Houdini with a six point lead near the halfway point of TCEC
Posted: Tue Nov 28, 2017 4:10 pm
Chessdom Write Up:
Statement by team Komodo
"During Game 1 of the Superfinal Anton Mihailov, TCEC’s tournament director, emailed us noting significantly lower nodes per second from the version of Komodo we had used in Stage 2, 1959.00. Anton checked TCEC’s server for correct settings and found nothing amiss there.
We began our investigation by making Windows versions of many of the promotions between version 1959.00 and 1970.00 (the Superfinal version), using 23 threads on a 24 core machine. In this testing we found a roughly 8% drop in speed when running 1960.00 versus 1959.00, and similar speed drops in all the versions tested up through 1970.00. We examined the code change between 1959.00 and 1960.00 which was a minor addition to some backward pawn score (two or three lines of code).
Finding nothing in the code that could explain a significant slowdown, we recompiled the source code turning off PGO (profile-guided optimization). This did not help. Then we recompiled turning off LTO (Link Time Optimization). This resulted in restoration of the lost speed. The compiler we used throughout was MinGW 4.7.3.
It is not completely clear what exactly happened between Stage 2 version 1959.00 and 1960.00. There was no evidence of this speed reduction in 1959.00 or Komodo 11.2.2, our latest public release. We speculate that the speed reduction happens once some size of compiled code is exceeded.
It is important to point out that the approximately 8% speed reduction we noted on our best hardware (24 cores) is apparently as high as 23% on TCEC’s 44-core machine based on Komodo’s relative nodes per second vs. Houdini in Stage 2. A difference of this magnitude is what Anton noted in his original email signalling that there might be a problem.
In summary, there is indeed a slowdown in the version now running in TCEC, which appears to be due to a compiler bug. We believe that the compiler bug is more significant than any program changes since 11.2.2.
We accept full responsibility for not discovering the problem before the start of the final. Although the bug has probably cost us some points it probably does not fully explain the current five point score deficit."
http://www.chessdom.com/houdini-with-a- ... t-of-tcec/
Statement by team Komodo
"During Game 1 of the Superfinal Anton Mihailov, TCEC’s tournament director, emailed us noting significantly lower nodes per second from the version of Komodo we had used in Stage 2, 1959.00. Anton checked TCEC’s server for correct settings and found nothing amiss there.
We began our investigation by making Windows versions of many of the promotions between version 1959.00 and 1970.00 (the Superfinal version), using 23 threads on a 24 core machine. In this testing we found a roughly 8% drop in speed when running 1960.00 versus 1959.00, and similar speed drops in all the versions tested up through 1970.00. We examined the code change between 1959.00 and 1960.00 which was a minor addition to some backward pawn score (two or three lines of code).
Finding nothing in the code that could explain a significant slowdown, we recompiled the source code turning off PGO (profile-guided optimization). This did not help. Then we recompiled turning off LTO (Link Time Optimization). This resulted in restoration of the lost speed. The compiler we used throughout was MinGW 4.7.3.
It is not completely clear what exactly happened between Stage 2 version 1959.00 and 1960.00. There was no evidence of this speed reduction in 1959.00 or Komodo 11.2.2, our latest public release. We speculate that the speed reduction happens once some size of compiled code is exceeded.
It is important to point out that the approximately 8% speed reduction we noted on our best hardware (24 cores) is apparently as high as 23% on TCEC’s 44-core machine based on Komodo’s relative nodes per second vs. Houdini in Stage 2. A difference of this magnitude is what Anton noted in his original email signalling that there might be a problem.
In summary, there is indeed a slowdown in the version now running in TCEC, which appears to be due to a compiler bug. We believe that the compiler bug is more significant than any program changes since 11.2.2.
We accept full responsibility for not discovering the problem before the start of the final. Although the bug has probably cost us some points it probably does not fully explain the current five point score deficit."
http://www.chessdom.com/houdini-with-a- ... t-of-tcec/