The EU funded MANGO project, in which many core architectures for future HPC systems are explored, mixes high-end processors, graphics processors and FPGAs to meet most challenging application needs.
MANGO is an EU funded project under the H2020-FETHPC-2014 program. The project is coordinated by Universitat Politécnica de Valéncia (Spain).
Other partners are Centro Regionale Information Communication Technology (Italy), Politecnico di Milano (Italy), University of Zagreb (Croatia), PRO DESIGN Electronic GmbH (Germany), Thales Communications & Security S.A.S. (France), École Polytechnique Fédérale de Lausanne (Switzerland), EATON (France) and Philips Healthcare (Netherlands). This project aims to build a large heterogeneous High Performance Computing (HPC) prototype solution based on state-of-the-art high-end processors (CPUs), Graphics Processing Units (GPUs), low-power (ARM) processors and Field Programmable Gate Arrays (FPGAs). This prototype has a common interconnect structure to efficiently reach all components and innovative cooling and power monitoring systems.
MANGO (“explore Manycore Architectures for Next-GeneratiOn HPC systems") aims for a reconfigurable HPC system which meets the requirements of different applications which all have individual needs. The challenge is to find the most efficient combination of the available processing resources for each application. Performance, power-consumption and predictability are the criteria to be looked at (3P model). The usability shall be optimized by improving programmability, integration and Quality of Service (QoS). The system is made up by three parts:
The first part, Hardware, consists of general purpose nodes (GNs) made of blades with CPUs and GPUs as well as heterogeneous nodes (HNs) made of ARM cores and FPGAs. A common interconnect infrastructure allows to efficiently reach all elements. Secondly, the software stack includes a Linux-based OS, compilers and programming models for different processing resources and applications as well as a resource manager which analyzes the data to be processed and assigns the best combination of the processing cores. The resource manager is controlled by a program and by data from currently running processes such as temperature, power, process status, etc.
Finally, applications run real case studies to verify novel architecture. These case studies include transcoding applications, medical imaging applications as well as security, big data and communication applications.
Hardware description of the MANGO system
The general purpose nodes (GNs) of the MANGO system are represented by a scalable server with standard blades with Intel processors (CPUs) and GPUs. The hetero -geneous nodes (HNs) consist of single-chip Xilinx Zynq-7000 and Xilinx Virtex-7 and UltraScale FPGA modules from PRO DESIGN’s proFPGA product family. Zynq FPGAs have an embedded ARM processor and an FPGA part while Virtex devices are pure FPGA fabrics with highest possible gate count.
The FPGA modules can freely be plugged and mixed on the proFPGA quad Motherboard (and later on a dedicated chassis) which holds up to 4 such FPGA modules. Multiple such motherboards can be connected together to even larger systems. This modularity gives the flexibility to upgrade or downgrade the HN setup in steps of 1 FPGA module and allows scaling the HN system up to an almost unlimited number of FPGA modules which allows to almost creating any type and size of processing unit cluster.
Dieser Beitrag ist urheberrechtlich geschützt. Sie wollen ihn für Ihre Zwecke verwenden? Kontaktieren Sie uns über: support.vogel.de/ (ID: 44308847 / English)