Parallel logic programming (PLP) systems have obtained good performance on traditional bus-based shared-memory architectures. However, the scalable multiprocessors being developed today pose new challenges. Our experience with a sophisticated PLP system, Andorra-I, demonstrates that indeed performance suffers greatly on modern architectures. In order to improve performance, we perform a detailed analysis of the cache behaviour of all Andorra-I data structures via executiondriven simulation of a DASH-like multiprocessor. Based on this analysis we optimise the Andorra-I code using 5 different techniques. Our results show that the techniques provide significant performance improvements, leading to the conclusion that PLP systems can and should perform well on modern scalable multiprocessors.