Logo: to the web site of Uppsala University

This thesis deals with how to develop scientific computing software that runs efficiently on multicore processors. The goal is to find building blocks and programming models that increase the productivity and reduce the probability of programming errors when developing parallel software.

In our search for new building blocks, we evaluate the use of hardware transactional memory for constructing atomic floating point operations. Using benchmark applications from scientific computing, we show in which situations this achieves better performance than other approaches.

Driven by the needs of scientific computing applications, we develop a programming model and implement it as a reusable library. The library provides a run-time system for executing tasks on multicore architectures, with efficient and user-friendly management of dependencies. Our results from scientific computing benchmarks show excellent scaling up to at least 64 cores. We also investigate how the execution time depend on the task granularity, and build a model for the performance of the task library.

Modern computer architectures, with multicore CPUs and GPUs or other accelerators, make stronger demands than ever on writers of scientific code. As a rule of thumb, the fastest, most efficient program consists of labor-intensive code written by expert programmers for a certain application on a particular computer. This thesis deals with several algorithmic and technical approaches towards effectively satisfying the demand for high-performance parallel programming without incurring such a high cost in expert programmer time. Effective programming is accomplished by writing performance-portable code where performance-critical functionality is provided either by external software or at least a balance between maintainability/generality and efficiency.