1st Call for Short Term Collaboration Grants for HPC

Closed

Start: July 1, 2018
Deadline: Sept. 30, 2018
Budget: EUR 40 000

Funding for Short Term Collaboration Exchanges in HPC

Eurolab4HPC is a Coordination Action funded by the European Commission under H2020 with the bold commitment to build the foundation for a European Research Center of Excellence in High-Performance Computing (HPC) Systems.

A funded Short Term Collaboration typically lasts 3 months and needs to end by 31 December 2019. The objective of the instrument is to foster tight integration among research teams from multiple communities working across all levels: hardware, architectures, programming and applications, in alignment with the goals of the Eurolab4HPC Vision. Each proposal should contain representatives from multiple research communities across the whole system stack, including strong commitment from the user communities. The grants will provide funding for short cross-discipline visits, targeted at PhD students, post-doctoral researchers and academics to work on a common project in different layers of the compute stack. Funding is available for independent small studies and benchmarking activities to capture requirements, metrics and opportunities to improve the platforms and technologies. A short term collaboration grant has a maximum budget of EUR 5 000 and will be paid as a daily allowance of EUR 55 for max. 91 days. Half of the Eurolab4HPC grant will be paid upon receival of a grant agreement signed by the grantee, the host and home institution. The second half of the grant will be paid upon submission of a public summary of the activity that will be published on the Eurolab4HPC website.

Only proposals where either the home or host institution of the applicant is a Eurolab4HPC member, are eligible for the grants. A list of Eurolab4HPC members can be found on eurolab4hpc.eu/community.

The evaluation be done by at least two external independent experts selected by the steering committee under a Non-Disclosure Agreement (NDA), and without conflict of interest with the proposers. The evaluation process will be composed of five steps: (i) eligibility check, (ii) review by external experts, (iii) consensus meeting chaired by the partner in charge of the call, (iv) decision to grant the best proposals by the steering committee, (v) feedback about the outcome of the evaluation process to all the proposers. All applicants will receive information about the outcome of their application by 15 November 2018.

See the attached instructions and proposal template for more details. The proposals have to be submitted via the Eurolab4HPC website no later than on 30 September 2018. The funding period after proposal acceptance is expected to start at the earliest on 15 November 2018.

Funded projects

Title	Team
HPC for Neutron Transport in a Fusion Reactors Core	From Comisión Nacional de Energía Atómica, CNEA, Argentina; to Barcelona Supercomputing Center (BSC)	Alejandro Soba
Scaling Large Machine Learning Models for HPC	From Barcelona Supercomputing Center (BSC); to Tokyo Institute of Technology	Albert Kahira	Scaling Large Machine Learning Models for HPC Albert Kahira From Barcelona Supercomputing Center (BSC); to Tokyo Institute of Technology Parallelism in deep learning training is inevitable. Mostly because of the need to accelerate training and recently due to larger models and bigger datasets. Though data parallelism is the most common parallelisation method, numerous strategies have been suggested, each with a promise to accelerate the training process. However, there is no clear way to determine which approach to use for a given model or dataset. By analysing memory, communication and performance, we have developed an oracle that suggests the best parallelisation strategy given a set of parameters such as model, dataset and system specifications. We have also performed a cocktail of tests with standard datasets and with newer emerging scientific datasets such as Cosmoflow on the ABCI supercomputer.
Processing in Memory for Data Intensive Applications	From University of Malaga; to ETH Zürich	Ivan Fernandez Vega	Processing in Memory for Data Intensive Applications Ivan Fernandez Vega From University of Malaga; to ETH Zürich From the general-purpose cores approach, I obtained more than 3.6x better performance using a PIM approach than a typical host configuration. Using the custim accelerator based on RTL logic, the speedup reached 18x. Those results, including energy consumption estimation that we obtained have been written in a paper for a conference, which is going to be submitted to HPCA.
Themis high performance simulation framework	From University Grenoble Alpes/Inria; to Imperial College	Christopher Eldred	Themis high performance simulation framework Christopher Eldred From University Grenoble Alpes/Inria; to Imperial College During this mobility, we have been working on two publications covering (some of) the work performed at Imperial College: One publication covering, amongst other aspects, preconditioners for mimetic Galerkin difference elements based on using operators from the lowest order spaces. A particulat emphasis will be mixed problems that commonly occur in geophysical applications. A second publication abour variational formulations for fluid dynamics that use split exterior calculus.