optimization & scalability€¦ · programming models & tools cloud computing optimization...
TRANSCRIPT
Investigating the relationship between complex appli ca tion domains, evolving hardware, and programming paradigms on large-scale systems.
Optimization & Scalability
4HLRS
HLRSHigh Performance Computing Center Stuttgart
The High Performance Compu-ting Center of Stuttgart (HLRS) of the University of Stuttgart is the first National Supercompu-ting Center in Germany and is offering services to both acade-mic users and industry. Apart from the operation of supercom-puters HLRS activities include teaching and training in distribu-ted systems, software enginee-ring and programming models, as well as development of new technologies. HLRS is an acti-ve player in the European rese-arch arena with special focus on Scientific Excellence and Indust-rial Leadership initiatives.
Our Network: HLRS is tightly connected to academia and indus-try through long term partners-hips with global market players such as Porsche and T-Systems, as well as worldwide companies, HPC centres and Universities. Particular attention is given to collaboration with Small and Me-dium Enterprises (SMEs).
Our Infrastructure: HLRS ope-rates a CRAY XC40 supercom-puter (peak performance > 7 Pe-taFlops), as well as a variety of smaller systems, reaching from clusters to cloud resources.
ProgrammingModels & Tools
CloudComputing
Optimization & Scalability
Energy Efficiency
Exascale Computing
Services
Big Data, Analytics & Management
Visualization
Featured Topics
5HLRS
Director HLRSProf. Dr. Michael Resch
Our Experience: HLRS has been at the forefront of regional, nati-onal and European research and innovation over the last 20 years. During this time, HLRS has parti-cipated successfully in more than 90 European research and inno-vation projects.
Our Expertise: HLRS is a lea-ding innovation center, applying software engineering methods to HPC and Cloud for the bene-fit of multiple application domains such as automotive, engineering, health, mobility, security, and energy. Thanks to the close inter-action with industry, the center’s capabilities and expertise sup-port the whole lifecycle of simu-lation covering research aspects, pre-competitive development and preparation for production. The HLRS innovation group, which actively examines and tests new technologies, can bring into pro-jects expertise on leading edge technologies hardware and scale up data analysis techniques.
6Optimization & Scalability
Optimization & Scalability
The use of supercomputers is mainly motivated by two goals: solving large-scale computations not possible on smaller systems and shortening times to solution where time is the critical factor. To achieve both goals, simulation codes have to be highly optimized for supercomputer hardware. Required optimizations must tar-get different aspects. The first is the efficient use of all single hardware components of the supercomputer. Therefore, specialists modify codes to utili-ze hardware characteristics to their full capabilities. The second optimization aspect targets the scalability of simulations, such as the capability of using growing amounts of compute resources in parallel. Thus, it is important that computational work is distributed across compute resources evenly so that a good load balance is achie-
ved, and that the communication between single compute nodes of a supercomputer does not domi-nate the simulation. With these goals in mind, HLRS and its researchers are dedica-ted to investigate the relationship between complex application do-mains as well as evolving hard-ware and programming paradig-ms through cutting-edge scientific projects such as POP and Phan-tom. To this end, HLRS staff not only researches new optimization strategies in established applica-tion domains, such as structural mechanics, fluid dynamics, che-mical processing or quantum phy-sics, but also tries to address the complexity of modern computing hardware—especially as it relates to moving toward exascale sys-tems—by developing and testing new parallel programming models and languages.
7Optimization & Scalability
Project Overview
POP - Performance Optimisation and Productivity (A Centre of Excel-lence in Computing Applications)
Smart-DASH - Smart Data Struc-tures and Algorithms with Sup-port for Hierarchical Locality
PHANTOM - Cross-layer and mul-ti-objective Programmin g approacH for next generAtioN heTerogeneous parallel cOMputing systems
EXASOLVERS - Extreme Scale Solvers for Coupled Problems
Page 8
Page 10
Page 12
Page 14
CATALYST - Combining HPC and High Performance Data Analytics for Academia and Industry
Page 16
Page 18Beam Me - Developing speed-up methods from Applied Mathematics and Com-puter Science for the optimizati-on of energy systems models
8POP
POP
High performance computing is a fundamental tool for the progress of science and engineering and as such for the economic competiti-veness. The growing complexity of parallel computers is leading to a situation where code owners and users are not aware of the de-tailed issues affecting the perfor-mance of their applications. The result is often an inefficient use of computing resources. Code de-velopers often do not have suffi-cient insight in its detailed causes in order to address the problem properly. The objective of POP is to operate a Center of Excellence in perfor-mance optimisation and producti-vity and to share our expertise in the field with the computing com-munity. In particular, POP will offer the service of precisely assessing the performance of computing applications of any sort, from a few hundred to many thousands of processors. Also, POP will show users the specific issues affecting the performance of their code and the best way to allevia-
Performance Optimisation and Productivity (A Centre of Excellence in Computing Applications)
te them. POP will target and offer such services to code owners and users from all domains, including infrastructure operators, acade-mic and industrial users.The estimated population of such applications in Europe is 1500
9POP
and within the project lifetime POP has the ambition of serving over 150 such codes. The Added Value of POP’s services is the sa-vings generated in the operation and use of a code, which will re-sult in a significant Return on In-vestment (fixing a code costs less than running it below its optimal levels) by employing best-in-class services and release capacity for resolving other priority issues. POP will be a best-in-class centre. By bringing together the Europe-an world-class expertise in the area and combining excellent aca-demic resources with a practical, hand-on approach, it will improve the access to computing appli-
ContactDr. José Gracia
Christoph Niethammer
Phone: +49 (0) 711/ 685-87208
+49 (0) 711/ 685 87203
E-Mail: [email protected]
Further Informationwww.pop-coe.eu
cations, thus allowing European researchers and industry to be more competitive.
Project Partners � Barcelona Supercomputing Center, Spain
� Numerical Algorithm Group,UK � RWTH Aachen � HLRS � Teratec, FR � Forschungszentrum Jülich
Project Information � Funding Organisation: EU-H2020
� Runtime: 10.2015 - 03.2018
10Smart-DASH
Smart-DASH
Smart-DASH is a collaborative research project funded for 3 years by the German Research Foundation (DFG) as part of the priority programme „Software for Exascale Computing – SPPEXA“ (2013-2019). Smart-DASH is a follow-up to the project “DASH”. Smart-DASH aims to ease the efficient programming of future supercomputing systems for da-ta-intensive applications. These systems will be characterized by their extreme scale and a mul-ti-level hierarchical organization. Smart-DASH adopts the concept of Partitioned Global Address Spa-ce (PGAS) and provides the C++ template library “DASH” for distri-buted containers such as multidi-mensional arrays, lists and hash tables. Unlike other PGAS appro-aches, DASH will furthermore al-low a developer either to control and explicitly take advantage of the hierarchical data layout of glo-bal data structures, or, to rely on
Smart Data Structures and Algorithms with Support for Hierarchical Locality
smart handling of data-locality by the runtime system. The runtime and C++ template library will be extended to support a task-ba-sed execution model which al-lows to specify data-dependencies amongst distributed tasks on the global address space. We will develop ‘smart’ data struc-tures that capture frequently en-countered application scenarios to enable a productive transition onto new hardware platforms and assist in code modernization ef-forts. To address fault-tolerance and reliability, we will explore con-cepts for the redundant storage of data items and with the DASH data dock we will explore the usa-ge of the PGAS approach in ge-neral and NVRAM in particular for the coupling of applications. Several case studies will explore the utility of these new features in the context of important scientific problem classes.
11Smart-DASH
Within the project HLRS will lead in the development of the core runtime system, in particular the tasking and communication backend, respectively. Additio-nally, HLRS will contribute to the user-facing C++ template library, in particular when related to the tasking model.
ContactDr. José Gracia
Phone: +49 (0) 711/ 685-87208
E-Mail: [email protected]
Further Informationwww.dash-project.org
www.sppexa.de
Project Partners � LMU Munich, MNM Team � IHR, University of Stuttgart � TU-Dresden, ZIH � HLRS (Germany)
Project Information � Runtime: 08.2016 – 07.2019 � Funding Organisation: DFG
12PHANTOM
Cross-layer and multi-objective Programmin g approacH for next generAtioN heTerogeneous parallel cOMputing systems
PHANTOM
without considering the required development efforts. Furthermo-re, the challenge is getting even more complex when taking the assumption that all this hardware is working in a collaborative way within the common infrastruc-ture (which might range from the “server-on-chip” to the cluster-like distributed systems). So the term „Cloud“ has become common for such restrictive hardware do-mains like embedded systems as well.
Modern software applications have to cope with a great variety of hardware platforms, ranging from the commodity Intel’s and low-power ARM’s CPUs to the ac-celerators like NVIDIA’s GPU, re-configurable-logic systems like Xi-linx’s FPGA or dedicated systems like Movidius’ Myriad2. Consequently, the selection of the best possible platform for the specific application, which has to fulfil the user-imposed functional and non-functional requirements, is a very challenging task, even
13PHANTOM
ContactDr. Alexey Cheptsov
Phone: +49 (0) 711 / 685-60470
E-Mail: [email protected]
Further Informationwww.phantom-project.org
PHANTOM is a EU-H2020 pro-ject (under grant agreement No. 688146) that has the ambition to provide a platform, which allows the components constituting the application (within the specified control- and data-flow) to be exe-cuted in heterogeneous, parallel, and distributed hardware environ-ments (see the figure on the left) without any hardware-specific ad-aptation of the source code.
Role of HLRS � Use case provider: Dynamic simulation of aero- and gas-dy-namic processes in real-time
� Technology provider: Monito-ring and Resource Manage-ment Framework, Parallelisa-tion Toolkit
Project Partners � The Open Group (UK) � Easy Global Market (France) � GMV (Portugal) � Intecs (Italy) � HLRS (Germany) � University of York (UK) � Unparallel Innovation (Portu-gal)
� WINGS ICT Solutions (Greece)
Project Information � Runtime: 12.2015 – 11.2018 � Funding Organisation: EU H2020
14EXASOLVERS
Optimization and inverse prob-lems (Trier University)By means of inverse problems, it is possible to determine simu-lation parameters that can’t be measured due to e.g. subminia-ture structures, inaccessible en-vironments, etc. However, usage of the aforementioned methods for optimization and inverse prob-lems provides further potential to use exascale systems efficiently.
Uncertainty quantification (RWTH Aachen)The group from Aachen uses low rank hierarchical tensors to quan-tify uncertainties of simulations, which allows to further increase the amount of parallelism that can be used efficiently.
Exascale systems will be charac-terized by bil lion-way parallelism. Computing on such extreme sca-les requires suitable methods. The ExaSolvers 2 project hence investigates such methods:
Parallel adaptive multigrid (G-CSC, University Frankfurt)The multigrid method is of opti-mal complexity and hence suited for extreme scale parallelism. The group from Frankfurt develops their own parallel multigrid frame-work ug4 which also adapts mesh resolution in order to increase the solution efficiency.
Time parallelization (ICS, USI Lugano)In transient simulations, not only the simulation domain but also the investigated time frame can be divided and handled on diffe-rent execution units in parallel in order to efficiently use the massi-ve parallelism of future systems.
EXASOLVERSExtreme Scale Solvers for Coupled Problems
15EXASOLVERS
Energy efficiency (HLRS, University Stuttgart)Due to their massive parallelism, Exascale systems will require huge amounts of energy. We hence in-vestigate methods to increase the energy effi ciency of such systems on multiple levels, i.e. algo rithmic efficiency, efficiency-aware imple-mentation as well as adaption of hardware parameters (e.g. redu-cing the CPU’s core frequency, known as Dynamic Voltage and Frequency Scaling).
A collaboration with the Japane-se ADVENTURE project has been established in order to deploy the
ContactBjörn Dick
Dr. Ralf Schneider
Phone: +49 (0) 711/ 685-87189
+49 (0) 711/ 685-87236
E-Mail: [email protected]
Further Informationwww.hlrs.de/about-us/research/
current-projects/exasolvers
performance engineering experti-se of the project partners from Japan on codes developed by the ExaSolvers 2 project. In return, ADVENTURE is going to integra-te our methods into their frame-work.In order to assess the developed methods, a simulation of trans-dermal drug delivery through the human skin with detailed resolu-tion of the lipid scale is used as benchmark application.
Project Information � Runtime: 05.2016 - 04.2019 � Funding Organisation: DFG
16CATALYST
Combining HPC and High Performance Data Analytics for Academia and Industry
CATALYST
As the majority of today’s data analytics algorithms are oriented towards text processing (e.g. bu-si-ness analytics) and graph analy-sis (e.g., social network studies), we are further in need to evaluate existing algorithms with respect to their applicability for the enginee-ring domain. Thus, CATALYST will examine future concepts for both hardware and software. The first case study conducted in collaboration with Cray Inc. addresses the performance va-riations of our Cray XC40 sys-tem. Performance variability on HPC platforms is a critical issue with serious implications for the users: irregular runtimes prevent users from correctly assessing performance and from efficiently planning allocated machine time. Consequently, monitoring today’s IT infra-structures has actually be-come a big data challenge on its own. The analysis workflow used to identify the causes of runtime variations consists of three steps including different configuration parameters:
At the High Performance Com-puting Center Stuttgart (HLRS), customers tend to execute more and more data-intensive applica-tions. Since it no longer becomes feasible that data is processed and analysed manually by domain experts, HLRS and Cray Inc. have launched the CATALYST project to advance the field of data-intensi-ve computing by converging HPC and Big Data in order to allow a seamless workflow between com-pute-intensive simulations and data-intensive analytics. For that purpose, Cray Inc. designed the Urika-GX data analytics hardware, which supports Big Data techno-logies and furthermore, enhan-ces the analysis of semantic data. This system has been installed as an extension of Hazel Hen, the current HPC-flagship system of HLRS. The main objective of CATALYST is to evaluate the hardware as well as the software stack of the Ur-ika-GX and its usefulness with a particular focus on applications from the engineering domain.
17CATALYST
ContactMichael Gienger
Phone: +49 (0) 711 / 685-63824
E-Mail: [email protected]
With the help of this workflow, 470 so called „Victim“ applications have been identified that suffered from the particular behaviour of 3 „Aggressors“. Consequently, HLRS took this information and approached the responsible sta-keholders in order to optimise their applications in general. So not only the performance of these applications has been improved, but also the entire system perfor-mance in production.
Further Informationwww.hlrs.de/en/about-us/
research/current-projects/
data-analytics-for-hpc
Outlook � Big Data application evaluation � Close cooperation with part-ners from both, industry and academia
� Seamless integration of the Big Data system into our exis-ting HPC infrastructure
� Develop and evaluate practical case studies to advertise the solution
Project Information � Runtime: 10.2016 – 09.2019 � Funding Organization: Ministry of Science, Research and the Arts Baden-Württemberg
� Partners: HLRS, Cray Inc. & Daimler AG (associated)
18Beam Me
delling. Severe simplifications of the models regarding complexity are required to allow them to sol-ve in a reasonable amount of time – with significant influence on the validity of results and reliability of the models in general. Within this project the consorti-um of researchers from different research fields (system analysis, mathematics, operations rese-arch and informatics) develop new strategies to increase com-putational performance of ener-gy system models and to apply energy system models to high performance computing. Within the project, the High Performan-ce Computing Center Stuttgart (HLRS) supports the development of the ESM-Software with relevant algorithmic strategies in order to make sure the framework runs on HPC systems. The first ESMs have already been applied to two of Germany’s fastest supercom-puters – HLRS’ Hazel Hen and the Jülich Supercomputing Cen-ter’s JUQUEEN. The project aims at identifying efficient strategies
Developing speed-up methods from Applied Mathematics and Computer Science for the optimization of energy systems models
Beam Me
Energy system models (ESM) are widely used in research and in-dustry to analyze today’s and fu-ture energy systems and poten-tial pathways for the European energy transition. Current studies address future policy design, ana-lysis of technology pathways and the analysis of future energy sys-tems. To analyze these questions and support the transformation of today’s energy systems, ESM are required to become increa-singly complex in order to provide valuable quantitative insights for policy makers and industry. Espe-cially when analyzing uncertainty and integration of large shares of renewable energies, ESM requi-res a detailed implementation of the underlying electricity system. Due to the increasing complexity of the models, for research insti-tutions and industries all over Eu-rope applying ESM becomes more and more difficult, as boundari-es with regard to computational power of today’s decentralized workstations impose significant constraints to energy market mo-
19Beam Me
ContactDmitry Khabi
Phone: +49 (0) 711 / 685-65734
E-Mail: [email protected]
Further Informationwww.beam-me-projekt.de
and developing best-practice gui-de with the recommendations for the application of ESMs to high performance computing. The pro-ject consists of the following ob-jectives:
� Identification, implementation and comparison of strategies for optimizing computing time for high resolution energy sys-tem models.
� Standardization of parallelizati-on strategies for different ty-pes of large scale energy sys-tem models.
� Development and adaptation of mathematical algorithms for parallel solving of energy market models on distributed HPC platforms
� Development of new complex energy system models that cannot be solved today. The-se models address new rese-arch questions especially with regard to robust scenario de-velopment and decisions un-der uncertainty.
Project Information � Runtime: 12.2015 – 11.2018 � Funding Organisation: Federal Ministry for Economic Affairs and Energy
High Performance Computing Center Stuttgart (HLRS) University of Stuttgart Nobelstrasse 19 | 70569 Stuttgart | Germany Phone: +49 (0)711 / 685 87 269 Fax: +49 (0)711 / 685 87 209 Mail: [email protected] www.hlrs.de
Editor: Lena Bühler, Eric Gedenk, Dr. Bastian Koller Design: Janine Jentsch, Ellen Ramminger Picture Credits: Cover and Interior shot: Bohris Lehner for HLRS Back cover shot: Simon Sommer for HLRS © HLRS 2018