Monte Carlo Simulations of Spin Glasses on Cell Broadband Engine

Several large-scale computational scientific problems require high-end computing systems to be solved. In the recent years, design of multi-core architectures delivers on a single chip tens or hundreds Gflops of peak computing performance, with high power dissipation efficiency, and it makes available computational power previously available only on high-end multi-processor systems. The aim of this Ph.D. thesis is to study the capability of multi-core processors for scientific programming, analyzing sustained performance, issues related to multicore programming, data distribution, synchronization, in order to define a set of guideline rules to optimize scientific applications for this class of architectures. As an example of multi-core processor, we consider the Cell Broadband Engine (CBE), developed by Sony, IBM and Toshiba. The CBE is one of the most powerful multi-core CPU current available, integrating eight cores and delivering a peak performance of 200 Gflops in single precision and 100 Gflops in double precision. As case of study, we analyze the performances of CBE for Monte Carlo simulations of the Edwards-Anderson Spin Glass model, a paradigm in theoretical and condensed matter physics, used to describe complex systems characterized by phase transitions (such as the para-ferro transition in magnets) or model “frustrated” dynamics. We descrive several strategies for the distribution of data set among on-chip and off-chip memories and propose analytic models to find out the balance between computational and memory access time as a function of both algorithmic and architectural parameters. We use the analytic models to set the parameters of the algorithm, like for example size of data structures and scheduling of operations, to optimize execution of Monte Carlo spin glass simulations on the CBE architecture.