Koivisto issues about koivisto

[Luecx]

We still cannot open those images. you can just attach them to your message as normal images.
If you are on windows you can use: Win+Shift+S to take a screenshot which you can very simply paste into your text with Ctrl+V.

An example:

[Gargantua1981]

[Gargantua1981]

It worked this time around.

[Luecx]

With "It worked this time around" you mean the cpu usage went up to 100%? I tried to reproduce and checked it on my Ryzen 3950x and my cpu usage was constantly at 100%. I am organising some machine with higher core counts so i can reproduce. Once I have done this, we can compare the results. very interesting :)

[Gargantua1981]

No I meant it worked with the attachments. I suppose to have engine use 100% or 95% of its CPU some adjustments to the RAM must be made. But unfortunately I not have a clue how to do so. I am of that opinion that it is best not to experiment much trying to fix the probelm on your own, if one not know how to do so..

[Luecx]

my ram is also throtteled to 2.33 Mhz instead of 3.6. I dont think the RAM plays an important role here although could be possible with those high core counts. Once I have a machine I can verify.

[Luecx]

Do other engines like Stockfish go up to 100%?

[Gargantua1981]

Yes engines like Stockfish, ShashChess, Raubfisch go by 100%.
Komodo has same problem as Koivisto, that it only uses about 63-64% of its CPU

[rainydew]

CB based GUI may cause some performance loss. How about the CPU usage when using Arena or Cutechess?

[ChrisWhittington]

GUIs are definitely responsible for performance loss, but this might not be that. Each thread (I don’t know how Koivisto data is stored at that level) has a memory footprint, and the overall footprint will be 64 x that. At the point this blows through L3 cache there will be some memory accesses that get slow. I don’t know how the windows processor usage app works out processor usage, but if it thinks CPU is inactive waiting for RAM, then that would show on the graphic. I guess it’s possible SF or some others are engineered with low enough memory footprint.

…

Sent from my iPhone

On 22 Mar 2022, at 17:22, Rainy Chan ***@***.***> wrote:  CB based GUI may cause some performance loss. How about the CPU usage when using Arena or Cutechess? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.

[Luecx]

@ChrisWhittington Hey Chris, You might be right. I do basics of programming but I am not that familiar with caching. I did observe that Stockfish or similar engines allocate a lot more memory, especially for history tables, than koivisto. I am not sure how one would measure the memory footprint beside just checking how much memory is being allocated per thread. Could you give me a hint on how to investigate this?

Greetings,
Finn

[ChrisWhittington]

Well, my knowledge on this is not expert, i prefer logic programming to worrying about hardware specifics, but I guess a lazy way would be fire up Koivisto with various n_thread settings and take a look (while it’s thinking) at the memory usage in the windows app the OP used. That way gives a rough idea of RAM being grabbed per thread. That’s not quite the same as RAM being often read/write, but I guess it is at least an upper bound. IIRC my big AMD 64x (same as OP?) has L3 cache of 256Mb, and that gets shared by all threads (I think only L1 and L2 cache are per cpu). At 64 threads that’s 4Mb per thread and it’s pretty easy to use that up with some tables and quite possibly a 1024 wide accumulator for each search ply.

…

Sent from my iPhone

On 23 Mar 2022, at 12:11, Luecx ***@***.***> wrote:  @ChrisWhittington Hey Chris, You might be right. I do basics of programming but I am not that familiar with caching. I did observe that Stockfish or similar engines allocate a lot more memory, especially for history tables, than koivisto. I am not sure how one would measure the memory footprint beside just checking how much memory is being allocated per thread. Could you give me a hint on how to investigate this? Greetings, Finn — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.

[Luecx]

I have been talking with other people about this too. Have you used tablebases? It could be that the TB code blocks other threads using a mutex. Incase you are using syzygy-tablebases, try disabling them and check again :)

[Gargantua1981]

I am using some of the 7 men set tbs, about 7-8 TB. If there was a slowdown in TB code than all engines would be impacted. But so far only Koivisto and Komodo have been using reduced CPU. Koivisto does not support syzygy tablebases so I doubt that will help solve the reduced CPU performance.

[Luecx]

huh? Koivisto does support the 6-men TB. Also this could cause the slowdown. It may be that Koivisto accesses the TB too often which in return causes the slowdowns.

[Gargantua1981]

Well in Infinity GUI once I try and use Koivisto there is no tbs path. In Infinity GUI one must manually add tbs path to the engine.
Koivisto does not have this option available. So basically it is not using tbs.

[Gargantua1981]

Koivisto has extreme time management that even with +2 sec increment it runs out of time.

Here is the game

PS: if possible can Koivisto team please add Move Overhead option ( the same option Stockfish has) in its parameters so one can increase it so engine not uses too much of its time and play faster ?

In the current game I have lost on time after 40 moves or so in a draw game. The game was 16+2 game on Infinity chess server.

[Luecx]

Added MoverOverhead.
Also added a select which lets you select what type of overhead you specified. it can be PerMove which indicates that every move will be searches for x less seconds. Alternatively PerGame will lead to the engine having x seconds left on the clock at the end of the game.