EU has funded the aCG project
Published:
A new project aCG is funded by EuroHPC under the 2023 Call for Inno4scale Innovation Studies. The project aims to make supercomputing better adapted for the future.
Funding: EU (EuroHPC), 2023 Call for Inno4scale Innovation Studies
Project full name: Innovative Algorithms for Applications on European Exascale Supercomputers
Project acronym: aCG
This project will develop and test an adaptive Conjugate Gradient (aCG) algorithm to address communication and synchronization latencies in large-scale GPU-based supercomputers. By facilitating CPU-free execution, aCG circumvents potential delays caused by standard CG algorithms that are CPU-controlled, in which the host manages every operation during GPU execution. aCG can thus leverage the new hardware and software capabilities of GPUs, such as device-initiated communication, persistent kernels, device-side synchronization, and thread block specialization, to minimize host-induced overheads and implement various communication-hiding techniques.
By using performance modelling for dynamically switching between different CG variants, the aCG can achieve optimal CG performance. The aCG API will be designed for compatibility with widely recognized linear algebra libraries, to simplify the switch from CG to aCG within existing applications. Numerous HPC applications can benefit from this innovation study, and CPU-free execution has the potential to unlock the capabilities of GPU-based supercomputers by harnessing new hardware features to improve both performance and energy efficiency.
The project aCG presents a forward-looking and potentially disruptive method for programming iterative solvers on accelerator-rich supercomputers.
Duration: 1 year
Partners: Simula Research Lab and Koç University in Istanbul, Turkey.
The Simula partner will provide its expertise in performance optimization, modeling and GPU acceleration of advanced numerical computations that rely on iterative solvers (such as the CG method) to solve sparse linear systems.
Simula participants: Prof. Xing Cai (PI), Dr. Johannes Langguth (Senior Research Scientist) and Dr. James Trotter (Postdoc).
The KU partner will contribute with essential expertise in multi-GPU programming, communication optimizations, performance modelling and scaling. From KU there are two key personnel assigned to the innovation study.
KU participants: Assoc. Professor Didem Unat (PI) and Dr. Aditya Sasongko (Postdoc).
