diff --git a/data/talks.yaml b/data/talks.yaml index ff1eea8bebb3618ab1461178311a6e765ccdbb70..32eb05fbecf2c461743e65600e19d63e0e1d15d0 100644 --- a/data/talks.yaml +++ b/data/talks.yaml @@ -4,6 +4,44 @@ # references. Nominally we will use /talks. cards: + - meeting: Supercomputing Frontiers Europe 2018 + location: Warsaw, Poland + date: March 2018 + title: "Individual time-stepping in cosmological simulations: A challenge for strong scaling and domain decomposition algorithms" + author: Matthieu Schaller + abstract: "For many decades cosmological simulations have used individual time-steps for the particles (or cells) that are evolved forward in time. Given the large range in time-step sizes + required for accurate integration, this allows substantial gains in time to solution compared to the more widely adopted in science technique of using a single time-step for the whole + computational domain. However, only limited work has been done to improve domain decomposition techniques to make them aware of this multiple time-stepping scheme. This results in domains + perfectly suited for time-steps where most particles require updating but leads to terrible performance with large levels of imbalance for the steps where only a small fraction of particles are + being updated. This generates poor overall strong-scaling performance and leads to a poor utilisation of the systems. + + In this paper, we will introduce some of these short-comings and present a solution based on task-based parallelism and a weighted graph domain decomposition algorithm implemented in the + open-source cosmological code SWIFT. By carefully weighting the different tasks in an overall cluster-wide graph, we can divide the work and not the data optimally. If we further penalise the + creation of communication tasks we can achieve a domain decomposition where no MPI communication is required for the time-steps with few active particles. This effectively eliminates most imbalance + issues and leads to optimal performance not just for time-steps with large number of updates but also for the ones with few particles to update. We achieve much better strong-scaling performance than + commonly used codes and achieve a time-to-solution 30 times smaller than the de-facto standard Gadget code." + links: + - href: "SuperComputingFrontiers_2018.pdf" + name: Slides + + - meeting: SIAM Conference on Parallel Procecssing for Scientific Computing 2018 + location: Tokyo, Japan + date: March 2018 + title: "Using Task-Based Parallelism, Asynchronous MPI and Dynamic Workload-Based Domain Decomposition to Achieve Near-Perfect Load-Balancing for Particle-Based Hydrodynamics and Gravity" + author: Matthieu Schaller + links: + - href: "SIAM_2018.pdf" + name: Slides + + - meeting: Intel HPC DevCon 2017 + location: Denver, USA + date: November 2017 + title: "Task-based Calculation + Task-based MPI + Task-based I/O = Maximal Performance" + author: Matthieu Schaller + links: + - href: "HPC_DevCon_2017.pdf" + name: Slides + - meeting: ParCo Conference 2017 location: Bologna, Italy date: September 2017 diff --git a/talks/HPC_DevCon_2017.pdf b/talks/HPC_DevCon_2017.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0619b78af1fbc0c47a1536006f35323f560a5a05 Binary files /dev/null and b/talks/HPC_DevCon_2017.pdf differ diff --git a/talks/SIAM_2018.pdf b/talks/SIAM_2018.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5caae781b1a3b718326af88d85ca81b50cc5f56 Binary files /dev/null and b/talks/SIAM_2018.pdf differ diff --git a/talks/SuperComputingFrontiers_2018.pdf b/talks/SuperComputingFrontiers_2018.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd99f29e5c6fde805ff1b655f37a37fc79a36ccb Binary files /dev/null and b/talks/SuperComputingFrontiers_2018.pdf differ