Optimal parameters for Komodo 9

[titanD]

I am trying to optimize Komodo 9 for "4 minutes per move" play.

Currently I've only loaded it into Arena and am playing around with it.

I'm using a 3930k cpu with 16 GB ram.

It's set to use all 12 logical cores.

Is there any more I need to do? I'm using the default 128 MB hash size.

Is this too big/small or close enough to whatever's optimal that it wouldn't make much difference?

And is there a problem suite that gives some information about the time it took another engine/hardware to solve the problems so I can compare?

[titanD]

I don't understand how hash utilization is reporting.

Even if I set 4096 MB hash size, when I do an infinite analysis, it eventually gets up to 97%

Ok no problem.

But if I reset the board and start over, it's still at 97%

How exactly is this happening? I would think hash utilization would grow the longer it analyzed.. and then carry over to subsequent moves.

Why would it still be using a bunch of the hash size even when it's completely out of the tree it was in when it generated the table?

[zullil]

I am trying to optimize Komodo 9 for "4 minutes per move" play.

Currently I've only loaded it into Arena and am playing around with it.

I'm using a 3930k cpu with 16 GB ram.

It's set to use all 12 logical cores.

Is there any more I need to do? I'm using the default 128 MB hash size.

Is this too big/small or close enough to whatever's optimal that it wouldn't make much difference?

And is there a problem suite that gives some information about the time it took another engine/hardware to solve the problems so I can compare?

You need to read the following files that likely came with your copy of Komodo:

Code: Select all

komodo-settings.txt 
README.html  
setHash.txt 

In particular, the Komodo programmer recommends that you set Threads = 6.

[titanD]

Two 15 second analyses with 12 cores (hyperthreading) = 152 million nodes searched average

Two with 6 cores (no hyperthreading) = 95 million nodes searched average

Do you think you could speak more to their suggestion since it seems to hurt performance? They didn't elaborate in the txt doc other than to say they recommended you don't use hyperthreading

[zullil]

Two 15 second analyses with 12 cores (hyperthreading) = 152 million nodes searched average

Two with 6 cores (no hyperthreading) = 95 million nodes searched average

Do you think you could speak more to their suggestion since it seems to hurt performance? They didn't elaborate in the txt doc other than to say they recommended you don't use hyperthreading

Rather than looking at nodes searched, see if using 12 cores gets you to a fixed depth faster than 6 cores does. I'll let the Komodo team elaborate.

[Zenmastur]

I don't understand how hash utilization is reporting.

Even if I set 4096 MB hash size, when I do an infinite analysis, it eventually gets up to 97%

Ok no problem.

But if I reset the board and start over, it's still at 97%

How exactly is this happening? I would think hash utilization would grow the longer it analyzed.. and then carry over to subsequent moves.

Why would it still be using a bunch of the hash size even when it's completely out of the tree it was in when it generated the table?

If your getting 10M nps and analysis time is 240 seconds the program can use around 24Gb of cache. If you can give it 24Gb fine, if not give it as much as you have available. I wouldn't go out and buy extra memory for it unless you do a lot of analysis because the difference between 12Gb and 24Gb isn't that great. I wouldn't worry too much about how full the TT is. If it's empty it worthless to the program. If it's full it means it contains the maximum data it can hold so the program is getting the maximum benefit from it.

Regards,

Zen

[lkaufman]

I am trying to optimize Komodo 9 for "4 minutes per move" play.

Currently I've only loaded it into Arena and am playing around with it.

I'm using a 3930k cpu with 16 GB ram.

It's set to use all 12 logical cores.

Is there any more I need to do? I'm using the default 128 MB hash size.

Is this too big/small or close enough to whatever's optimal that it wouldn't make much difference?

And is there a problem suite that gives some information about the time it took another engine/hardware to solve the problems so I can compare?

For four minutes per move and many cores the default Hash is way too small. Probably you should use 12 GB on your system, certainly at least 6. Regarding 6 vs 12 threads you cannot go by nodes per second or even by time to depth for Komodo; you would have to run a match between two different machines for at least a thousand games to prove anything. But my opinion is that you will get best results by turning off hyperthreading and using just six threads. If you need to keep hyperthreading on for any reason, it might well be better to set threads to 12. If you don't make use of hyperthreading by setting threads to 12 instead of 6, having hyperthreading on is pretty clearly going to weaken performance.

[titanD]

If you don't make use of hyperthreading by setting threads to 12 instead of 6, having hyperthreading on is pretty clearly going to weaken performance.

I'm not sure I understand what you mean here.

How would I "not make use of hyperthreading by setting threads to 12 instead of 6"?

[lkaufman]

If you don't make use of hyperthreading by setting threads to 12 instead of 6, having hyperthreading on is pretty clearly going to weaken performance.

I'm not sure I understand what you mean here.

How would I "not make use of hyperthreading by setting threads to 12 instead of 6"?

I meant that if you are setting threads to 6 as I recommend, hyperthreading on should perform worse than hyperthreading off (the old saying applies, "use it or lose it"). The real question is whether using 6 with hyperthreading off or 12 with hyperthreading on is better. We believe that 6 with hyperthreading off is the better choice, but we don't have the setup to prove this.

[shrapnel]

We believe that 6 with hyperthreading off is the better choice, but we don't have the setup to prove this.

This is correct. HT On may or may not help SF, but it certainly does NOT help Komodo.
On my 8-Core i7 5960X I get greater Depths faster with 8 Threads than while using all 16 threads. More wins too.
In fact, I have disabled HT in BIOS itseslf; this gives me faster speeds too, with more stable overclocks.