Odd behavior of the latest stockfish

[Dann Corbit]

I can't figure out how to embed an image, so here is an image of memory in use by SF:
https://drive.google.com/file/d/1909dhO ... sp=sharing
I allocated 32 GB RAM (as is seen by the COMMIT SIZE) but only a couple GB is in use after a couple of minutes.
This is very different from the old SF, which would have filled the 32 GB in the same amount of time.
I would think that the exponential expansion of the tree would demand more hash usage.
Does anyone know why the hash usage climbs so slowly?
Also, the NPS is about half. I guess it is the "only NNUE" eval and larger node collection.
Do others see the same thing?

[Dann Corbit]

I am getting an average of 35,338,489 in analyzing positions from analyzing the positions of:

[pgn][Event "Superbet Rapid Poland"] [Site "Warsaw POL"] [Date "2023.05.21"] [EventDate "2023.05.21"] [Round "1"] [Result "1-0"] [White "R Wojtaszek"] [Black "M Carlsen"] [ECO "A40"] [WhiteElo "2670"] [BlackElo "2853"] [PlyCount "76"] 1. d4 b5 2. e4 Bb7 3. Bd3 Nf6 4. Nd2 c5 5. c3 cxd4 6. cxd4 e6 7. Ngf3 Nc6 8. O-O a6 9. Re1 Be7 10. d5 exd5 11. e5 Nh5 12. Nb3 g6 13. Bh6 Rg8 14. Be3 Ng7 15. Nc5 Qc7 16. Rc1 Ne6 17. Nxb7 Qxb7 18. Bf1 Bb4 19. Re2 Ne7 20. Rec2 Nf5 21. Bg5 Be7 22. Bf6 Nfg7 23. Qd3 Bxf6 24. exf6 Nh5 25. Qc3 Rd8 26. a3 Qb8 27. g3 g5 28. Re1 g4 29. Nd4 Rg5 30. Bd3 Re5 31. Nf5 Rxe1+ 32. Qxe1 Nxf6 33. Qb4 d6 34. Qc3 Nd7 35. Qh8+ Ndf8 36. Qf6 Ng6 37. Re2 Qc7 38. Ng7+ 1-0 [/pgn]
Which is ironically a reversed Polish {Orangutan} game played in Poland against a Polish person.

[Ajedrecista]

Hello Dann:

I can't figure out how to embed an image, so here is an image of memory in use by SF:
https://drive.google.com/file/d/1909dhO ... sp=sharing
I allocated 32 GB RAM (as is seen by the COMMIT SIZE) but only a couple GB is in use after a couple of minutes.
This is very different from the old SF, which would have filled the 32 GB in the same amount of time.
I would think that the exponential expansion of the tree would demand more hash usage.
Does anyone know why the hash usage climbs so slowly?
Also, the NPS is about half. I guess it is the "only NNUE" eval and larger node collection.
Do others see the same thing?

Maybe a recent patch or an unlucky run? OTOH, you can embed images with IMG tags:

Code: Select all

[img]http://image_url[/img]

This time did not work with your URL, so I had to drag the image to a new tab in Google Chrome, where I got the following URL:

Code: Select all

https://lh3.googleusercontent.com/drive-viewer/AITFw-xPmqMREgFcYEFfqwtLPySYAWOt_0c3oYGeVOiTRuz6Yfu-BUZDyT5AXvEi6Y-KcW57RgPLI2eIjTyBYr0KI-JBukfd4Q=w1280-h913

You can even resize the image in the forum just toying with the end of the URL (w1280-h913).

The image will dissapear from the forum (or show an error message) once the URL of the image is down, of course.

Good luck with the solution of the problem!

Regards from Spain.

Ajedrecista.

[syzygy]

I can't figure out how to embed an image, so here is an image of memory in use by SF:
https://drive.google.com/file/d/1909dhO ... sp=sharing
I allocated 32 GB RAM (as is seen by the COMMIT SIZE) but only a couple GB is in use after a couple of minutes.

There should be no need to look at an image of memory, since you can look at hashfull.

This is very different from the old SF, which would have filled the 32 GB in the same amount of time.
I would think that the exponential expansion of the tree would demand more hash usage.
Does anyone know why the hash usage climbs so slowly?
Also, the NPS is about half. I guess it is the "only NNUE" eval and larger node collection.

You already gave at least part of the answer. If NPS is half, hash usage will grow at half the speed.

Note that the tree does not grow exponentially, at least not exponentially in time. Hash usage, as measured by hashfull or by your method, will grow linearly at first and then drop off (filling the last few % will take a lot longer than the first few % for reasons that are hopefully obvious).

Have you tried comparing the growth of hashfull vs total nodes for the old and new versions?

[Dann Corbit]

Have you tried comparing the growth of hashfull vs total nodes for the old and new versions?

No, that is an excellent suggestion. Thanks for the idea.