Lecture Notes in Computer Science 11887

Founding Editors

Gerhard GoosKarlsruhe Institute of Technology, Karlsruhe, Germany

Juris HartmanisCornell University, Ithaca, NY, USA

Editorial Board Members

Elisa BertinoPurdue University, West Lafayette, IN, USA

Wen GaoPeking University, Beijing, China

Bernhard SteffenTU Dortmund University, Dortmund, Germany

Gerhard WoegingerRWTH Aachen, Aachen, Germany

Moti YungColumbia University, New York, NY, USA

More information about this series at http://www.springer.com/series/7407

Michèle Weiland • Guido Juckeland •

Sadaf Alam • Heike Jagode (Eds.)

High PerformanceComputingISC High Performance 2019 International WorkshopsFrankfurt, Germany, June 16–20, 2019Revised Selected Papers

123

EditorsMichèle WeilandUniversity of EdinburghEdinburgh, UK

Guido JuckelandHelmholtz-Zentrum Dresden-RossendorfDresden, Sachsen, Germany

Sadaf AlamSwiss National Supercomputing CentreLugano, Ticino, Switzerland

Heike JagodeUniversity of Tennessee at KnoxvilleKnoxville, TN, USA

ISSN 0302-9743 ISSN 1611-3349 (electronic)Lecture Notes in Computer ScienceISBN 978-3-030-34355-2 ISBN 978-3-030-34356-9 (eBook)https://doi.org/10.1007/978-3-030-34356-9

LNCS Sublibrary: SL1 – Theoretical Computer Science and General Issues

© Springer Nature Switzerland AG 2019, corrected publication 2019, 2020Chapters “Sarus: Highly Scalable Docker Containers for HPC Systems” and “A Multitenant ContainerPlatform with OKD, Harbor Registry and ELK” are licensed under the terms of the Creative CommonsAttribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/). For further details seelicense information in the chapters.This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of thematerial is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,broadcasting, reproduction on microfilms or in any other physical way, and transmission or informationstorage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology nowknown or hereafter developed.The use of general descriptive names, registered names, trademarks, service marks, etc. in this publicationdoes not imply, even in the absence of a specific statement, that such names are exempt from the relevantprotective laws and regulations and therefore free for general use.The publisher, the authors and the editors are safe to assume that the advice and information in this book arebelieved to be true and accurate at the date of publication. Neither the publisher nor the authors or the editorsgive a warranty, expressed or implied, with respect to the material contained herein or for any errors oromissions that may have been made. The publisher remains neutral with regard to jurisdictional claims inpublished maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AGThe registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

A separate workshop day attached to the International Supercomputing Conference(ISC) High Performance, formerly known as the International SupercomputingConference, was first added to the technical program in 2015 under the leadership ofBernd Mohr (Forschungszentrum Jülich GmbH). Supported by the success of the pastfour years, ISC High Performance renewed and further extended the workshop pro-gram in 2018, under the leadership of Sadaf Alam (Swiss National SupercomputingCenter) and Heike Jagode (University of Tennessee Knowville). Michele Weiland(The University of Edinburgh) and Guido Juckeland (Helmholtz-ZentrumDresden-Rossendorf) joined the team as proceedings chair and deputy chair, respec-tively, and managed the organization of the proceedings for the workshops.

The 21 workshops at ISC High Performance 2019 provided a focused, in-depthplatform with presentations, discussion, and interaction on topics related to all aspectsof research, development, and application of large-scale, high-performance experi-mental and commercial systems. Workshop topics included HPC computer architectureand hardware; programming models, system software, and applications; solutions forheterogeneity, reliability, and power efficiency of systems; virtualization and con-tainerized environments; big data and cloud computing; artificial intelligence; as wellas international collaborations. Workshops were selected with a peer-review process byan international committee of 16 experts in the field from Europe, the United States,and Asia.

Since 2016, ISC High Performance has provided a platform for workshops withtheir own call for papers and individual peer-review process through an early deadlinein December 2017. In all, 17 workshop proposals were submitted before this deadlinefrom organizers all over the world; 9 full-day and 8 half-day workshops were acceptedafter a rigorous review process in which each proposal received 3 reviews. Addition-ally, each reviewer was given the possibility to discuss all the submissions. Workshopswithout a call for papers were invited to submit their proposals in February 2019. Forthis second deadline, a further four workshop proposals were accepted by thecommittee following the same rigorous peer-review process as for workshops withproceedings.

The 21 workshops were held on Thursday, June 20, 2019, at the Frankfurt MarriottHotel. The workshop proceedings volume collects all the accepted papers of theworkshops from the call for papers. Each chapter of the book contains the accepted and

revised papers for one of the workshops. For some workshops, an additional prefacedescribes the review process for that workshop and provides a summary of theoutcome.

June 2019 Michèle WeilandGuido Juckeland

Sadaf AlamHeike Jagode

The original version of the book was revised: For detailed information see correctionchapter. The correction to the book is available at https://doi.org/10.1007/978-3-030-34356-9_50

vi Preface

Organization

Program Committee

Christoph Angerer NVIDIA, USAHartwig Anzt University of Tennessee, USAAbdulrahman Azab University of Oslo, NorwayShane Canon Lawrence Berkeley National Lab, USASunita Chandrasekaran University of Delaware, USAValeriu Codreanu SURFsara, The NetherlandsJens Domke RIKEN Center for Computational Science (R-CCS),

JapanSteffen Frey Visualisierunsginstitut der Universität Stuttgart,

GermanySiddhartha Jana Intel, USAGuido Juckeland Helmholtz-Zentrum Dresden-Rossendorf (HZDR),

GermanyMozhgan Kabiri Chimeh The University of Sheffield, UKJulian Martin Kunkel University of Reading, UKScott Lathrop Shodor, USAJay Lofstead Sandia National Laboratories, USAPiotr Luszczek University of Tennessee Knoxville, USADavid Martin Argonne National Laboratory, USAMichael Mascagni Florida State University, USAAnastassios Nanos National Technical University of Athens, GreeceJan F. Nygård The Cancer Registry of Norway, NorwayDirk Pleiter Jülich Supercomputing Centre,

University of Regensburg, GermanySwaroop Pophale Oak Ridge National Laboratory, USAMichael Ringenburg Cray Inc., USAMartin Schulz Technical University of Munich, GermanyHari Subramoni The Ohio State University, USANitin Sukhija Slippery Rock University of Pennsylvania, USAThomas Theussl King Abdullah University of Science and Technology,

Saudi ArabiaMohamed Wahib AIST, TokyoTech OIL, JapanMichele Weiland The University of Edinburgh, UKAndrew Younge Sandia National Laboratories, USA

Short Papers

Preface to the First International Workshopon Legacy Software Refactoring

for Performance

Artur Podobas1, Jens Domke1, and Wahib Attia Mohamed2

1 RIKEN Center for Computational Science, Kobe, Japan2 National Institute of Advanced Industrial Science and Technology,

Tokyo, Japan

1 Workshop Background and Description

The First International Workshop on Legacy Software REFACtoring for Performance(REFAC’19) is dedicated to the much needed shift in focus from hardware to softwareto achieve performance gains. Modernizing hardware has too long been the primarymethod of accelerating legacy software, and close to half of the expected performanceimprovement in legacy codes can be attributed to improve processor technology. Morethan half of this improvement was based on Moore’s law and its observation thattransistors will continue to become smaller every few (originally two) years. Theremaining hardware improvements came from architectural innovations, such as deepercache hierarchies, the migration to more exotic architectures (e.g. GPUs), or theutilization of larger and wider vector-units (SIMD), as well as scaling the HPC systemsup by giving them more processors and cores. Unfortunately, we are no longer seeingthe consistent technology scaling that Gordon Moore observed. Instead, the technologyscaling has significantly slowed down, and is expected to continue only for a few moreyears. Consequently, in the so-called Post-Moore era, the “performance road” forksthree-ways, yielding the following alternatives: (i) architectural innovations willattempt to close the performance gap, and an explosion of diverging architecturestailored for specific science domains will emerge, (ii) alternative materials and tech-nologies (e.g. non-CMOS technologies) allow the spirit of Moore’s law to continue fora foreseeable future, or (iii) we abandon the von-Neumann paradigm together andmove to a neuromorphic or quantum-like computer (which, in time, might or might notbecome practical). Independent of which direction we will end up taking in the future,the following will hold: Software and algorithmic optimization will be transferableto the first two out of the three identified directions. It is thesearchitecture-oblivious software optimizations that are the primary scope of theproposed workshop.

The workshop offered a forum for technical contributions, invited presentations,and discussions in the following areas:

– All types of general-purpose processor legacy-software optimizations for HPC

– Changes to (collective) communication algorithms or implementations to enable theuse of different numerical methods (e.g.: Lagrangian vs. Eulerian)

– Accelerating of pre-/post-processing in scientific workflows or axillary tools– Improved maintainability/performance through existing production libraries– Revisiting and applying modern compiler (flag) techniques, performance analysis

tools, moderate usage of OpenMP, etc., for performance gains– Manual code refactoring, such as loop transformations or changing data structures,

to acknowledge the shifting ratio in memory vs. compute capabilities of modernarchitectures, and

– Using mixed or adaptive precision wherever possible.

2 Technical Paper Reviewing Process

The workshop adopted a single-blind reviewing process and was reviewed by thetechnical program committee (TPC). The TPC consisted of ten (excluding organizers)international experts from workshop-relevant domains: Andreas Knüpfer (TU Dresden),Anshu Dubey (ANL), Barna Bihari, (LLNL), Bernd Mohr (Jülich SupercomputingCentre), Dali Wang (ORNL), Daniel Molka (German Aerospace Center), Didem Unat(Koç University), Hisashi Yashiro (RIKEN R-CCS), Saurabh Chawdhary (ANL), andSeyong Lee (ORNL).

The workshop organizers partook in the reviewing process. Each technical papercontribution was reviewed by at least four TPC members, by sampling from ourreviewer pool the most knowledgeable experts for the research area of the paper. Thereview process resulted in two accepted technical papers:

1. “Asynchronous AMR on Multi-GPUs”, M.N. Farooqi, T. Nguyen, W. Zhang,A. Almgren, J. Shalf and D. Unat

2. “Modernizing Titan2D, a parallel AMR Geophysical Flow Code to SupportMultiple Rheologies and Extendability”, N.A. Simakov, R.L. Jones-Ivey,A. Akhavan-Safaei, H. Aghakhani, M.D. Jones and A.K. Patra

which were allowed to make minor changes to address the reviewers’ feedback.

3 Workshop Summary

The well-attended workshop took place in the Frankfurt Marriott Alabaster 2 confer-ence room on Thursday, June 20th, and featured (apart from the two accepted papers) akeynote by Benoit Marchand (NYU) and two invited speakers, reporting their expe-riences and approaches for refactoring legacy software.

The presentations were performed in an interactive style, where the audience could(and did) ask questions at any time. This forum-like format encouraged discussions andinformation exchange throughout the workshop and helped contributing to the successof the workshop. Participants continued to discuss even during the breaks, signifyingthe importance and timeliness of the workshop. The end of the workshop featured a30-min free-form debate, where workshop participants actively discussed emerging

xii A. Podobas et al.

issues ranging from algorithms to middleware and runtime systems. This free-formdiscussion successfully brought different researcher and practitioners together, somewhom continued the conversation even after the workshop ended. We and theparticipants are looking forward to the next REFAC’20, and hope this year’s workshopkickstarts a broader effort across our HPC community to modernize and speed-up ourdated production codes.

Preface to the First International Workshop on Legacy Software Refactoring xiii

P^3MA Workshop 2019

1 Workshop Summary

The 4th International Workshop on Performance Portable Programming models forManycore or Accelerators (P^3MA) co-located with ISC 2019 was held at Frankfurt,Germany on June 20. The workshop solicited papers on topics covering feature sets ofprogramming models (including, but not limited to, C++, OpenCL, OpenMP,OpenACC, and various DSLs), their implementations, and experiences with theirdeployment in HPC applications on multiple architectures, performance modeling andevaluation tools, asynchronous task and event-driven execution/scheduling. Wereceived 3 submissions in total. All submitted manuscripts were peer reviewed by atleast 3 reviewers. The review process was single blind, i.e., authors were known toreviewers. Submissions were evaluated on correctness, originality, technical strength,significance, quality of presentation, and interest and relevance to the conference. Wechose 3 papers to be published in the workshop proceedings, Springer-Verlag LectureNotes in Computer Science (LNCS) volumes. The workshop was held for a full day onJune 20 at ISC and brought together researchers, vendors, users and developers tobrainstorm aspects of heterogeneous computing and its various tools and techniques.Around 45 attendees were present. All of the accepted papers were presented at theworkshop with topics ranging from using high-level programming models forheterogeneous systems to experiences porting legacy code to accelerators. The work-shop organizers also conducted a timely panel moderated by Sunita Chandrasekaranthat brought together expert developers from Alpaka (Sergei Bastrakov, HZDR)Kokkos (Christian Trott, Sandia National Lab), SYCL (Morris Hafner, Codeplay),OpenMP (Johannes Doerfert, Argonne National Lab) and OpenACC (Michael Wolfe,PGI/NVIDIA). The panel focused on how an application developer could make aninformed choice about using a programming model when there is more than a fewavailable. The first paragraphs that follows a table, figure, equation etc. does not havean indent, either.

2 Organizing Committee

Steering Committee

Matthias Muller RWTH Aachen University, GermanyBarbara Chapman Stony Brook University, USAOscar Hernandez ORNL, USADuncan Poole OpenACC

Torsten Hoefler ETH, ZurichMichael Wong Codeplay Software Ltd., CanadaMitsuhisa Sato University of Tsukuba, JapanMichael Klemm OpenMPKuan-Ching Li Providence University, Taiwan

Program Chair(s)

Sunita Chandrasekaran University of Delaware, USASwaroop Pophale Oak Ridge National Laboratory, USAArghya Chatterjee Georgia Institute of Technology, USA

Program Committee

Ferrol Aderholdt Middle Tennessee State University, USASridutt Bhalachandra Argonne National Laboratory, USAKyle Friedline University of Delaware, USAAxel Huebl Technische Universität Dresden, GermanyAdrian Jackson The University of Edinburgh, UKGabriele Jost NASA Ames Research Center/Supersmith, USAAndreas Knüpfer Technische Universität Dresden, GermanyJohn Leidel Texas Tech University, USAKelvin Li IBM, CanadaPiotr Luszczek University of Tennessee Knoxville, USANaoya Maruyama Lawrence Livermore National Laboratory, USAJohn Pennycook Intel, USASuraj Prabhakaran Intel GmbH, GermanyRobert Searles University of Delaware, USARay Sheppard Indiana University Bloomington, USAShuaiwen Song Pacific Northwest National Laboratory, USAXiaonan Tian NVIDIA, USAAntonino Tumeo Politecnico di Milano, ItalyVeronica Larrea Vergara Oak Ridge National Laboratory, USACheng Wang University of Houston, USARengan Xu Dell EMC, USA

P^3MA Workshop 2019 xv

4th International Workshop on In SituVisualization (WOIV’19)

Steffen Frey1, Peter Messmer2 and Thomas Theußl3

1 University of Stuttgart2 NVIDIA3 KAUST

1 Introduction

Large-scale HPC simulations with their inherent I/O bottleneck have made in situ anessential approach for data analysis. In situ coupling of analysis and visualization to alive simulation circumvents writing raw data to disk. Instead, data abstracts are gen-erated that capture much more information than otherwise possible.

The “Workshop on In Situ Visualization” series provides a venue for speakers toshare practical expertise and experience with in situ visualization approaches. This 4thedition of the workshop, WOIV’19, took place as a full-day workshop on 20 June 2019in Frankfurt, Germany, after two half-day workshops in 2016 and 2017 and a full-dayworkshop in 2018. The goal of the workshop in general is to appeal to a wide-rangingaudience of visualization scientists, computational scientists, and simulation develop-ers, who have to collaborate in order to develop, deploy, and maintain in situ visual-ization approaches on HPC infrastructures.

For WOIV’19 we again also encouraged submissions on approaches that did notlive up to their expectations. With this, we expected to get first-hand reports on lessonslearned. Speakers should detail if and how the application drove abstractions or otherkinds of data reductions and how these interacted with the expressiveness and flexi-bility of the visualization for exploratory analysis or why the approach failed.

2 Organization of the Workshop

The workshop content was built on two tracks:

Invited talks experts in the field were invited to share their thoughts and insightsResearch paper presentations authors were required to submit a full paper before the

workshop, which was then reviewed for inclusion in the conference proceedings

After the submission deadline in mid-May 2019, five manuscripts were submitted.Having the time of a full-day workshop at their disposal, the organizing committee wasable to select four of these five submissions for presentation at the workshop. After afull review cycle by an international program committee after the workshop

presentations, all four papers were selected for inclusion in these conference pro-ceedings. Additionally, five internationally recognized researchers agreed to each givean invited talk at the workshop. Slides for all presentations can be downloaded from theworkshop web page at http://woiv.org.

2.1 Organizing Committee

Steffen Frey University of StuttgartPeter Messmer NVIDIAThomas Theußl KAUST

2.2 Programm Committee

Hadrien Calmet BSC-CNSJose Camata Federal University of Juiz de ForaHank Childs University of OregonJens Henrik Goebbert Jülich Supercomputing CentreSamuel Li National Center for Atmospheric ResearchKenneth Moreland Sandia National LaboratoriesBenson Muite University of TartuGuido Reina University of StuttgartJoachim Pouderoux KitwareTom Vierjahn Westphalian University of Applied SciencesMax Zeyen Kitware

3 Workshop Summary

3.1 Invited Talks

Burlen Loring and Silvio Rizzi presented the in situ visualization framework SENSEI,which addresses the challenge of toolchain dependent application instrumentation: Inorder to use in situ visualization the HPC applications needs to be instrumented toexpose the simulation data to the visualization application. So far, no single standardinterface has emerged, requiring the developers to instrument their applications withdifferent interfaces depending on the targeted in situ visualization toolchain. SENSEItries to consolidate these different APIs, offering a single mechanism for applicationinstrumentation independent of the selected visualization toolchain. This should help tolower the adoption threshold for in situ visualization and enable a broader range ofapplications to benefit from it.

In a next presentation, Peter Messmer gave an update on visualization efforts atNVIDIA. Specifically, the arrival of Turing GPUs with hardware accelerated raytracing put a new emphasis on ray tracing for scientific visualization. Not only does thistechnology help to produce powerful images for outreach and education, but it alsoprovides better visual cues for e.g. depth and therefore helps a scientist in the day to

4th International Workshop on In Situ Visualization (WOIV’19) xvii

day work to better understand the spatial relationship of objects in a scene. With theintegration of these technologies into popular scientific visualization tools like Para-View, these technologies become available to a broad range of scientific visualizationusers. In addition to these latest developments on hardware accelerated ray tracing,Peter raised the question whether there is a need for a standardization efforts for raytracing in scientific visualization in order to free ISV’s from developing their ownimplementations while offering optimal performance on each platform.

In a subsequent presentation, Mike Ringenburg from Cray talked about the con-currently ongoing workshop series on interactive HPC. This conference series has beenestablished to bring together domain scientist, HPC center managers and middlewaredevelopers to address the challenges of using HPC systems in an interactive manner.With visualization benefitting significantly from a high degree of interactivity, it wasinteresting to see the parallels and differences in the two areas. While the focus of thein situ vis workshop series is on the visualization aspect, both interactive or in a batchfashion, the interactive HPC workshops are more interested in the middleware andcenter policy aspects of interactivity, specifically in light of the convergence of HPCand data sciences. However, despite the differences in the two target audiences, the talkand the subsequent discussion showed that there clearly are common aspects ofinterest.

Next, Joao Barbosa from TACC presented a summary of the IXPUG In SituHackathon 2019. The Hackathon intends to provide a forum to bring simulationdevelopers together with visualization experts for hands-on implementation of in situanalysis. It took first place in 2017 in Austin, Texas hosted at TACC, then in 2018 inChicago, Illinois hosted at Argonne National Lab, and finally in 2019 Santa Fe, NewMexico hosted by TACC and LAN. Joao presented some statistics of these events with30–40 participants each, noting that they were successful because they provided adistraction-free environment and a good mix of new and returning participants. Heconcluded by presenting selected results of the last Hackathon.

Finally, Silvio Rizzi, on behalf of Jim Ahrens, talked about the development anduse of in situ visualization and analysis approaches for the U.S. Exascale ComputingProject (ECP), an accelerated research and development project funded by the USDepartment of Energy (DOE). It is a seven-year, $1.7 B R&D effort that launched in2016 with the participation of six core DOE National Laboratories (Argonne, LawrenceBerkeley, Lawrence Livermore, Los Alamos, Oak Ridge, Sandia), with staff from mostof the 17 DOE national laboratories taking part in the project. While ECP applicationstarget national problems, Silvio focussed on the in-situ aspects. He presented anoveriew of ECP Software Technology Data and Visualization projects and an outline ofseveral algorithms and subsystems planned within these projects.

3.2 Research Papers

In the first research presentation, Zhang and Entezari discussed their paper “In-SituData Reduction via Incoherent Sensing”. In this context, they presented study for datareduction in an in-situ situation based on compressed sensing. The volumetric data isfirst sparsified using wavelet, curvelet, and surfacelet transforms, then compressedsensing techniques are used for compression and reconstruction. The authors

xviii S. Frey et al.

demonstrate the effectiveness of their method with examples from chemistry andastrophysics and examine how different methods for compressing the data influence thequality of the reconstruction.

Next, Sarton et al., in their paper “Distributed out-of-core approach for in-situvolume rendering of massive dataset”, present a GPU-based volume renderingapproach suitable for out-of-core/in-situ visualization of large volumetric data sets. Thevolume is divided into bricks and managed with a page table layout. An exemplary2-CPU and 4-GPU server is used for storing, handling, and rendering the volume, theresult is then streamed to a thin client for visualization. By providing each GPU with alocal brick cache, and by using multi-resolution volume data the system is capable ofproviding interactive frame rates, which is demonstrated on two volumetric datasets.

In their paper “In-Situ Processing in Climate Science”, Röber and Engels, describethe current efforts of leveraging in situ visualization at the German Climate ComputingCentre (DKRZ). The paper describes the Catalyst in situ library being integrated intothe ICON model and used to generate images and extract geometric features. It alsodescribes some work to write hierarchical data and to compress data with waveletdecomposition. The authors describe a workflow where a special adaptor is developedto fit ICON model to Catalyst, and then the model output is directly visualized throughParaView. The authors also describe a progressive visualization approach where lowerresolution data is used for exploration, whereas the original resolution data is used fordetailed examination.

Finally, Hummels and van Kooten, in their paper “Leveraging NVIDIA Omniversefor In Situ Visualization”, propose a method that allows interactive, high-qualityvisualization of distilled simulation geometry. Omniverse is NVIDIA’s collaborationplatform for 3D production pipelines. It is integrated with a number of commerciallyavailable 3D software packages and game engines and enables content creators to workon different aspects of models or entire scenes simultaneously. By integrating Para-View and Catalyst with the Omniverse, the visualization geometry becomes immedi-ately accessible to a number of 3D content authoring and rendering tools without therequirement of invasive software changes or tedious postprocessing and conversionworkflows.

The workshop was concluded by a panel dicsussion on the “Future of In-situ atISC”, covering a variety of different topics. Among others, we talked about the differentdirections the workshop could take in future iterations, considering the landscape ofrelated workshops and events in the field of in situ visualization. Here, especially thestrong inclusion of simulation and domain scientists, as well as the focus on sharingexperiences gained in the development process were identified as important aspects. Inaddition, the role of interactivity in in-situ visualization was discussed, to identifycollaboration potential with the interactivity workshop in future iterations.

4th International Workshop on In Situ Visualization (WOIV’19) xix

Contents

On the Use of Kernel Bypass Mechanisms for High-PerformanceInter-container Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Gabriele Ara, Luca Abeni, Tommaso Cucinotta, and Carlo Vitucci

Continuous-Action Reinforcement Learning for Memory Allocationin Virtualized Servers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Luis A. Garrido, Rajiv Nishtala, and Paul Carpenter

Container Orchestration on HPC Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . 25Marco Enrico Piras, Luca Pireddu, Marco Moro, and Gianluigi Zanetti

Data Pallets: Containerizing Storage for Reproducibility and Traceability . . . . 36Jay Lofstead, Joshua Baker, and Andrew Younge

Sarus: Highly Scalable Docker Containers for HPC Systems . . . . . . . . . . . . 46Lucas Benedicic, Felipe A. Cruz, Alberto Madonna, and Kean Mariotti

Singularity GPU Containers Execution on HPC Cluster . . . . . . . . . . . . . . . . 61Giuseppa Muscianisi, Giuseppe Fiameni, and Abdulrahman Azab

A Multitenant Container Platform with OKD, Harbor Registry and ELK . . . . 69Jarle Bjørgeengen

Enabling GPU-Enhanced Computer Vision and Machine LearningResearch Using Containers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

Martial Michel and Nicholas Burnett

Software and Hardware Co-design for Low-Power HPC Platforms . . . . . . . . 88Manolis Ploumidis, Nikolaos D. Kallimanis, Marios Asiminakis,Nikos Chrysos, Pantelis Xirouchakis, Michalis Gianoudis,Leandros Tzanakis, Nikolaos Dimou, Antonis Psistakis,Panagiotis Peristerakis, Giorgos Kalokairinos, Vassilis Papaefstathiou,and Manolis Katevenis

Modernizing Titan2D, a Parallel AMR Geophysical Flow Code to SupportMultiple Rheologies and Extendability. . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

Nikolay A. Simakov, Renette L. Jones-Ivey, Ali Akhavan-Safaei,Hossein Aghakhani, Matthew D. Jones, and Abani K. Patra

Asynchronous AMR on Multi-GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113Muhammad Nufail Farooqi, Tan Nguyen, Weiqun Zhang,Ann S. Almgren, John Shalf, and Didem Unat

Batch Solution of Small PDEs with the OPS DSL. . . . . . . . . . . . . . . . . . . . 124Istvan Z. Reguly, Branden Moore, Tim Schmielau, Jacques du Toit,and Gihan R. Mudalige

Scalable Parallelization of Stencils Using MODA . . . . . . . . . . . . . . . . . . . . 142Nabeeh Jumah and Julian Kunkel

Comparing High Performance Computing AcceleratorProgramming Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Swaroop Pophale, Swen Boehm, and Verónica G. Vergara Larrea

Tracking User-Perceived I/O Slowdown via Probing . . . . . . . . . . . . . . . . . . 169Julian Kunkel and Eugen Betke

A Quantitative Approach to Architecting All-Flash Lustre File Systems . . . . . 183Glenn K. Lockwood, Kirill Lozinskiy, Lisa Gerhardt, Ravi Cheema,Damian Hazen, and Nicholas J. Wright

MBWU: Benefit Quantification for Data Access Function Offloading . . . . . . 198Jianshen Liu, Philip Kufeldt, and Carlos Maltzahn

Footprinting Parallel I/O – Machine Learning to ClassifyApplication’s I/O Behavior. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

Eugen Betke and Julian Kunkel

Adventures in NoSQL for Metadata Management . . . . . . . . . . . . . . . . . . . . 227Jay Lofstead, Ashleigh Ryan, and Margaret Lawson

Towards High Performance Data Analytics for Climate Change . . . . . . . . . . 240Sandro Fiore, Donatello Elia, Cosimo Palazzo, Fabrizio Antonio,Alessandro D’Anca, Ian Foster, and Giovanni Aloisio

An Architecture for High Performance Computing and Data SystemsUsing Byte-Addressable Persistent Memory . . . . . . . . . . . . . . . . . . . . . . . . 258

Adrian Jackson, Michèle Weiland, Mark Parsons,and Bernhard Homölle

Mediating Data Center Storage Diversity in HPC Applicationswith FAODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

Patrick Widener, Craig Ulmer, Scott Levy, Todd Kordenbrock,and Gary Templet

Predicting File Lifetimes with Machine Learning . . . . . . . . . . . . . . . . . . . . 288Florent Monjalet and Thomas Leibovici

An I/O Analysis of HPC Workloads on CephFS and Lustre . . . . . . . . . . . . . 300Alberto Chiusole, Stefano Cozzini, Daniel van der Ster,Massimo Lamanna, and Graziano Giuliani

xxii Contents

Enabling Fast and Highly Effective FPGA Design Process Using the CAPISNAP Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

Alexandre Castellane and Bruno Mesnet

Scaling the Summit: Deploying the World’s Fastest Supercomputer. . . . . . . . 330Verónica G. Vergara Larrea, Wayne Joubert, Michael J. Brim,Reuben D. Budiardja, Don Maxwell, Matt Ezell, Christopher Zimmer,Swen Boehm, Wael Elwasif, Sarp Oral, Chris Fuson, Daniel Pelfrey,Oscar Hernandez, Dustin Leverman, Jesse Hanley, Mark Berrill,and Arnold Tharrington

Parallelware Tools: An Experimental Evaluation on POWER Systems . . . . . . 352Manuel Arenaz and Xavier Martorell

Performance Evaluation of MPI Libraries on GPU-Enabled OpenPOWERArchitectures: Early Experiences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

Kawthar Shafie Khorassani, Ching-Hsiang Chu, Hari Subramoni,and Dhabaleswar K. Panda

Exploring the Behavior of Coherent Accelerator Processor Interface (CAPI)on IBM Power8+ Architecture and FlashSystem 900 . . . . . . . . . . . . . . . . . . 379

Kaushik Velusamy, Smriti Prathapan, and Milton Halem

Porting Adaptive Ensemble Molecular Dynamics Workflowsto the Summit Supercomputer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397

John Ossyra, Ada Sedova, Arnold Tharrington, Frank Noé,Cecilia Clementi, and Jeremy C. Smith

Performance Comparison for Neuroscience Application Benchmarks . . . . . . . 418Andreas Herten, Thorsten Hater, Wouter Klijn, and Dirk Pleiter

Evaluating POWER Architecture for Distributed Training of GenerativeAdversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432

Ahmad Hesam, Sofia Vallecorsa, Gulrukh Khattak,and Federico Carminati

A Study on the Performance of Reproducible Computations . . . . . . . . . . . . . 441Nico Bombace and Michèle Weiland

Three Numerical Reproducibility Issues That Can Be Explainedas Round-Off Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452

Michael Mascagni

Training Multiscale-CNN for Large Microscopy Image Classificationin One Hour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463

Kushal Datta, Imtiaz Hossain, Sun Choi, Vikram Saletore, Kyle Ambert,William J. Godinez, and Xian Zhang

Contents xxiii

Benchmarking Deep Learning Infrastructures by Means of TensorFlowand Containers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478

Adrian Grupp, Valentin Kozlov, Isabel Campos, Mario David,Jorge Gomes, and Álvaro López García

MagmaDNN: Towards High-Performance Data Analytics and MachineLearning for Data-Driven Scientific Computing . . . . . . . . . . . . . . . . . . . . . 490

Daniel Nichols, Nathalie-Sofia Tomov, Frank Betancourt,Stanimire Tomov, Kwai Wong, and Jack Dongarra

Open OnDemand: HPC for Everyone . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504Robert Settlage, Alan Chalker, Eric Franz, Doug Johnson, Steve Gallo,Edgar Moore, and David Hudak

Highly Interactive, Steered Scientific Workflows on HPC Systems:Optimizing Design Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514

John R. Ossyra, Ada Sedova, Matthew B. Baker, and Jeremy C. Smith

The Role of Interactive Super-Computing in Using HPC for UrgentDecision Making. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528

Nick Brown, Rupert Nash, Gordon Gibb, Bianca Prodan, Max Kontak,Vyacheslav Olshevsky, and Wei Der Chien

Deep Learning at Scale for Subgrid Modeling in Turbulent Flows:Regression and Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541

Mathis Bode, Michael Gauding, Konstantin Kleinheinz,and Heinz Pitsch

Using FPGAs to Accelerate HPC and Data Analyticson Intel-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561

Thomas Steinke, Estela Suarez, Taisuke Boku, Nalini Kumar,and David E. Martin

Exploring the Acceleration of the Met Office NERC Cloud ModelUsing FPGAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567

Nick Brown

Acceleration of Scientific Deep Learning Models on HeterogeneousComputing Platform with Intel® FPGAs. . . . . . . . . . . . . . . . . . . . . . . . . . . 587

Chao Jiang, Dave Ojika, Thorsten Kurth, Prabhat, Sofia Vallecorsa,Bhavesh Patel, and Herman Lam

In-Situ Data Reduction via Incoherent Sensing . . . . . . . . . . . . . . . . . . . . . . 601Kai Zhang and Alireza Entezari

In-Situ Processing in Climate Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612Niklas Röber and Jan Frederik Engels

xxiv Contents

Distributed Out-of-Core Approach for In-Situ Volume Renderingof Massive Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623

Jonathan Sarton, Yannick Remion, and Laurent Lucas

Leveraging NVIDIA Omniverse for In Situ Visualization . . . . . . . . . . . . . . . 634Mathias Hummel and Kees van Kooten

Hands-On Research and Training in High Performance Data Sciences,Data Analytics, and Machine Learning for Emerging Environments. . . . . . . . 643

Kwai Wong, Stanimire Tomov, and Jack Dongarra

Correction to: High Performance Computing . . . . . . . . . . . . . . . . . . . . . . . C1Michèle Weiland, Guido Juckeland, Sadaf Alam, and Heike Jagode

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657

Contents xxv