ACTS is a free and open-source software project for track reconstruction in particle physics experiments. As a modernized version of the particle tracking code used by the ATLAS experiment at the CERN Large Hadron Collider, the project is focused on adoption of modern C++ standards, usability in multi-threaded workflows, and increased use of vectorization and accelerators (gpgpu).
The fitting step of track reconstruction, in which particle track hypotheses are confronted to experimental data, uses the Kalman Filter algorithm in the 5-dimensional space of possible track parameters. This entails performing linear algebra operations on a large amount of 5x5 matrices, a problem which has only received a limited amount of scrutiny from the linear algebra community in the past. We are thus evaluating the relative performance of multiple linear algebra toolkits on this problem, using realistic input data from typical ACTS use cases.
Until now, ACTS has only been able to perform computations on CPUs. In this project, we aim to investigate the use of extra accelerator hardware, such as GPUs and Intel’s Xeon Phis. We would like to privilege the desirable property of sharing one single codebase across all target hardware configurations, and to this end we want to first focus on the Kokkos toolkit, which promises to permit the development of performance-portable codebases by easily adapting several key characteristics, including data layout, to the target hardware architecture.
We want to start by writing a Kokkos-based version of the Kalman Filter benchmark, and see how well it performs with respect to CPU-based versions. We also want to evaluate the portability of this version by testing it on the various accelerators available in the GridCL/ACP facility (Xeon Phi 5110P, AMD FirePro S9170, NVIDIA K20, GeForce GTX Titan and hopefully V100). After this, we want to integrate kokkos-based computations in the ACTS codebase, and see how well it performs in a more realistic use case.
In the context of a key step of particle hunting, this project gives the opportunity to implement 5x5 linear algebra optimizations, and evaluate them with a large set of cutting-edge accelerators (gpgpu).
C++, Computing with GPU, Linear Algebra.