WCCC 2016 Live games

[Laskos]

Small correction. Komodo is using 48 cores of a 60 core server. Early testing suggested it runs faster than using more cores.

Faster meaning NPS?

[Jesse Gersenson]

Is Komodo going to play with Erdo's opening book again?

As far as I know, the book is from Erdo and Larry.

[bob]

Small correction. Komodo is using 48 cores of a 60 core server. Early testing suggested it runs faster than using more cores.

Faster meaning NPS?

Hopefully time to depth...

[Laskos]

Small correction. Komodo is using 48 cores of a 60 core server. Early testing suggested it runs faster than using more cores.

Faster meaning NPS?

Hopefully time to depth...

It would be pretty bad if even NPS on 60 cores is lower than on 48 cores.

[velmarin]

As in the other posts are the games, here the press Forum
Two images of the room.

[mjlef]

Small correction. Komodo is using 48 cores of a 60 core server. Early testing suggested it runs faster than using more cores.

Faster meaning NPS?

"Faster" is a combination of factors. Komodo does not reach the same speeds on a NUMA machine as you would expect from the same nuber of cores as a on a non-NUMA machine. Not having a lot of time to access this 4 NUMA node machine, in the very limited testing we could do, 48 cores worked best. Too many processors needing to access memory via the NUMA bus seems to slow it down. Also, 48+ cores accessing a single SSD could be an issue. Hopefully, I will be able to get access to this or similar machines when I have more time and can try to optimize Komodo for this type of machine. I have tried various thread locking to NUMA node schemes, but they have not yet let me use all the cores effectively. The OS just seems to know more about who is "busy" and moves the threads around better than Komodo can so far. Also, the OS does not seem to "lock" the thread to a NUMA node very quickly. The 2 NUMA node machines have a lot less NUMA issues. But Komodo eventually gets up to speed if you let the OS move things around a bit.

If anyone knows a programming library for linux and Windows that gives real time CPU usage information of threads/processes, let me know. I am hoping to make Komodo smarter. I am sure I will figure it out giving enough access and time. I wish memory speeds would increase with CPU speeds, but that has stalled it seems. Asking a user to manually choose where to put things is too complicated for a commercial product. I want to to be simple for the user.

For anyone interested, using 48 cores, Komodo gets about 50 million nodes per second on its first move/search. Jonny gets an amazing 1 billion nodes per second!

[Werewolf]

For anyone interested, using 48 cores, Komodo gets about 50 million nodes per second on its first move/search. Jonny gets an amazing 1 billion nodes per second!

5x faster than Deep Blue.

So on the Komodo machine, what type of Xeon setup is it? Quad socket?

[Vinvin]

For anyone interested, using 48 cores, Komodo gets about 50 million nodes per second on its first move/search. Jonny gets an amazing 1 billion nodes per second!

5x faster than Deep Blue.

So on the Komodo machine, what type of Xeon setup is it? Quad socket?

I think the speed peak for DeepBlueII was 1 billion nodes per second too.

[Jesse Gersenson]

Which type of raid array would solve the issue of 60 cores accessing to the ssd at the same time?

I think the speed peak for DeepBlueII was 1 billion nodes per second too.

That isn't a peak number for Jonny, which hit 2 billion nodes per second last year vs. Komodo.

[Jesse Gersenson]

Which type of raid array would solve the issue of 60 cores accessing to the ssd at the same time?

I think the speed peak for DeepBlueII was 1 billion nodes per second too.

That isn't a peak number for Jonny, which hit 2 billion nodes per second last year vs. Komodo.

(Perhaps someone at the tournament could pass this along -- Hyatt's name is misspelled here https://icga.leidenuniv.nl/?page_id=1676 )