Adaptive Loop Tiling for a Multi-cluster CMP

Loop tiling is a fundamental optimization for improving data locality. Selecting the right tile size combined with the parallelization of loops can provide additional performance increases in the modern of Chip MultiProcessor (CMP) architectures. This paper presents a runtime optimization system which automatically parallelizes loops and searches empirically for the best tile sizes on a scalable multi-cluster CMP. The system is built on top of a virtual machine and targets the runtime parallelization and optimization of Java programs. Experimental results show that runtime parallelization and tile size searching are capable of improving performance for two BLAS kernels and one Lattice-Boltzmann simulation, despite overheads.