The DEEP-SEA project builds on the results of the previous DEEP, DEEP-ER and DEEP-EST projects. It produces an integrated SW stack for Modular Supercomputing Architecture (MSA) systems, which provides application developers with convenient and powerful tools to exploit the benefits of MSA for heterogeneous HPC and combined HPC/AI applications and workflows. The project interacts closely with the IO-SEA and RED-SEA projects, which themselves focus on the I/O and network aspects of MSA systems.
The DEEP-SEA stack includes leading programming models for node-level and system-level (distributed-memory) parallelism, adopts ParaStation Modulo as a powerful process management and system provisioning tool, and introduces the novel concept of “optimization cycles” which guide users through the use of advanced performance analysis and optimization tools. An important focal topic is to support malleable applications and adapt the set of allocated resources to the actual demand. To this end, DEEP-SEA prototypes extensions to Slurm and production MPI implementations. A second cross-cutting topic is the support for multi-level memory hierarchies, which combine a fast memory tier with a larger yet slower memory level.
On the system side, the DEEP-SEA project has upgraded the DEEP-EST prototype with a number of nodes connected by the Atos BXI network, and is in the process of adding a few nodes with two memory levels (DRAM and HBM).
This DEEP prototype consists of a 50-node dual-processor Intel® Xeon® Cluster module using InfiniBand EDR, a 75-node Booster with one NVIDIA V100 GPU each coupled to a small x86 host CPU and its own EDR InfiniBand interconnect, a data analytics module with 16 high-end dual-processor Intel® Xeon® nodes with dual-processor Intel® Xeon® nodes. A conventional, disk-based storage module and an innovative, NVMe flash module support low-latency, high-bandwidth parallel I/O via the BeeGFS parallel file system. The Cluster and Booster modules use Megware’s blade and warm water cooling technologies, and Megware-developed power meters report the overall power and energy use of each node with a time resolution of a few milliseconds. ParaStation MPI provides seamless and transparent communication across the whole prototype.
Together with the lessons learnt from the previous DEEP projects, eight complex, real-world HPC applications from the fields of computational fluid dynamics, earth system simulation, molecular dynamics, neutron transport, seismic imaging (app1, app2), and weather forecasting provided co-design input for the DEEP-SEA SW stack. The space weather application introduced AI/ML components.
The DEEP-SEA project started on April 1, 2021 and is scheduled to end on March 31, 2024. The consortium consists of fourteen partners from eight European countries, with Forschungszentrum Jülich acting as the coordinator. The project budget was 15.1 M€, and the indicative European Union and national funding amounts to this budget.
The DEEP-SEA Project is receiving funding from the European High-Performance Computing Joint Undertaking (EuroHPC JU) under grant agreement n° 955776. The EuroHPC JU receives support from the European Union’s Horizon 2020 research and innovation programme and Germany, France, Spain, Greece, Belgium, and Sweden.
The Scientific Advisory Board (SAB) is a council of international academic and industrial experts that was created with the aim to provide technical, practical, societal and legal recommendations to the DEEP-SEA project, relative to the development and the dissemination of its activities. The SAB was founded in 2021 and currently has 9 members from institutions and companies that are well recognized in the HPC community and beyond. The members will be involved throughout the project from the initial phases to the test, validation and impact assessments. The DEEP-SEA members keep regular contact with the SAB in order to be informed of their technology roadmaps, get access to pre-release technologies, and ensure that developments in the project are in line with emerging trends in HPC.
Confirmed members of the SAB:
