Implementation and Optimization of the Lattice Boltzmann Method for the Jackal DSM System

The goal of this thesis is to implement and optimize the lattice–Boltzmann method for the distributed shared memory system Jackal. The lattice Boltzmann method is a new and promising alternative to the traditional Navier–Stokes solvers in the domain of computational fluid dynamics. It is based on cellular automata and operates on a domain, were all cells are normalized in space. Because of the great need for memory and computing time most computational fluid dynamic problems cannot be solved on a single computer in sufficient time. Therefore the solvers have to be parallelised. Because of the implicit parallelism of the lattice Boltzmann method this promises a good speed–up. For parallelization a distributed shared memory approach is instead of the traditional Message Passing Interface approach. The distributed shared memory system provides the programmer a global address space on top of the distributed memory of the individual nodes of the cluster. Hence, all data can be addressed regardless of which node allocated them. The Jackal system implements such a distributed shared memory system for Java programs by cooperation of it’s compiler and runtime system. The Jackal system is also capable of automatically distribute a multi–threaded program on a cluster. Our performance evaluation shows that by optimizing the lattice Boltzmann method the sequential program we are able to achieve 5.5 millions of lattice updates per second for a 1000 × 1000 grid. By further optimizing the sequential program in a distributed shared memory specific way we are able to achieve a speed–up of 3.4 in both 2D and 3D test case. This results in a total of 2 millions of lattice updates per second.