Lattice Field Theory at the Exascale Frontier
Lead Research Organisation:
Science and Technology Facilities Council
Department Name: The Hartree Centre
Abstract
Abstracts are not currently available in GtR for all funded research. This is normally because the abstract was not required at the time of proposal submission, but may be because it included sensitive information such as personal details.
Publications
| Description | 1. New Algorithms Developments: One of the objectives is a new algorithms developments for LQCD. Hartree Centre's team has developed a novel MCMC Matrix Inversion for LQCD matrices. Currently we are investigating the effectiveness of MCMCMI as a preconditioner of matrices arising from LQCD and then we are solving the resulting system using GMRES. This has demonstrated that the preconditioner is effective in the sense that it will reduce the number of steps required to achieve a desired precision. We are currently improving an existing C++ test set to utilise the Tpetra package of Trilinos for complex-valued matrices, thereby merging computation and testing of the preconditioner into one project. Furthermore, along with other projects, we are working on an implementation of the method that can use a linear operator function instead of a matrix. Result: Novel hybrid stochastic/deterministic parallel algorithms based MCMCMI for solving systems arising from LQCD were developed. The approach can be applied to matrices and to linear operators function instead when fully developed. Can be applied to similar problems arising in other areas and generalised due course. 2. Training: One of the objectives of the project is to develop a roadmap for a broad community to be able to exploit exascale facilities where training is a crucial component. Hartre Centre is the pioneering UK supercomputing centre that predominantly works with industry has contributed to this objective by organising and delivering a Parallel Programming and Data Parallelism School for Lattice QCD and HEP communities Impact: Our Exit Questionnaire is based on the well-established Kirkpatrick model of evaluation of professional training. The questionnaire is purposefully very concise and focusing solely on the learning outcomes of the courses. We were specifically interested in the responses of the attendees on the "Have you been introduced to a technique or a method that you would like to apply in your research/work to study further?" and 94% of the respondents to the Exit questionnaire have answered positively to it, while 70% of them have given further details on how they intend to do that. The program consisted of several modules: Module 0 (pre-course session): Hartree Driving Licence (how to run your code on Scafell Pike) assessment is through optional on-line test. This is a session on Scafell Pike Customer On-boarding. So far, five participants from Lattice and HEP communities have attempted the test and two have passed it. Module 1 Introduction to parallel computing, algorithms and GPU programming: This module was designed of two parts, Introduction Advanced Computer Architectures and Algorithms and GPU Programming based on the Nvidia DLI Ambassadorship program. The assessment of the first part was through lightning talks by the attendees on " How to apply a presented Parallel Design Techniques or Performance tools to their PhD project". The issued discussion with all the lecturers help attendees clear any questions they had at the moment. The GPU training had a number of practical session and provided access to the Nvidia resources. Interested attendees can take the Nvidia assessed tests and get certified. Module 2 Further mathematical concepts on communication avoiding and hybrid algorithms: Overview was given on Communication avoiding algorithms and Stochastic and hybrid stochastic/deterministic algorithms, as well as particular UQ concepts. This module was designed as Team work exercise on three topics with mentor working with each team. The teams were given time and further material to scope and present a possible solution to a given problem. The teams have presented their proposals in front of a panel and the rest of the teams. Open discussion was held after each presentation. The three topics were: • Monte Carlo method: parallelisation approaches • Using Modern Numerical Libraries • Krylov subspace methods: parallelisation exercises Module 3 OpenMP, MPI and Hybrid programming (optional): the module introduced with practical exercises OpenMP and MPI & Hybrid programming environments. The final session was open discussion and coding surgery. Results are three-fold: 2.1. Our training contributed to attendees learning new parallel and communication avoiding algorithmic techniques as well as UQ concepts and advanced parallel programming approaches and tools so that they can use these efficiently in their research and work. In fact in the exit questionnaire 94% of the attendees that answered it said that they have "been introduced to a technique or a method that they would like to apply in your research/work to study further" and 70% of them have given further details on how they intend to do that. 2.2. The approach we have taken will enable us to develop further the material and re-use components/modules of the training for wider set of scientific, engineering communities as well as communities from healthcare and environmental areas. 2.3. The advanced teaching and learning methodology developed and implemented for the on-line teaching and delivery of the school will be enhanced and used for other on-line courses we are developing both for academia and industry. 3. Benchmarking Hartree Centre Exa-LAT team is currently involved in QCD code SOMBRERO benchmarking and performing an initial performance analysis of the code. The scaling of the parallel code has been studied for a set of datasets on Archer 2. For runs involving large numbers of MPI tasks the inter-node scaling is generally very efficient. However, the intra-node parallel efficiency reduces when nodes are fully occupied by MPI tasks, suggesting that memory access contentions are limiting performance. Profiling tools are being used to try to understand better this behaviour. Future work will involve analysing and understanding SOMBRERO parallel performance on a range of architectures (e.g. Scafell Pike) and working with code developers to address or mitigate the parallel bottlenecks. Results: Applying advanced benchmarking strategies to the LQCD codes. |
| Exploitation Route | Use of results/Impact: 1. One of the objectives is a new algorithms developments for LQCD was met, novel hybrid MCMCMI method for tackling linear systems arising from LQCD problems have been developed. We and Hartree Centre team are using this approach to solve linear systems arising from variety of problems and are aiming to apply it in the future to linear operators functions instead and any matrices that this method can be suitable to apply to. 2. Our training enabled: - The attendees to learn new parallel and communication avoiding algorithmic techniques as well as UQ concepts and advanced parallel programming approaches and tools so that they can use these efficiently in their research and work (see above). - Us to develop further the material and re-use components/modules of the training for wider set of scientific, engineering communities as well as communities from healthcare and environmental areas. - to develop and implement advanced teaching and learning methodology for on-line teaching and delivery of the school which we will enhance and use for other on-line courses we are developing both for academia and industry. Impact: Novel material developed in the subject areas outlined above that we will re-use and adapt for wider communities bot in academia and industry together with novel teaching and learning methodology for on-line and mixed delivery. |
| Sectors | Digital/Communication/Information Technologies (including Software),Education,Energy,Environment,Other |
| Description | New Algorithmic Developments impact: Novel hybrid stochastic/deterministic parallel algorithms based MCMCMI for solving systems arising from LQCD were developed. The approach can be applied to matrices and linear operators function instead when fully developed. Can be applied to similar problems and generalised due course to be applied to problems arising in scientific and industrial problems in wide range of areas. Organised Parallel Programming and Data Parallelism School for Lattice QCD and HEP communities Impact: Novel material developed in the subject areas outlined below can be re-used and adapted for wider communities both in academia and industry together with novel teaching and learning methodology for on-line and mixed delivery. Our Exit Questionnaire is based on the well-established Kirkpatrick model of evaluation of professional training. The questionnaire is purposefully very concise and focusing solely on the learning outcomes of the courses. We were specifically interested in the responses of the attendees on the "Have you been introduced to a technique or a method that you would like to apply in your research/work to study further?" and 94% of the respondents to the Exit questionnaire have answered positively to it, while 70% of them have given further details on how they intend to do that. The program consisted of several modules: Module 0 (pre-course session): Hartree Driving Licence (how to run your code on Scafell Pike) assessment is through optional on-line test. This is a session on Scafell Pike Customer On-boarding. So far, five participants from Lattice and HEP communities have attempted the test and two have passed it. Module 1 Introduction to parallel computing, algorithms and GPU programming: This module was designed of two parts, Introduction Advanced Computer Architectures and Algorithms and GPU Programming based on the Nvidia DLI Ambassadorship program. The assessment of the first part was through lightning talks by the attendees on " How to apply a presented Parallel Design Techniques or Performance tools to their PhD project". The issued discussion with all the lecturers help attendees clear any questions they had at the moment. The GPU training had a number of practical session and provided access to the Nvidia resources. Interested attendees can take the Nvidia assessed tests and get certified. Module 2 Further mathematical concepts on communication avoiding and hybrid algorithms: Overview was given on Communication avoiding algorithms and Stochastic and hybrid stochastic/deterministic algorithms, as well as particular UQ concepts. This module was designed as Team work exercise on three topics with mentor working with each team. The teams were given time and further material to scope and present a possible solution to a given problem. The teams have presented their proposals in front of a panel and the rest of the teams. Open discussion was held after each presentation. The three topics were: • Monte Carlo method: parallelisation approaches • Using Modern Numerical Libraries • Krylov subspace methods: parallelisation exercises Module 3 OpenMP, MPI and Hybrid programming (optional): the module introduced with practical exercises OpenMP and MPI & Hybrid programming environments. The final session was open discussion and coding surgery. |
| First Year Of Impact | 2021 |
| Sector | Digital/Communication/Information Technologies (including Software),Education,Other |
| Impact Types | Societal |
| Description | ExaTEPP: Theoretical and Experimental Particle Physics at the Exascale Frontier |
| Amount | £163,318 (GBP) |
| Funding ID | EP/X01939X/1 |
| Organisation | Engineering and Physical Sciences Research Council (EPSRC) |
| Sector | Public |
| Country | United Kingdom |
| Start | 11/2022 |
| End | 11/2024 |