University of Houston

Compiler cost models or performance models are analytical models to estimate the costs of executing a specific section of codes, such as computation-intensive loops. It is often used in the compiler to decide whether certain optimization has performance benefits, thus to guide the optimizations. The main drawback of compiler cost models is that their performance estimations do not take into account the performance impact from the interference or contention for resources between parallel threads such as the false sharing effects, the competition to use shared cache or memory bus. Exascale systems such as many-core and heterogeneous architectures that have become ubiquitous and widely used by many different users provide with increasing number of cores and the decrease of average memory and bus bandwidth per cores. On these systems, shared resource interference and contention will have significant performance impacts for applications. Thus, it is very importa nt for compilers' and performance estimating tools' cost models to accurately estimate the execution performance of applications on these architectures by considering concurrent utilization of limited shared resources. This dissertation will aim at defining compile time cost models with regards to the contention of using shared resources on current many-core and heterogeneous platforms, including the shared cache and memory bandwidth. The defined cost models will be used by compiler, tools and runtime system for making compiler transformation decisions, for helping making auto-tuning decisions, and for runtime dynamic optimizations or feed-back driven optimizations.