Fast Autotuning Configurations of Parameters in Distributed Computing Systems Using Ordinal Optimization

Conventional autotuning configuration of parameters in distributed computing systems using evolutionary strategies increases integrated performance notably, though at the expense of consuming too much measurement time. An ordinal optimization (OO) based strategy is proposed in this work, combined with neural networks to improve system performance and reduce measurement time, which is fast enough to autotune configurations for distributed computing applications. The method is compared with a well known evolutionary algorithm called Covariance Matrix Algorithm (CMA). Experiments are carried out using high dimensional rastrigin functions, which show that OO can reduce one to two orders of magnitude of simulation time while at the cost of an acceptable scope of optimization performance. We also carried out experiments using a real application system with three-tier web servers. Experimental results show that OO can reduce 40% testing time on average at a reasonable and slight cost of optimization performance.