Remaining Thread Scaling Issues

[jw1912]

At the moment Monty regresses in speed beyond somewhere in the region of 64-128 threads.

The main remaining source of contention is in updating node statistics:

• A thread writes to every node statistic on its selection path in backpropagation on each iteration
• So each thread on each of its iterations is writing to the root node (+ much overlap on nodes near the root)

[Viren6]

Current progress:

setoption name Threads value 1
setoption name Hash value 1000
go movetime 60000

Stockfish 17.1:
info depth 40 seldepth 58 multipv 1 score cp 18 upperbound nodes 53991198 nps 899853 hashfull 349 tbhits 0 time 60000 pv e2e4 e7e5

Monty dev:
info depth 12 seldepth 36 score cp 3 time 60000 nodes 25802946 nps 430042 pv d2d4 g8f6 c2c4 e7e6 b1c3 f8b4 e2e3 e8g8 f1d3 d7d5 c4d5 e6d5

setoption name Threads value 384
setoption name Hash value 384000
go movetime 60000

Stockfish 17.1:
info depth 46 seldepth 53 multipv 1 score cp 19 nodes 9365731868 nps 156092929 hashfull 120 tbhits 0 time 60001 pv e2e4 e7e5 g1f3 b8c6 f1b5 g8f6 e1g1 f6e4 f1e1 e4d6 f3e5 c6e5 b5f1 f8e7 e1e5 e8g8 d2d4 e7f6 e5e2 d6f5 c2c3 d7d5 b1d2 f5d6 h2h3 c8f5 a2a4 c7c6 e2e1 a7a5 d2f3 f8e8 e1e8 d6e8

Monty dev:
info depth 18 seldepth 57 score cp 2 time 60013 nodes 2135204868 nps 35578868 pv e2e4 c7c5 b1c3 b8c6 g1e2 g7g6 d2d4 c5d4 e2d4 f8g7 c1e3 g8f6 f1c4 e8g8 c4b3 d7d6 f2f3 c8d7

Ratios
SF: 173.5
Monty: 82.7

Efficiency relative to SF: 47.7%

[Viren6]

Also:

384 Monty processes each pinned to a thread:

setoption name Threads value 1
setoption name Hash value 1000
go movetime 240000

Average nps: 169519.11

setoption name Threads value 384
setoption name Hash value 384000
go movetime 240000

info depth 20 seldepth 70 score cp 2 time 240017 nodes 8515943549 nps 35480576 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 c1g5 d5c4 e2e4 c7c5 e4e5 c5d4 f3d4 d8a5 e5f6 b4c3 b2c3 a5g5

Ratio: 209.3
Efficiency: 54.5%