A Monte Carlo tree search algorithm for optimization of load scalability in database systems
Omondi, Allan Odhiambo
MetadataShow full item record
Variable environmental conditions and runtime phenomena require developers of complex business information systems to expose configuration parameters to system administrators. This allows system administrators to intervene by tuning the bottleneck configuration parameters in response to current changes or in anticipation of future changes in order to maintain the system’s performance at an optimum level. However, these manual performance tuning interventions are prone to error and lack of standards due to varying levels of expertise and over-reliance on inaccurate predictions of future states of a business information system. The purpose of this research was therefore to investigate on how to design an algorithm that proactively reconfigures bottleneck parameters without over-relying on an accurate model of a stochastic environment. This was done using a comparative experimental research design that involved quantitative data collection through simulations of different algorithm variants. The research built on the theoretical concepts of control theory and decision theory, coupled with the estimation of unknown quantities using principles of simulation-based inferential statistics. Subsequently, Monte Carlo Tree Search, with a variant of the selection stage, was used as the foundation of the designed algorithm. The selection stage was variated by applying a “lean Last Good Reply with Forgetting” (lean-LGRF) strategy and first tested in the context of a strategy board game, Reversi. The lean-LGRF selection strategy applied over 1,000 playouts against the baseline Upper Confidence Bound applied to Trees (UCT) selection strategy recorded the highest number of wins. On the other hand, the Progressive Bias selection strategy had a win-rate of 45.8% against the UCT selection strategy. Lastly, as expected, the UCT selection strategy had a win-rate of 49.7% (an almost 50-50 win-rate) against itself. The results were then subjected to a Chi-square (χ2) test which provided evidence that the variation technique applied in the selection stage of the algorithm had a significantly positive impact on its performance. The superior selection variant was then applied in the context of a distributed database system. This also provided compelling results that indicate that applying the algorithm in a distributed database system resulted in a response-time latency that was 27% lower than the average response-time latency and a transaction throughput that was 17% higher than the average transaction throughput.