The NAStJA Framework: Non-Collective Scalable Global Communications

Authors: Marco Berghoff (Karlsruhe Institute of Technology), Ivan Kondov (Karlsruhe Institute of Technology)

Abstract: In recent years, simulations in various areas of science and engineering have proven to be very useful. To efficiently deploy simulation codes on current and future high-performance computer systems, high node level performance, scalable communication and the exclusion of unnecessary calculations are an absolute must when developing new solvers.

We have introduced the NAStJA framework, a block-based MPI parallel solver for algorithms, based on regular grid methods, i.e., stencil codes. NAStJA has a dynamic block adaptation, which modifies the calculation domain around the region in which the calculation is currently taking place. The creation and deletion of blocks are autonomously managed within local neighborhoods. Collective all-gather communication is avoided by using a multi-hop network to distribute information across the entire domain that greatly improves the application scaling. In this contribution, we present applications that can benefit from this adaptive method and scaling tests demonstrating the excellent scalability.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF


Back to Poster Archive Listing