header news2

This website uses cookies to manage authentication, navigation, and other functions. By using our website, you agree that we can place these types of cookies on your device.

View Privacy Policy

Supercomputing research project DEEP-EST recently ended, showing the effectiveness of a modular approach towards future exascale systems.

Jülich/Munich, April 15, 2021 – The project “DEEP Extreme Scale Technologies”, funded by the European Union, has come to a successful conclusion. DEEP-EST started in 2017 to develop a modular architecture for High-Performance Computing (HPC). Leveraging a co-design approach with real-life applications, the DEEP-EST project paved the way for future exascale systems that will bring supercomputing to a new level.


One major goal in HPC is to build exascale systems – systems that are capable of calculating 1018 FLOPS (Floating Point Operations per Second). In Europe, different research projects work on this challenge, with different approaches and different focal points. DEEP (Dynamical Exascale Entry Platform) projects, a series on EU funded research projects on future supercomputing, are part of this community. On March 31, 2021, the latest DEEP project was completed successfully: DEEP-EST.

DEEP-EST (EST: Extreme Scale Technologies) started to develop the prototype of a modular architecture for future exascale HPC systems. This system-level concept dynamically adapts to the needs of different types of applications. Not all workloads in today’s supercomputing are the same: Many combine data analytics, sheer number-crunching, machine learning and more. It is hard to optimise system usage, energy efficiency, and execution performance with a single architecture. A modular system enables running workloads on the hardware that’s best suited for the specific case, adapting the system to the needs of next generation of applications.


Modular Supercomputing Architecture

The Modular Supercomputing Architecture (MSA) developed in DEEP-EST comprises three different compute modules interconnected by a high-performance fabric:

  1. Cluster Module (CM): The Cluster Module is a general-purpose cluster with 50 nodes, using dual Intel® Xeon® Scalable processors with high single-thread performance and warm water cooling.
  2. Extreme Scale Booster (ESB): The Extreme Scale Booster provides 75 nodes, each equipped with a single Intel® Xeon® Scalable processor and one NVIDIA® Tesla® V100 GPGPU.
  3. Data Analytics Module (DAM): The 16 Data Analytics Modules are each equipped with dual Intel® Xeon® Scalable processors, one NVIDIA® Tesla® V100 GPGPU, and one Intel® Stratix® 10 FPGA per node.

All modules are connected with a high-performance network. Besides standard Infiniband for CM and Ethernet for DAM, the project leverages a new low latency fabric developed by DEEP-EST partner Extoll. The Fabri3 network provides six 100Gbit/s links per node and enables 3D grids and tori to be configured.


Realistic Co-Design Approach

On the software side, the focus was on ease of programming, resource management and job scheduling. The DEEP-EST software stack allows defining optimal resource allocations for each combination of workloads, supports adaptive scheduling and, enables dynamical reservation of the resources. To prove this concept with real-world codes, a co-design approach was chosen. Six different application teams from European partners ported their codes to the DEEP-EST system to support development and validate the technologies. Additionally, an Early Access Programme (EAP) enabled academic users from outside the project to intensively test the new architecture with their applications.


The DEEP-EST prototype follows a modular design that adapts dynamically to the requirements of different workloads.


DEEP-EST was completed successfully on March 31, 2021. The hardware prototype is located at the Jülich Supercomputing Centre and will be used for further projects. The DEEP series of projects goes on with DEEP-SEA (Software for Exascale Architectures): DEEP-SEA aims to deliver the programming environment for future European exascale systems. The new project started on April 1, 2021 with a planned term of 3 years.