CGOS Bot Operators Community
Posted: Fri Sep 01, 2017 2:33 am
Hello everyone,
as I finally managed to find some time to get a bot running on CGOS, I thought it might be nice to have a place to share information and tips on how to set up different bots with CGOS or provide hardware information.
For those, who don't about CGOS yet: It's a computer go server that allows free access for bots to compete against each other with 15min sudden death time limits on 19x19.
Main page: http://www.yss-aya.com/cgos/
Current standings: http://www.yss-aya.com/cgos/19x19/standings.html
BayesElo standings after at least 100 games: http://www.yss-aya.com/cgos/19x19/bayes.html
The main page gives a good overview, but if there are any questions, feel free to ask here.
Right now, personally I'm running the Rn4204_4790_980Ti bot. That's Rn 4.20.4 from https://github.com/zakki/Ray on an i7 4790k, GTX 980Ti and 16GB RAM on Windows 10, with the following command line configuration:
ray.exe --time 880 --size 19 --thread 8 --komi 7.5 --superko --tree-size 262144 --no-debug --no-early-pass
Pondering is disabled as there is a bug in 4.20.4 that leads to timeouts otherwise. I'm looking to add --reuse-subtree for the next matches. The tree size is already at max for the amount of RAM available. The time is set at 880 seconds to leave a bit of buffer for the 900 seconds time limit, but I assume this is actually overridden by GTP by the server.
Also, I'm not using the TCL script, but the Python client, as Rn doesn't completely obey the --no-debug flag and outputs some debug info on move generation when running in GPU mode, which confuses the client and terminates the session. I edited the Python client to filter out anything but the actual move response so I could run Rn with GPU.
As the CPU only has 4 real cores, I ran with 4 threads on some test sessions, but Rn would only utilize the CPU up to about 70%, and I had to bump the thread count up to 8 to get 100% utilization.
I'm looking forward to the next Leela release and giving that a spin on CGOS.
Some bot information
Rn4204_4790_980Ti
Bot: Rn 4.20.4 - https://github.com/zakki/Ray
Operator: Satorian
Hardware: i7 4790k, GTX 980Ti, 16GB RAM
OS: Windows 10
Params: ray.exe --time 880 --size 19 --thread 8 --komi 7.5 --superko --tree-size 262144 --no-debug --no-early-pass
LeelaLight 1Thread NoGPU 10.000 playouts ELO 2811
LeelaGPU 4Threads GPU (GTX 750Ti) 50.000 playouts ELO 3063 (+252)
LeelaMax 8Threads GPU (GTX 1070) 100.000 playouts ELO 3314 (+251) (+503)
Bot: Leela - https://sjeng.org/leela.html
Operator: kris
Hardware: CPU i7 2600K, 16GB RAM
Rn.4.20.4-1c1g - 3417 Elo
Bot: Rn 4.20.4 - https://github.com/zakki/Ray
Operator: cyan
Hardware: GTX 1060
Leela-0.10.5-16t1g
Bot: Leela 0.10.5 - Pre-release version - https://sjeng.org/leela.html
Operator: roy7
HardwareRyzen 1700 and a 1080Ti
as I finally managed to find some time to get a bot running on CGOS, I thought it might be nice to have a place to share information and tips on how to set up different bots with CGOS or provide hardware information.
For those, who don't about CGOS yet: It's a computer go server that allows free access for bots to compete against each other with 15min sudden death time limits on 19x19.
Main page: http://www.yss-aya.com/cgos/
Current standings: http://www.yss-aya.com/cgos/19x19/standings.html
BayesElo standings after at least 100 games: http://www.yss-aya.com/cgos/19x19/bayes.html
The main page gives a good overview, but if there are any questions, feel free to ask here.
Right now, personally I'm running the Rn4204_4790_980Ti bot. That's Rn 4.20.4 from https://github.com/zakki/Ray on an i7 4790k, GTX 980Ti and 16GB RAM on Windows 10, with the following command line configuration:
ray.exe --time 880 --size 19 --thread 8 --komi 7.5 --superko --tree-size 262144 --no-debug --no-early-pass
Pondering is disabled as there is a bug in 4.20.4 that leads to timeouts otherwise. I'm looking to add --reuse-subtree for the next matches. The tree size is already at max for the amount of RAM available. The time is set at 880 seconds to leave a bit of buffer for the 900 seconds time limit, but I assume this is actually overridden by GTP by the server.
Also, I'm not using the TCL script, but the Python client, as Rn doesn't completely obey the --no-debug flag and outputs some debug info on move generation when running in GPU mode, which confuses the client and terminates the session. I edited the Python client to filter out anything but the actual move response so I could run Rn with GPU.
As the CPU only has 4 real cores, I ran with 4 threads on some test sessions, but Rn would only utilize the CPU up to about 70%, and I had to bump the thread count up to 8 to get 100% utilization.
I'm looking forward to the next Leela release and giving that a spin on CGOS.
Some bot information
Rn4204_4790_980Ti
Bot: Rn 4.20.4 - https://github.com/zakki/Ray
Operator: Satorian
Hardware: i7 4790k, GTX 980Ti, 16GB RAM
OS: Windows 10
Params: ray.exe --time 880 --size 19 --thread 8 --komi 7.5 --superko --tree-size 262144 --no-debug --no-early-pass
LeelaLight 1Thread NoGPU 10.000 playouts ELO 2811
LeelaGPU 4Threads GPU (GTX 750Ti) 50.000 playouts ELO 3063 (+252)
LeelaMax 8Threads GPU (GTX 1070) 100.000 playouts ELO 3314 (+251) (+503)
Bot: Leela - https://sjeng.org/leela.html
Operator: kris
Hardware: CPU i7 2600K, 16GB RAM
Rn.4.20.4-1c1g - 3417 Elo
Bot: Rn 4.20.4 - https://github.com/zakki/Ray
Operator: cyan
Hardware: GTX 1060
Leela-0.10.5-16t1g
Bot: Leela 0.10.5 - Pre-release version - https://sjeng.org/leela.html
Operator: roy7
HardwareRyzen 1700 and a 1080Ti