0 to 9) especially on short time control.


With all those descriptions mandatory just few people would publish tests, mostly because they don't have the knowledge about these settings/will to write them down.

You can already see that the guy I asked for them didn't answer anymore, so let's stay to the essential.

Thank you... we are better together !)

About your comment... if Large Pages is enabled,
Normally I never suggest LP ON for running tours..
Yes..LP ON can be useful for Eng analyzing,
But not so good idea for running long tours...!

On other hand, let's say we run tours with LP ON:
And let's say that X chess engine is NOT capable to use LP,
Where the other opponent is capable to use LP ...

And after running long-hours matches, such as 24 hours..
Then there is a possibility the hardware speed to be down..
And just imagine how much the test cond. will be fair...?!
Besides, the quality of played games may be low/down too...

For these reasons,
It safer LP to be disabled for Tour organizers...

Greetings )

massimilianogoi wrote: ↑Wed May 29, 2024 5:41 pm
About that lower draw ratio with 50 processes on a 36 hyperthreaded core, I'm pretty sure that the most of them are engine "errors" (jumps from ~|0.xy| to |1.ab|,

As I stated earlier.. in one of my previous postings:
My old machine (2x E5-2686) has 72 threads (not 36 threads, but 36 Cores)
You may know, logical threads and physical cores are two different things..!
Where I tested the engines with close to 70% CPU usage.
Sure I refer for the matches when HT is enabled...

In other words, on my old machine too:
It is safer the HT to be disabled...and then
Even with 80%+ CPU usage.. no any problem issue..

But if running with HT ON (on 2x E5-2686), then
It is safer to be played with less than 50% CPU usage...

So as conclusion,
Sure it all depends.. but if playing with 50%+ logical threads:
1st. The engines do not gain...even I can say they suffer...
2nd our produced games may be poor, not good quality etc


Greetings

Sedat Canbaz wrote: ↑Wed May 29, 2024 8:38 pm

massimilianogoi wrote: ↑Wed May 29, 2024 5:41 pm
About that lower draw ratio with 50 processes on a 36 hyperthreaded core, I'm pretty sure that the most of them are engine "errors" (jumps from ~|0.xy| to |1.ab|,

As I stated earlier.. in one of my previous postings:
My old machine (2x E5-2686) has 72 threads (not 36 threads, but 36 Cores)
You may know, logical threads and physical cores are two different things..!

I know they are two different things. The past years I had the feeling that using the hyperthreading for chess engines was kinda like overloading it, thus engines more subjected to errors. I don't know why this happens, I just know it does.

Many other factors contribute to move choice mistakes: short time per game, an uncorrect hashtable size (too short or too big with short timed games), other concurrent processes, badly programmed engine, et cetera.

Sedat Canbaz wrote: ↑Wed May 29, 2024 8:38 pm...
...
So as conclusion,
Sure it all depends.. but if playing with 50%+ logical threads:
1st. The engines do not gain...even I can say they suffer...
2nd our produced games may be poor, not good quality etc

One can compare games made this way with sweets made in factories: tons of them but cheap instead of a hand crafted cake made by a pastry chef - whole different things.

In my opinion there's nothing better in quality than games made in correspondence chess where players use computers.