Repository navigation
Home
Welcome to the Eurohack21 wiki!
This page contains material for the Eurohack21 GPU Hackathon, held online from Sep. 20 to Sep. 29, 2021. The goal of the GPU Hackathon is to provide application development teams a conducive environment to jump-start development of accelerator-capable versions of their applications by providing access to the best mentors from the GPU community from supercomputing centers, vendors and universities.
Eurohack21 has a dedicated Slack workspace where we are going to share comments, material etc. during the hackathon. Each team will also have its own slack channel for coordinating. Please make sure to register to the workspace!
The teams are expected to do an initial and a final short presentations during the hackathon. More details and templates will be shared in the Slack channel.
GBS simulations are used to study plasma turbulence at the boundary of tokamak devices, a region characterized by a large drop of temperature, from about 100-million-degrees in the core to a few hundred degrees at the machine walls. GBS is currently capable of simulating medium size tokamaks such as TCV on Tier-0 CPU-only supercomputers. However, a much larger grid is required to simulate large-scale tokamaks such as ITER and DEMO, an essential step to improve our prediction capability. Therefore, accelerators are mandatory to complete the simulations in a feasible time. In a typical GBS simulation, there are two main components called plasma and neutral, which are evolved in parallel. Usually, a neutral step takes a few hundred plasma steps. The main bottlenecks are the evaluation of the right-hand side of the Braginskii equations and the resolution of the linear system for Poisson, Ampere, and neutrals.
The ICON model is a climate model that solves the equations of motion of the atmosphere and oceans, i.e. a fluid dynamics problem.
Palabos is a general software library for computational fluid dynamics based on the lattice Boltzmann method. It is massively parallel and runs on CPU clusters with the help of MPI. The code is open-source and receives community contributions, although the core development team is situated in Geneva.
The GEMPIC library implements a geometric particle in cell method for the solution of the Vlasov-Maxwell system for plasma physics application. It is built on the AMReX library and leverages AMReX parallelisation capabilities to make the code HPC scalable.
The solver of the research group, URANOS, is a general-purpose Finite Difference (FD) fluid dynamics code suitable for Direct Numerical Simulations (DNS) and Large-Eddy Simulations (LES) of compressible flows over complex geometries. In-house code developed at the Industrial Engineering Department of the University of Padova, URANOS deals with the time-dependent compressible Navier-Stokes system of equations in a conservative formulation and takes advantage of a peculiar sharp-interface Immersed Boundary Method (IBM) to account for embedded (even moving) complex geometries.
The code we plan to use in this project is a semi-implicit, multidimensional, fully parallelized electromagnetic Particle-In-Cell (PIC) code (Gonzalez-Herrero et al. 2018). In the PIC algorithm, macroparticles representative of several real plasma particles are moved under the influence of the self-consistent electromagnetic fields. The latter are calculated on a grid via Maxwell’s equations, where the source terms are obtained from the particles (Dawson 1983). In our semi-implicit code both motion equations for the particle and Maxwell’s equations for the field are solved adopting an implicit discretization in time. It can be demonstrated that the adopted spatial and temporal discretization conserves energy exactly (Lapenta et al. 2017).
GmgPolar is a C++ code implementing a geometric multigrid method for the solution of a Poisson-like equation defined on a disk-like shape. This equation is classically obtained as a subproblem from the simulation of plasma in Tokamak reactors performed with the gyrokinetic code GyselaX. The simulation of plasma in Tokamak reactors involves high costs in terms of computation and memory and requires the use of efficient high performance computing (HPC) techniques. For this reason, a GPU implementation would be the final goal.
The E2E group developed four standalone applications for particle reconstruction using CNNs: the electron/photon with 1 image layer, quark/gluon with 3 image layers, the top jet classifier with 8 image layers, and the tau classifier with 14 image layers. Recently, the E2E group is implementing GNN models to these E2E applications, as graph representation learning can handle irregular grids with non-Euclidean geometry, encode physics knowledge via graph construction and introduce relational inductive bias into data-driven learning systems.