Torch v1 64-bit testing

[Modern Times]

timemargin is set to 2000. But J.Oster in german forum said, cutechess uses this timemargin only if cutechess is used with a fixed time per move level, not, when using a "regular" playing-level (like my 3min+1sec)...
So, J.Oster could be right...

Yes I think he is probably right. I always wondered how that works for different types of time controls.

[Rebel]

I haven't checked my own games just yet, I will do so later. But for my testing I limit concurrency to the core count, so it is hopefully less likely to be happening for me.

I restarted the testrun yesterday with Torch Move Overhead set to 500. Right now 2250 games played - all fine.
Result on wednesday/thursday.

Is it possible to get a daily update, no rating list, just a simple pgn statistic or so.

[RubiChess]

timemargin is set to 2000. But J.Oster in german forum said, cutechess uses this timemargin only if cutechess is used with a fixed time per move level, not, when using a "regular" playing-level (like my 3min+1sec)...
So, J.Oster could be right...

Yes I think he is probably right. I always wondered how that works for different types of time controls.

I'm pretty sure timemargin works for every timecontrol apart from "infinite". That's what I read from this code https://github.com/cutechess/cutechess/ ... r.cpp#L130

Why shouldn't it work for different time controls? For every time control and every move to make, you have a "deadline" when time forfeit should be declared. Add timemargin to this deadline and that's it.

Regards, Andreas

PS. I myself have solved some time forfeit problems in the past with some "slow" computer by using this timemargin parameter of cutechess. And I (almost) never use st=... time per move control.

[pohl4711]

I haven't checked my own games just yet, I will do so later. But for my testing I limit concurrency to the core count, so it is hopefully less likely to be happening for me.

I restarted the testrun yesterday with Torch Move Overhead set to 500. Right now 2250 games played - all fine.
Result on wednesday/thursday.

Is it possible to get a daily update, no rating list, just a simple pgn statistic or so.

Sorry, but my ratinglist is one of only two lists worldwide, testing Torch. I want to publish a complete testrun and a valid first test-result of Torch. I hope for your understanding. If all works as expected, the result will go online wednesday noon...

[AndrewGrant]

I restarted the testrun yesterday with Torch Move Overhead set to 500. Right now 2250 games played - all fine.
Result on wednesday/thursday.

Great news. What is your cutechess overstep margin setting ?

timemargin is set to 2000. But J.Oster in german forum said, cutechess uses this timemargin only if cutechess is used with a fixed time per move level, not, when using a "regular" playing-level (like my 3min+1sec)...

`timemargin=x` is applied to any type of condition, whether that is set time, fischer time, cyclic, etc. I'm guessing your timemargin is somehow not actually set to what you think, or its not being communicated to cutechess, or something to that effect.

Torch's default MoveOverhead is 100ms. OpenBench's default ( on master ) timemargin is =250. I've not seen a time loss in tens of millions of games. Furthermore, you stopped seeing time losses once setting Torch's overhead to 500ms. Both of those should be far eclipsed by a timemargin=2000, if it was indeed being used.

At least this is my thought on the matter. Note that the machines I run these games on, do nothing but play chess. Which is not realistic. People's home machines, especially laptops, are much more variable. At any moment a program might update itself, the OS might check a package list, etc etc. So margins must be higher.

Not that any of this is of particular importance.

[Joerg Oster]

I restarted the testrun yesterday with Torch Move Overhead set to 500. Right now 2250 games played - all fine.
Result on wednesday/thursday.

Great news. What is your cutechess overstep margin setting ?

timemargin is set to 2000. But J.Oster in german forum said, cutechess uses this timemargin only if cutechess is used with a fixed time per move level, not, when using a "regular" playing-level (like my 3min+1sec)...

`timemargin=x` is applied to any type of condition, whether that is set time, fischer time, cyclic, etc. I'm guessing your timemargin is somehow not actually set to what you think, or its not being communicated to cutechess, or something to that effect.

Torch's default MoveOverhead is 100ms. OpenBench's default ( on master ) timemargin is =250. I've not seen a time loss in tens of millions of games. Furthermore, you stopped seeing time losses once setting Torch's overhead to 500ms. Both of those should be far eclipsed by a timemargin=2000, if it was indeed being used.

At least this is my thought on the matter. Note that the machines I run these games on, do nothing but play chess. Which is not realistic. People's home machines, especially laptops, are much more variable. At any moment a program might update itself, the OS might check a package list, etc etc. So margins must be higher.

Not that any of this is of particular importance.

Thanks to you and A. Matthies for clarification.
At least, the cutechess readme could be more precise on this.

[pohl4711]

Since restart, my Torch-testrun is close to half-time (around 7000 games of 15000 played). No timelosses or any other problems right now.

[Modern Times]

Since restart, my Torch-testrun is close to half-time (around 7000 games of 15000 played). No time losses or any other problems right now.

It will be very interesting to see your results on Wednesday or thereabouts. I may be wrong but I think Torch might very close to Stockfish at standard chess at blitz on one thread.