Leto wrote:If anyone's interested here are the results I got with my 12 core Xeon and my 4 thread HP laptop, using the 2017-5-22 asmfishw:
Intel Xeon X5650:
Total time: 169889 ms
Nodes Searched: 3593687704
Nodes/Second: 21153151
HP Pavilion Intel i3-3130M:
Total time: 507486 ms
Nodes Searched: 195599135
Nodes/Second: 3854291
How many threads did you use on the Xeon? I did not see instruction on the benchmarking site about setting Threads = physical cores. If you have hyper-threading enabled, you can likely get a higher NPS using perhaps 16 threads. Of course, what this means is open to interpretation, but people seem to like seeing big numbers.
Laskos wrote:Even if 0.955 is quite imprecise and varying across the engines, the function itself is tame. Milos result was 2.4 with 0.955 efficiency comparing Ryzen to that 224-threaded Xeon. With much higher 0.975 efficiency, fitted to what you say you got with Houdini, this factor becomes 3.4, pretty far from your factor of 6 you calculated. And I think that Amdahl's law is the one to be applied to parallel search with alpha-beta.
Chess engine alpha-beta is a lot more complex than the tasks for which Amdahl's law usually is applied.
I see no fundamental reason why the coefficient to use in Amdahl's formula could not depend on the number of threads.
For example:
- 0.955 for 8 threads.
- 0.975 for 24 threads
- 0.990 for 80 threads
This, of course, would imply that Amdahl's law is not a very good model of multi-threaded chess engines.
I will end my contribution to this thread by once again saying that only well-controlled tests with high number of threads will allow us to move beyond the current, rather idle speculations .
I pute the table with Lazy Texel doublings in a more readable way, and added effective speedup per doubling for the number of cores and Amdahl's efficiency if assumed to be a law here. The numbers I added are not perfect, because doubling time line is not equal in base strength to doubling cores line (only the first row), but quite close.
It's not Amdahl by fair shot, it's not even your formula. As Peter showed, the smaller speedup for 32 vs 16 cores of 1.62 is due mainly to NUMA=2 configuration. The scaling seems to improve with cores, pretty crazy, close to Robert Hyatt crazy formula for effective speedup of 1 + 0.7 * (n_cores-1).
Maybe something else is in effect here.
Perhaps there is a fixed startup cost that is payed in full by the first engine that starts. So there is some baseline term that we should subtract.
Amdahl's law is basically pure math. It seems to me that if it is not obeyed what that really means is there is something in the model that we have not accounted for.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Leto wrote:If anyone's interested here are the results I got with my 12 core Xeon and my 4 thread HP laptop, using the 2017-5-22 asmfishw:
Intel Xeon X5650:
Total time: 169889 ms
Nodes Searched: 3593687704
Nodes/Second: 21153151
HP Pavilion Intel i3-3130M:
Total time: 507486 ms
Nodes Searched: 195599135
Nodes/Second: 3854291
How many threads did you use on the Xeon? I did not see instruction on the benchmarking site about setting Threads = physical cores. If you have hyper-threading enabled, you can likely get a higher NPS using perhaps 16 threads. Of course, what this means is open to interpretation, but people seem to like seeing big numbers.
Laskos wrote:It's not Amdahl by fair shot, it's not even your formula. As Peter showed, the smaller speedup for 32 vs 16 cores of 1.62 is due mainly to NUMA=2 configuration. The scaling seems to improve with cores, pretty crazy, close to Robert Hyatt crazy formula for effective speedup of 1 + 0.7 * (n_cores-1).
Some new information, I'm currently running Houdini 6 on a dual 20-core Xeon server in a match of 20 threads against 40 hyper-threads.
So basically you have (20 threads at 1x speed) versus (40 threads at 0.65x speed).
After 900 games, the 40 threads engine leads by about 10 Elo points.
The result is very similar to the 12 vs 24 threads match I reported above.
The "naive" formula - implying that every doubling of the number of threads produces about 20% of overhead - seems to hold quite well.
If I find the time, I'll also run a 40 threads vs 80 hyper-threads match.
Laskos wrote:It's not Amdahl by fair shot, it's not even your formula. As Peter showed, the smaller speedup for 32 vs 16 cores of 1.62 is due mainly to NUMA=2 configuration. The scaling seems to improve with cores, pretty crazy, close to Robert Hyatt crazy formula for effective speedup of 1 + 0.7 * (n_cores-1).
Some new information, I'm currently running Houdini 6 on a dual 20-core Xeon server in a match of 20 threads against 40 hyper-threads.
So basically you have (20 threads at 1x speed) versus (40 threads at 0.65x speed).
After 900 games, the 40 threads engine leads by about 10 Elo points.
The result is very similar to the 12 vs 24 threads match I reported above.
The "naive" formula - implying that every doubling of the number of threads produces about 20% of overhead - seems to hold quite well.
If I find the time, I'll also run a 40 threads vs 80 hyper-threads match.
I don't know how this Lazy stuff with alpha-beta can scale better than Monte Carlo search of AlphaGo (although on cluster), but it seems something like your 1.75^(doublings) is not naive anymore. Then there are NUMA and cluster issues. In that thread Peter studied them too.
Houdini wrote:I've stopped the match after 1200 games, final score is (+196 -149 =855) or about (+13±10 Elo) in favor of the 40 hyper-threads engine.
I'll start the 40 vs 80 threads match, it will take some time .
If the threads are hyperthreads, are we sure of what we are measuring?
IOW, is the power of a hyperthread core (when physical cores are already exhausted) equal to the power of a core when the physical core count has not been exhausted?
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Houdini wrote:I've stopped the match after 1200 games, final score is (+196 -149 =855) or about (+13±10 Elo) in favor of the 40 hyper-threads engine.
I'll start the 40 vs 80 threads match, it will take some time .
If the threads are hyperthreads, are we sure of what we are measuring?
IOW, is the power of a hyperthread core (when physical cores are already exhausted) equal to the power of a core when the physical core count has not been exhausted?
Hyper-threads are slower; when you double the number of threads you only get about 30% total node speed increase.
That's why I call it a match between (20 threads at 1x speed) and (40 threads at 0.65x speed).
.
