Drops a scalability bottleneck due to memory contention
of a single shared table across threads. The effect starts
to be sensible with a high number of threads. Specifically
we have a small regression with 7 threads both at 60 and
180 seconds TC:
10000 @ 60+0.6 th 7
ELO: -2.46 +-3.2 (95%) LOS: 6.5%
Total: 9896 W: 1037 L: 1107 D: 7752
5000 @ 180+0.6 th 7
ELO: -1.95 +-4.1 (95%) LOS: 17.7%
Total: 5000 W: 444 L: 472 D: 4084
We have a regression because counterMoveHistory table is
quite big and it takes time for a single thread to fill it.
Sharing the table yields to a higher fill rate and better
quality of moves and up to 7 threads the benefits of sharing
more then compensate the loss in speed due to contention.
Interestingly even with a 3X longer TC, so with more time
for the single thread to catch up, the improvment is quite
limited and below noise level. It seems we really need much
longer TC to saturate the table.
When we move to high threads number it's another story:
5000 @ 60+0.6 th 22
ELO: 3.49 +-4.3 (95%) LOS: 94.6%
Total: 4880 W: 490 L: 441 D: 3949
2000 @ 60+0.6 th 32
ELO: 8.34 +-6.9 (95%) LOS: 99.1%
Total: 2000 W: 229 L: 181 D: 1590
As expected the speed-up more than compensates the filling
rate, and we expect that with tournament TC, where single
thread is able to saturate the table, the difference will
be even stronger. For instance for TCEC 9 super-final time
control will be 180 minutes + 15 seconds and this scalability
improvement seems definitely the way to go.
So, summarizing:
GOOD:
Measured big improvement in high core scenario
Suitable for TCEC 9 superfinal (big hardware, very long TC)
Consistent and natural patch that extends to counterMoveHistory
what we already do for remaining history tables, that are all per-thread
Non functional change for the common case of a single core
Very simple (just 6 lines modified, no added ones)
BAD:
Small regression (within 2-3 ELO) with few threads and short TC
bench: 5341477
Currently root moves are copied to all teh threads
but are DTZ filtered only in main thread at the
beginning of teh search.
This patch moves the TB filtering before the
copy of root moves fixing issue #679https://github.com/official-stockfish/Stockfish/issues/679
No bench change.
And passed in do_move(), this ensures maximum efficiency and
speed and at the same time unlimited move numbers.
The draw back is that to handle Position init we need to
reserve a StateInfo inside Position itself and use at
init time and when copying from another Position.
After lazy SMP we don't need anymore this gimmick and we can
get rid of this special case and always pass an external
StateInfo to Position object.
Also rewritten and simplified Position constructors.
Verified it does not regress with a 3 threads SMP test:
ELO: -0.00 +-12.7 (95%) LOS: 50.0%
Total: 1000 W: 173 L: 173 D: 654
No functional change.
Make it explicit that those variables are not globals, but
are used only by main thread. I think it is a sensible
clarification because easy move is already tricky enough
and current patch makes the involved actors explicit.
No functional change.
Resolves#537
Instead of creating a running std::thread and
returning, wait in Thread c'tor that the native
thread of execution goes to sleep in idle_loop().
In this way we can simplify how search is started,
because when main thread is idle we are sure also
all other threads will be idle, in any case, even
at thread creation and startup.
After lazy smp went in, we can simpify and rewrite
a lot of logic that is now no more needed. This is
hopefully the final big cleanup.
Tested for no regression at 5+0.1 with 3 threads:
LLR: 2.95 (-2.94,2.94) [-5.00,0.00]
Total: 17411 W: 3248 L: 3198 D: 10965
No functional change.
Now that we don't have anymore TimerThread, there is
no need of this long class hierarchy.
Also assorted reformatting while there.
To verify no regression, passed at STC with 7 threads:
LLR: 2.97 (-2.94,2.94) [-5.00,0.00]
Total: 30990 W: 4945 L: 4942 D: 21103
No functional change.
Unfortunately std::condition_variable::wait_for()
is not accurate in general case and the timer thread
can wake up also after tens or even hundreds of
millisecs after time has elapsded. CPU load, process
priorities, number of concurrent threads, even from
other processes, will have effect upon it.
Even official documentation says: "This function may
block for longer than timeout_duration due to scheduling
or resource contention delays."
So retire timer and use a polling scheme based on a
local thread counter that counts search() calls and
a small trick to keep polling frequency constant,
independently from the number of threads.
Tested for no regression at very fast TC 2+0.05 th 7:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 32969 W: 6720 L: 6620 D: 19629
TC 2+0.05 th 1:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 7765 W: 1917 L: 1765 D: 4083
And at STC TC, both single thread
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 15587 W: 3036 L: 2905 D: 9646
And with 7 threads
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 8149 W: 1367 L: 1227 D: 5555
bench: 8639247
The only interesting change is the moving of
stack[MAX_PLY+4] back to its original position
in id_loop (now renamed Thread::search).
No functional change.
Rely on well defined behaviour for message passing, instead of volatile. Three
versions have been tested, to make sure this wouldn't cause a slowdown on any
platform.
v1: Sequentially consistent atomics
No mesurable regression, despite the extra memory barriers on x86. Even with 15
threads and extreme time pressure, both acting as a magnifying glass:
threads=15, tc=2+0.02
ELO: 2.59 +-3.4 (95%) LOS: 93.3%
Total: 18132 W: 4113 L: 3978 D: 10041
threads=7, tc=2+0.02
ELO: -1.64 +-3.6 (95%) LOS: 18.8%
Total: 16914 W: 4053 L: 4133 D: 8728
v2: Acquire/Release semantics
This version generates no extra barriers for x86 (on the hot path). As expected,
no regression either, under the same conditions:
threads=15, tc=2+0.02
ELO: 2.85 +-3.3 (95%) LOS: 95.4%
Total: 19661 W: 4640 L: 4479 D: 10542
threads=7, tc=2+0.02
ELO: 0.23 +-3.5 (95%) LOS: 55.1%
Total: 18108 W: 4326 L: 4314 D: 9468
As suggested by Joona, another test at LTC:
threads=15, tc=20+0.05
ELO: 0.64 +-2.6 (95%) LOS: 68.3%
Total: 20000 W: 3053 L: 3016 D: 13931
v3: Final version: SeqCst/Relaxed
threads=15, tc=10+0.1
ELO: 0.87 +-3.9 (95%) LOS: 67.1%
Total: 9541 W: 1478 L: 1454 D: 6609
Resolves#474
Start all threads searching on root position and
use only the shared TT table as synching scheme.
It seems this scheme scales better than YBWC for
high number of threads.
Verified for nor regression at STC 3 threads
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 40232 W: 6908 L: 7130 D: 26194
Verified for nor regression at LTC 3 threads
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 28186 W: 3908 L: 3798 D: 20480
Verified for nor regression at STC 7 threads
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 3607 W: 674 L: 526 D: 2407
Verified for nor regression at LTC 7 threads
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 4235 W: 671 L: 528 D: 3036
Tested with fixed games at LTC with 20 threads
ELO: 44.75 +-7.6 (95%) LOS: 100.0%
Total: 2069 W: 407 L: 142 D: 1520
Tested with fixed games at XLTC (120secs) with 20 threads
ELO: 28.01 +-6.7 (95%) LOS: 100.0%
Total: 2275 W: 349 L: 166 D: 1760
Original patch of mbootsector, with additional work
from Ivan Ivec (log formula), Joerg Oster (id loop
simplification) and Marco Costalba (assorted formatting
and rework).
Bench: 8116244
To sync UI with main thread it is enough a single
condition variable because here we have a single
producer / single consumer design pattern.
Two condition variables are strictly needed just for
many producers / many consumers case.
Note that this is possible because now we don't send to
sleep idle threads anymore while searching, so that now
only UI can wake up the main thread and we can use the
same ConditionVariable for both threads.
The natural consequence is to retire wait_for_think_finished()
and move all the logic under MainThread class, yielding the
rename of teh function to join()
No functional change.
Idea and original implementation by Stephane Nicolet
7 threads 15+0.05
ELO: 3.54 +-2.9 (95%) LOS: 99.2%
Total: 17971 W: 2976 L: 2793 D: 12202
There is no functional change in single thread mode
Use Mutex instead.
This is in preparaation for merging with master branch,
where we stilll don't have spinlocks.
Eventually spinlocks will be readded in some future
patch, once c++11 has been merged.
No functional change.
And use mutex instead. You may never want to do this.
It is a workaround to run c++11 on fishtest where many
machiens have HTenabled and this can be a problem when
number of cores set is higher than number of physical cores.
To disable spinlocks, just compile with -DNO_SPINLOCK flag
No functional change.
Calling lock.test_and_set() in a tight loop creates expensive
memory synchronizations among processors and penalize other
running threads. So syncronize only only once at the beginning
with fetch_sub() and then loop on a simple load() that puts much
less pressure on the system.
Reported about 2-3% speed up on various systems.
Patch by Ronald de Man.
No functional change.
It is reported to be defenitly faster with increasing
number of threads, we go from a +3.5% with 4 threads
to a +15% with 16 threads.
The only drawback is that now when testing with more
threads than physical available cores, the speed slows
down to a crawl. This is expected and was similar at what
we had setting the old sleepingThreads to false.
No functional change.
Initialization is more complex than what I'd like due
to MSVC compatibility that for some reason does not like:
std::atomic_flag lock = ATOMIC_FLAG_INIT;
No functional change.
It should be used slavesMask.count() instead.
Verified 100% equivalent when sp->allSlavesSearching:
dbg_hit_on(sp->allSlavesSearching, sp->slavesCount != sp->slavesMask.count());
No functional change.
And retire a redundant field. This is important also
from a concept point of view becuase we want to keep
SMP structures as simple as possible with the only
strictly necessary data.
Verified with
dbg_hit_on(sp->spLevel != level)
that the values are 100% the same out of more 50K samples.
No functional change.
Balance threads between split points.
There are huge differences between different machines and autopurging makes it very difficult to measure the improvement in fishtest, but the following was recorded for 16 threads at 15+0.05:
For Bravone (1000 games): 0 ELO
For Glinscott (1000 games): +20 ELO
For bKingUs (1000 games): +50 ELO
For fastGM (1500 games): +50 ELO
The change was regression for no one, and a big improvement for some, so it should be fine to commit it.
Also for 8 threads at 15+0.05 we measured a statistically significant improvement:
ELO: 6.19 +-3.9 (95%) LOS: 99.9%
Total: 10325 W: 1824 L: 1640 D: 6861
Finally it was verified that there was no (significant) regression for
4 threads:
ELO: 0.09 +-2.8 (95%) LOS: 52.4%
Total: 19908 W: 3422 L: 3417 D: 13069
2 threads:
ELO: 0.38 +-3.0 (95%) LOS: 60.0%
Total: 19044 W: 3480 L: 3459 D: 12105
1 thread:
ELO: -1.27 +-2.1 (95%) LOS: 12.3%
Total: 40000 W: 7829 L: 7975 D: 24196
Resolves#258
Import C++11 branch from:
https://github.com/mcostalba/Stockfish/tree/c++11
The version imported is teh last one as of today:
6670e93e50
Branch is fully equivalent with master but syzygy
tablebases that are missing (but will be added with
next commit).
bench: 8080602
This gives SF accurate PVs, such that the evaluation of the leaf node in
the PV matches the score backed up to the root (99% of the time.
q-search will use the value stored in the hash table instead of the eval
value sometimes).
One drawback is that fail-high/low only get a minimal 2 move PV.
It doesn't add any additional overhead to the non-PV codepath except an
extra eight bytes to the SearchStack structure in multi-threaded
searches.
A core part of this is not pruning based on TT score in PV nodes. This
was measured as not being a regression at multiple TCs, except for one
exception, fast TC with huge hash, which is not realistic for longer
searches.
STC - 1 thread, 128 mb hash
ELO: 1.42 +-3.1 (95%) LOS: 81.9%
Total: 20000 W: 4078 L: 3996 D: 11926
STC - 3 thread, 128 mb hash
ELO: -3.60 +-2.9 (95%) LOS: 0.8%
Total: 20000 W: 3575 L: 3782 D: 12643
STC - 3 thread, 8 mb hash
ELO: 0.12 +-2.9 (95%) LOS: 53.3%
Total: 20000 W: 3654 L: 3647 D: 12699
LTC - 3 thread, 32mb hash
ELO: 2.29 +-2.0 (95%) LOS: 98.8%
Total: 35740 W: 5618 L: 5382 D: 24740
Bench: 6984058
Resolves#102
- Currently broken
- Never been really useful
- Does not work well with new splitting model
Verified for no regression at STC with 3 threads:
LLR: 2.96 (-2.94,2.94) [-6.00,0.00]
Total: 81905 W: 12122 L: 12381 D: 57402
No functional change
After last Joona's patch there is no measurable
difference between the option set or unset.
Tested by Andreas Strangmüller with 16 threads
on his Dual Opteron 6376.
After 5000 games at 15+0.05 the result is:
1 Stockfish_14050822_T16_on : 3003 5000 (+849,=3396,-755), 50.9 %
2 Stockfish_14050822_T16_off : 2997 5000 (+755,=3396,-849), 49.1 %
bench: 880215
