Evaluation of two Formulations of the Conjugate Gradients Method with Transactional Memory
Martin Schindewolf, Björn Rocker, Wolfgang Karl, Vincent Heuveline · University Library Heidelberg · 2013
Transactional Memory (TM) offers new possibilities for algorithmic design. This paper evaluates TM implementations of two algorithmic variations of the wide-spread conjugate gradients method (CG) regarding their performance on multi-core CPUs employing TM. Through applying tools for TM that visualize the TM application behavior, we show that the main bottleneck for both is the waiting times at barriers and illustrate the implementation of reduction operations with TM in a beneficial way. Performance monitoring through using the PAPI interface uncovers the quantity and type of instructions that each algorithms requires. This basic work is the key for environment-aware numerics as well as a hint for software developers who plan to use TM.