Lazy SMP >4 Thread Slowdown

[nitrocan]

I have recently implemented lazy smp and it works remarkably well with up to 4 cores. One weird thing that I did notice though is that as soon as I start using 8/16 cores, the nps actually starts decreasing instead of increasing. Same with the depth that it searches per unit time. Any ideas on why this might be happening? Could it be that too many threads are contending for the transposition table? All ideas are very welcome, thanks!

[smatovic]

...in the case of AMD Threadripper/Epyc as cpu,

do you use all 4 memory channels?

http://talkchess.com/forum/viewtopic.ph ... 00&t=64980

Code: Select all

...Threadripper needs to be in quad channel to move from around 18'000 kn/s to over 30'000, it is was it is. 

Not sure about this, but, maybe each 4 core unit of Threadripper has to be treated as NUMA node...

https://www.anandtech.com/show/11551/am ... analysis/2

http://talkchess.com/forum/viewtopic.ph ... tion+table

--
Srdja

[abulmo2]

Good explanation here of what may happen:

https://www.youtube.com/watch?v=WDIkqP4JbkE

[nitrocan]

...in the case of AMD Threadripper/Epyc as cpu,

do you use all 4 memory channels?

http://talkchess.com/forum/viewtopic.ph ... 00&t=64980

Code: Select all

...Threadripper needs to be in quad channel to move from around 18'000 kn/s to over 30'000, it is was it is. 

Not sure about this, but, maybe each 4 core unit of Threadripper has to be treated as NUMA node...

https://www.anandtech.com/show/11551/am ... analysis/2

http://talkchess.com/forum/viewtopic.ph ... tion+table

--
Srdja

I'm not actually running it on my own machine at all, I use AWS instances for testing. The particular instance works perfectly fine with stockfish (nps increases almost linearly with number of threads) but not my engine haha.

Good explanation here of what may happen:

https://www.youtube.com/watch?v=WDIkqP4JbkE

Will take a look thanks!

[syzygy]

Shared cache lines. Use per-thread node counters, etc.

Are you using locked instructions or heavily accessed atomic variables?

On Linux, use "perf c2c" to find the cache lines that are shared, either because of a real shared variable or because of falsely shared variables (thread-specific ones that happen to sit in the same cache line).

[Dann Corbit]

I have recently implemented lazy smp and it works remarkably well with up to 4 cores. One weird thing that I did notice though is that as soon as I start using 8/16 cores, the nps actually starts decreasing instead of increasing. Same with the depth that it searches per unit time. Any ideas on why this might be happening? Could it be that too many threads are contending for the transposition table? All ideas are very welcome, thanks!

Perhaps there is a shared variable that is a bottleneck.

E.g. an eval hash (if global) might be better as a thread local storage so that each thread gets its own.

Besides the main hash table, what else is a public object in your program?

[nitrocan]

I have recently implemented lazy smp and it works remarkably well with up to 4 cores. One weird thing that I did notice though is that as soon as I start using 8/16 cores, the nps actually starts decreasing instead of increasing. Same with the depth that it searches per unit time. Any ideas on why this might be happening? Could it be that too many threads are contending for the transposition table? All ideas are very welcome, thanks!

Perhaps there is a shared variable that is a bottleneck.

E.g. an eval hash (if global) might be better as a thread local storage so that each thread gets its own.

Besides the main hash table, what else is a public object in your program?

That was pretty much exactly it. After making:

-Node counter
-Killer moves
-Counter moves
-History

thread specific, the scalability of threads is now working as intended. I'm sure there are other things I should look into as well but so far so good! Thanks everyone!

If anyone's interested, here's the pull request that I have made that addresses this issue:

https://github.com/nitrocan/sctr/pull/24/files