Barcelona Supercomputing Center: Horizon 2020 project concludes with major contributions to energy efficiency in supercomputing

19 Jan 2021 | Network Updates | Update from Barcelona Supercomputing Center
These updates are republished press releases and communications from members of the Science|Business Network

After three years of research, the BSC coordinated project LEGaTO concludes with major contributions to the main goal of energy efficiency in future HPC systems. The technologies developed in the project targeted one order of magnitude energy savings for five widely applicable use cases. In addition, six BSC European-funded projects will build on the outcomes of LEGaTO.

The project has attained the following results in the five use cases of healthcare, smart home, smart city, machine learning and secure IoT gateway:

  • The first use case is a pilot healthcare application. This healthcare application is biomarker discovery, which aims to identify potential disease markers. The need to speed-up biomarker analysis is more relevant now than ever before in the light of the recent health emergency that the world is facing. In particular, the use of FPGAs resulted in a phenomenal 822x speedup in this use case, which enables a new world of biomarker analysis.
  • LEGaTO’s target is to achieve large energy savings for the smart home use case so that assisted living scenarios such as the LEGaTO Smart Mirror application become possible to deploy in senior citizen homes. The use of shared-memory programming style on distributed GPUs led to 10x energy savings in the SmartMirror. At the start of the project, the SmartMirror required 600 Watts to operate, severely limiting its potential for field deployment. By the end of LEGaTO, SmartMirror required only 50 Watts thus enabling its deployment.
  • The smart city use case on operational urban pollutant dispersion modelling had a 7x gain in energy efficiency thanks to the use of GPUs.
  • Up to 16x gain in energy efficiency and performance was achieved in the machine learning use case, devoted to automated driving and graphics rendering, using the LEGaTO optimizer.
  • The Secure IoT Gateway was vital to simplify the complexity of communication of local devices to a network, and it supported the above mentioned use cases to achieve their goals by reducing the complexity of security.

“Besides technical and management leadership of the project, the BSC LEGaTO team has substantially contributed towards the LEGaTO goal of an order of magnitude energy-savings on heterogeneous compute devices leveraging the OmpSs write-once deploy-everywhere task based programming model and associated dataflow runtime” said Osman Unsal, Group Manager for the Department of Computer Architecture for Parallel Paradigms at the BSC and coordinator of LEGaTO.

Independently from the above use case optimizations, BSC researchers also achieved one order of magnitude energy savings through undervolting on FPGA accelerators in deep neural network applications. Additionally, the FTI (Fault Tolerance Interface) checkpointing solution was expanded to GPU and FPGA devices to support seamless reliability in heterogeneous HPC compute devices. Other accomplishments included overhauling the software stack to embed task-based applications in secure enclaves; and leveraging OmpSs@cluster to automatically partition the application in multiple nodes to run the smart mirror application with substantial energy efficiency.

LEGaTO has provided excellent outcomes from the OmpSs perspective, including the support for CUDA and OpenCL devices when they are available in the same machine. This effort includes combinations like Intel FPGAs and Nvidia CUDA GPUs. Large improvements have also been achieved on the support that OmpSs now has for Xilinx FPGAs (aka OmpSs@FPGA), and the implementation of the directory cache for GPUs and FPGAs within OmpSs-2.

Additionally, the experience with LEGaTO’s partners Technion and Maxeler has been successful in achieving the integration of the DFiant DSL kernels, and the support for Maxeler kernels at the runtime level. In collaboration with Maxeler, the development of the task graph analysis tool is very important for future application developments on FPGAs, to know specifically which parts of the application and with what shape can be better exploited on the FPGA. Last but not least important, it has been very useful to investigate the use of SGX for secure tasking, showing that OmpSs tasks can be executed securely in SGX Enclaves.


LEGaTO System Overview

In addition to other initiatives already using LEGaTO-developed technologies, six European BSC projects will continue the development of the results achieved in LEGaTO:

  • eProcessor, coordinated by BSC and with other three LEGaTO partners in the consortium, will leverage BSC’s work on energy efficient checkpointing (FTI) with the aim to deliver the first completely open source European full stack ecosystem based on RISC-V technology.
  • OmpSs@FPGA functionality is used and extended in the projects AMPERE (to work in real-time domains), EuroEXA (to support the new prototypes), and MEEP (to develop an FPGA-based emulation platform based on European-developed IP).
  • OmpSs@Cluster is used and further extended under the EPEEC project, which tackles upcoming overwhelmingly-heterogeneous exascale HPC systems.
  • FPGA undervolting techniques are being implemented and demonstrated in the Tetramax project LV-EmbeDL, led by BSC and the spin-off EmbeDL.

Lastly, LEGaTO also contributed to the report “Towards Resilient EU HPC Systems: A Blueprint”, which aims to spearhead a Europe-wide discussion on HPC system resilience and to help the European HPC community define best practices for resilience.

LEGaTO’s software in GitHub is available here:

About LEGaTO

The LEGaTO (Low Energy Toolset for Heterogeneous Computing) project is funded by the European Commission with a budget of more than €5 million and will last 3 years from its beginning on 1 December 2017. The partners of the project are Barcelona Supercomputing Center (BSC, Spain), Universität Bielefeld (UNIBI, Germany), Université de Neuchâtel (UNINE, Switzerland), Chalmers Tekniska Högskola AB (CHALMERS, Sweden), Machine Intelligence Sweden AB (MIS, Sweden), Technische Universität Dresden (TUD, Germany), Christmann Informationstechnik + Medien GmbH & Co. KG (CHR, Germany), Helmholtz-Zentrum für Infektionsforschung GmbH (HZI, Germany), TECHNION - Israel Institute of Technology (TECHNION, Israel), Maxeler Technologies Limited (MAXELER, United Kingdom).

Further information:

This article was first published on 18 January by BSC.

Never miss an update from Science|Business:   Newsletter sign-up