Transcript GPW08 - GeoCities

Running “Zen” on Computer
Clusters
HIDEKI KATO† and IKUO TAKEUCHI†
† The
University of Tokyo
November 13th, 2009
Contents
Background
Related Work
Parallel Monte-Carlo Tree Search
Our Architecture
Experiments
Conclusion and Future Work
Background
• Challenge
– Can beat human professional players?
• Interests
– How “Zen” on an HPC cluster is strong?
– Also, on a pc cluster at home?
– Distributed MCTS on the Internet
• Real world applications of MCTS
– Provides smarter planning for intelligent robots,
intelligent vehicles, etc
– Environment: Many small processors on a LAN
Related Work
• S. Gelly et al. introduced SMT PMCTS for sharedmemory SMP systems (2006)
• T. Cazenave et al. proposed and evaluated three PMCTS
algorithms on a 16 Intel Pentium-4 MPI cluster (2007)
• G. Chaslot et al. evaluated root, leaf and tree
parallelization on 2 x 8 core IBM Power5 (2008)
• S. Gelly et al. proposed a combination of tree and root
parallel MCTS for MPI clusters of shared-memory SMP
nodes (2008)
• H. Kato et al. proposed and evaluated a leaf parallel
MCTS on an asymmetrical pc cluster (2008)
Parallel Monte Carlo Tree Search
• Tree parallelization
– Symmetrical multithread parallelism on shared
memory multiprocessor systems
– Used by almost all MC Go programs
• Leaf parallelization
– For asymmetrical computer clusters
– Fudo Go
• Root parallelization
–
–
–
–
Shares search tree in part
Less communication
Best match with HPC clusters
MoGo, Many Faces of Go and Fuego
Parallel Monte Carlo Tree Search
(cont’d)
Tree parallelization
Leaf parallelization
Root parallelization
Global lock
Local locks
G. Chaslot, et al. 2007
Our Architecture
• Requirements
– Since “Zen” is a commercial product, less modification is better
– Root parallelization should be the best
– Can run on non-MPI environments
• Master manages GTP communications and conducts
exchanging root information
– Broadcasts incoming messages and sends answers back by the
majority rule to/from the slaves
– Slaves send their root information according to the master’s
message
• Master and slave programs are built into “Zen” for
convenience and shorter delay
Our Architecture (cont’d)
GTP
Master
Root
information
Broadcast
Majority rule
Master
Gather,
Average, &
Broadcast
Slave 1
Slave 1
Slave n
Slave n
Experiments
• No majority rule
• All results are Elo ratings against Zen (self-play) on
the same PC
Board size
Games
Komi
Time for a move
Interval
Node computers
9x9
~1500
7
0.3 s
0.1 s
2 to 32
Scalability
Winning rate of self-play vs. Number of node computers (9 x 9)
+400
HA8000
1 thread self-play 0.3s/move
Handcraft pc 1 thread self-play 0.3s/move
MoGo
Winning rate (Elo)
+300
+200
+100
0
-100
0
1
2
3
4
Log2 Number of node computers (master and slaves)
5
6
Conclusion and Future Work
• We have implemented a root parallelization
version of Zen by adding about 1,500 line of C++
code using Boost asio library
• Not good scalability on HA8000 HPC cluster on
9 x 9 board
• How about 19 x 19? How about other hardware?
• Benchmark on 19 x 19