Coupled cluster (CC) methods, while effective for some of the most complicated chemical problems,  are still very time consuming for large systems which can severely limit the use of these methods.  In addition, tuning the algorithms to large and complex massively parallel computers is becoming much more difficult.  Dataflow-based models with a task scheduling system provide one method to improve the the scalability and performance of CC codes. Using a dataflow-based representation, the CC methods can be expressed in a way that allows the developer to decompose the computation into fine-grained tasks for maximum parallelism, or task aggregates for higher locality. At the same time, the data distribution and load balancing are handled and tuned in a decoupled manner from the algorithm itself, and the communication between tasks is overlapped with computation.  In this talk, the integration of the task scheduling system PaRSEC (Parallel Runtime Scheduling and Execution Controller) with the NWChem Tensor Contraction Engine (TCE) CC will be discussed and contrasted with the more traditional, coarse-grained parallel methods.  A performance evaluation will be presented along with several applications to larger systems.

Use of dataflow-based execution to improve scalability and performance of coupled cluster codes

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.