Fast and Highly Scalable Bayesian MDP on a GPU Platform
Abstract
By employing Optimal Bayesian Robust (OBR), Bayesian Markov
Decision Process (BMDP) can be a power optimization method to
solve large problems. However, due to the “curse of dimensionality”,
the data storage limitation hinders the practical application of
BMDP. To overcome this impediment, we propose a novel Improved
Compressed Sparse Row (ICSR) data structure in this paper, and developed
the implementation of BMDP solver with ICSR technique
on a heterogeneous platform with GPU. The simulation results
demonstrate that our techniques achieve about a 5× reduction in
memory utilization over using full matrix, and an average speedup
of 4.1× over using full matrix. Additionally, we present a study of
the tradeoff between the runtime and the trends of result difference
between our ICSR techniques and using full matrix.
Decision Process (BMDP) can be a power optimization method to
solve large problems. However, due to the “curse of dimensionality”,
the data storage limitation hinders the practical application of
BMDP. To overcome this impediment, we propose a novel Improved
Compressed Sparse Row (ICSR) data structure in this paper, and developed
the implementation of BMDP solver with ICSR technique
on a heterogeneous platform with GPU. The simulation results
demonstrate that our techniques achieve about a 5× reduction in
memory utilization over using full matrix, and an average speedup
of 4.1× over using full matrix. Additionally, we present a study of
the tradeoff between the runtime and the trends of result difference
between our ICSR techniques and using full matrix.