Random node spikes

[gotogo]

has anyone noticed the random node spikes in the latest stockfish?
[/img]

[gladius]

has anyone noticed the random node spikes in the latest stockfish?

What do you mean by node spikes? Does that position cause the same issue when you run again? How about with a clear hash?

Thanks.

[gotogo]

What do you mean by node spikes?
I mean it happens every now and then. could happen any move and any position, it is a spike it just happens. I have many more images I saved after I noticed it. Nodes could be as high as 48000 plus to 15000 plus when it spikes.

How about with a clear hash?
I ran the engine for the first time and seen the spikes in the second game then I looked for them so I could document.
I figure someone could figure out what is going on I am just showing what I noticed.
if it is a bug it could be a very good bug if Stockfish team learns how to stabilize it. I was wondering if the buggy Houdini versions had this problem before Robert made Houdini stable?

[Joerg Oster]

Looks like a hash hit.
Stockfish finds the whole PV in its Transposition Table, and since memory access is very fast, you see those high numbers of nodes/second.

[zullil]

has anyone noticed the random node spikes in the latest stockfish?
[/img]

GUI problem? The numbers in the image don't make much sense to me.

Here's what a healthy Stockfish log (ie, no GUI) looks like, on a 2-core laptop:

Code: Select all

Searching: 2rr2k1/pb3pb1/2n1qnpp/1pp1p3/8/PPBPPN2/3NBPPP/1QR2RK1 w - - 0 18
infinite: 0 ponder: 0 time: 0 increment: 0 moves to go: 0
 1   +0.08   00:00     112  Ne4 Nxe4 dxe4 
 2   +0.08   00:00     868  Ne4 Nxe4 dxe4 
 3   +0.08   00:00    1291  Ne4 Nxe4 dxe4 
 4   +0.24   00:00    1753  Ne4 Nxe4 dxe4 a6 
 5   +0.38   00:00    3295  Ne4 Nxe4 dxe4 a6 Rfd1 
 6   +0.30   00:00   10381  Ne4 Qe7 Rfd1 b4 axb4 cxb4 
 7   +0.42   00:00   22470  Qb2 Re8 Ne4 Qe7 b4 Nxe4 dxe4 
 8   +0.40   00:00   32318  Ne4 Qe7 Qb2 Nxe4 dxe4 b4 axb4 cxb4 
 9   +0.44   00:00   42461  Ne4 Qe7 Qb2 Nxe4 dxe4 b4 axb4 cxb4 Bd2 Na5 Qa2 Rxc1 
                            Rxc1 
10   +0.44   00:00   46480  Ne4 Qe7 Qb2 Nxe4 dxe4 b4 axb4 cxb4 Bd2 Na5 Qa2 Rxc1 
                            Rxc1 
11   +0.38   00:00   68997  Ne4 Qe7 Qb2 Nxe4 dxe4 b4 axb4 cxb4 Bd2 Bf6 Bc4 Na5 
12   +0.51   00:00  135450  Ne4 Qe7 Bb2 Nxe4 dxe4 a6 a4 bxa4 bxa4 Bf6 Rfd1 Rxd1+ 
                            Rxd1 Rd8 Rxd8+ Nxd8 
13   +0.38   00:00  270274  Ne4 Bf8 Bd2 b4 axb4 cxb4 Nc5 Bxc5 Rxc5 e4 dxe4 Nxe4 
                            Bc4 
14   +0.30   00:00  499802  Ne4 Bf8 Bd2 b4 axb4 cxb4 Nxf6+ Qxf6 e4 Qe7 Rfe1 a5 
                            Qa2 f6 Be3 
15   +0.30   00:00  848791  Ne4 Bf8 Bd2 b4 axb4 cxb4 Nxf6+ Qxf6 Rfe1 a5 e4 Qe7 
                            Be3 f6 Rc2 h5 Rec1 
16   +0.22   00:03   3899K  Rfd1 a6 Ne4 Nd7 b4 cxb4 axb4 Bf8 Qb2 Be7 Bd2 f5 Nc5 
                            Nxc5 bxc5 Qd5 e4 fxe4 dxe4 Qxe4 Bxh6 Rxd1+ Bxd1 
17   +0.22   00:03   4566K  Rfd1 a6 Ne4 Nd7 b4 cxb4 axb4 Bf8 Qb2 Be7 Bd2 f5 Nc5 
                            Nxc5 bxc5 Qd5 e4 fxe4 dxe4 Qxe4 Bxh6 Rxd1+ Bxd1 
18   +0.02   00:06   7837K  Rfd1 a6 Ne4 Nd7 b4 Bf8 Qb2 cxb4 axb4 Be7 Bd2 f5 Nc5 
                            Nxc5 bxc5 Rd5 e4 Rxc5 Bxh6 Rxc1 Rxc1 f4 
19   +0.10   00:06   8806K  Rfd1 a6 Ne4 Nd7 b4 cxb4 axb4 Bf8 Qb2 f5 Nc5 Bxc5 
                            bxc5 Nxc5 Bxe5 Nxe5 Nxe5 Nd7 f4 Nxe5 fxe5 Rxc1 Rxc1 
                            Rc8 Rxc8+ Bxc8 
20   +0.14   00:07  10324K  Rfd1 a6 Ne4 Nd7 b4 cxb4 axb4 Bf8 Qb2 f5 Nc5 Bxc5 
                            bxc5 Nxc5 Bxe5 Nxe5 Nxe5 Nd7 f4 Nxe5 fxe5 Rxc1 Rxc1 
                            Rc8 Rxc8+ Bxc8 Bf3 
21   +0.20   00:34  46097K  Ne4 Qe7 Nxf6+ Bxf6 Nd2 Bg7 Rfe1 f5 b4 cxb4 axb4 Bf6 
                            Nf3 Qe6 Qb2 a6 Qa3 Qe7 Qb2 Qe6 
22   +0.20   01:06  88428K  Ne4 Bf8 Bd2 Nd7 b4 cxb4 axb4 a6 Bd1 Qe7 Nc5 Nxc5 
                            bxc5 e4 dxe4 Ne5 Ba5 Re8 Nxe5 Qxe5 Qa2 Bxc5 Bb3 Qxe4 
                            Bxf7+ Kh7 
23   +0.08   01:33 122698K  Ne4 Bf8 Rfd1 Nd7 b4 cxb4 axb4 a6 Qb2 f5 Nc5 Bxc5 
                            bxc5 Nxc5 Bxe5 Nxe5 Nxe5 Nd7 Nxd7 Qxd7 Qf6 Qd6 Qxd6 
                            Rxd6 Rxc8+ Bxc8 d4 Bb7 
24   +0.00   02:08 165805K  Ne4 Bf8 Rfd1 Nd7 b4 cxb4 axb4 a6 Qb2 f5 Nc5 Bxc5 
                            bxc5 Nxc5 Bxe5 Nxe5 Nxe5 Nd7 Rxc8 Rxc8 Nxd7 Qxd7 e4 
                            Kf7 Qd2 Kg7 Qb2+ Kf7 
25   +0.00   02:33 194432K  Ne4 Bf8 Rfd1 Nd7 b4 cxb4 axb4 a6 Qb2 Be7 Re1 Qd5 Bd2 
                            Rc7 Nc3 Qd6 Ne4 Qd5 
Nodes: 194432096
Nodes/second: 1266749
Best move: Ne4
Ponder move: Bf8

[bob]

Looks like a hash hit.
Stockfish finds the whole PV in its Transposition Table, and since memory access is very fast, you see those high numbers of nodes/second.

That generally REDUCES the NPS, as the HT doesn't count nodes for the tree below the hit.

[Sven]

Looks like a hash hit.
Stockfish finds the whole PV in its Transposition Table, and since memory access is very fast, you see those high numbers of nodes/second.

That generally REDUCES the NPS, as the HT doesn't count nodes for the tree below the hit.

Why is that a reason for reducing NPS? Searching N nodes with NPS1 takes N/NPS1 seconds. If you save one subtree of size H through a hash hit then you search (N-H) nodes with average speed NPS1 and 1 node with a neglible effort since it returns quickly with a hash hit. That should take around (N-H)/NPS1 seconds for N-H+1 nodes, so roughly same speed (but never slower).

Sven

[bob]

Looks like a hash hit.
Stockfish finds the whole PV in its Transposition Table, and since memory access is very fast, you see those high numbers of nodes/second.

That generally REDUCES the NPS, as the HT doesn't count nodes for the tree below the hit.

Why is that a reason for reducing NPS? Searching N nodes with NPS1 takes N/NPS1 seconds. If you save one subtree of size H through a hash hit then you search (N-H) nodes with average speed NPS1 and 1 node with a neglible effort since it returns quickly with a hash hit. That should take around (N-H)/NPS1 seconds for N-H+1 nodes, so roughly same speed (but never slower).

Sven

You get a fast NPS by avoiding useless work. If you do a move generation and get a hash probe cutoff on the first move searched, that's a lot of wasted time... It depends on how the engine is written, of course.