Best Practices Supporting Science at CSCS

Best Practices Supporting Science at CSCS Michele De Lorenzi (CSCS)

Case 1 – Simplify the usage of HPC resources GPU Accelerated In-Situ Visualization

§  Peter Messmer (NVIDIA Co-Design Lab for Hybrid Multicore Computing at ETH Zurich)

Best Practices Supporting Science at CSCS 3

HPC Workflow

Workstation

Viz Cluster Supercomputer

File System

Setup

Dump, Checkpointing Visualization,

Analysis

Analysis, Visualization


Traditional Workflow: Challenges

Workstation


File System

Setup

Dump, Checkpointing Visualization,

Analysis

Analysis, Visualization

Lack of interactivity prevents “intuition”

I/O becomes main simulation bottleneck

Viz resources need to scale with simulation

High-end viz neglected due to workflow complexity

Eliminating the IO Bottleneck: In-Situ ViZ

Workstation


Setup Analysis, Visualization


Eliminating the IO Bottleneck: In-Situ ViZ

Workstation


Setup Analysis, Visualization


Different Visualization Scenarios

§  1) Legacy workflow §  Separate compute & Viz system §  Communication via filesystem

§  2) In-situ partitioned or integrated §  Different nodes for viz & compute §  Communication via High Performance Network

§  3) In-situ co-processing

§  Compute and visualization on same node

HPC System

Compute

GPU

Viz System

Viz GPU

Filesystem

HPC System

Compute

GPU

Viz GPU Network

HPC System

Compute+Viz nodes


Typical Scientific Visualization system

§  Visualization is more than Rendering

§  Simulation: Data as needed in numerical algorithm

§  Filtering: Conversion of simulation data into data ready for rendering §  Typical operations: binning, down/up-sampling, iso-surface extraction, interpolation, coordinate transformation, sub-

selection, .. §  Sometimes embedded in simulation

§  Rendering: Conversion of shapes to pixels (Fragment processing)

§  Compositing: Combination of independently generated pixels into final frame

Visualization

Filtering Rendering Compositing Simulation


Visualization Dataflow

CPU

GPU

Simulation

Filtering Rendering


Fully GPU accelerated Visualization

CPU

GPU

Simulation Rendering

Filtering


FuLLy GPU accelerated Visualization

CPU

GPU

Simulation Rendering

Filtering Speedup


Bonsai With In-Situ Viz On Daint


Acknowledgement

§  Bonsai development: Evghenii Gaburov (SURFsara), Jeroen Bédorf (CWI)

§  OpenGL on Daint: Gilles Fourestey (CSCS), Nina Suvanphim (Cray)

§  OpenGL on Titan: Don Maxwell (ORNL)

Thomas Schulthess (CSCS) and Jack Wells (ORNL) for providing access to Daint and Titan for these experiments


Case 2: Software as a service Supporting the MARVEL community

14

MATERIALS’ REVOLUTION: COMPUTATIONAL DESIGN AND DISCOVERY OF NOVEL MATERIALS

EPFL-ETHZ-UNIBAS-UNIFR-UNIGE-USI-UZH-IBM-CSCS-EMPA-PSI MARVEL


Access Pattern

Piz Dora

External Login Access (ELA) CSCS


EPFL

AiiDA Server

/store

Repository access

MARVEL Researcher

Research Community

Traditional access (batch jobs)

Access through AiiDa

Scientific Community access

Case 3: Supporting scientific communities The Platform for Advanced Scientific Computing (PASC)

17

Swiss Platform for Advanced Scientific Computing

§  Promote joint effort to address key scientific issues in different domain sciences exploiting the potential of the next generation of HPC architectures

§  Interdisciplinary collaboration between §  domain scientists,

§  computational scientists,

§  software developers,

§  computing centres

§  hardware developers.

§  Builds on the principle of co-design involving all these actors

18 Best Practices Supporting Science at CSCS


PASC Organization

PASC instruments

Organized in four pillars:

§ Application Support Network (ASN)

§ Co-design projects

§ Institutional computing system pillar supporting the procurement of institutional computing systems

§ The training pillar supporting the Swiss Graduate Program Foundations in Mathematics and Informatics for Computer Simulations in Science and Engineering (FOMICS).


Application Support Network (ASN)

§  ASN are intended to coordinate and support the effort in different domain science areas to pursue the objectives of the project

§  The following Domain Science Networks are active today: § Climate - Climate and Atmospheric Modelling

§ Earth - Solid Earth Dynamics

§  Life Sciences - Life Sciences Across Scales

§ Materials - Materials Simulations

§ Physics - Plasma Physics, Astrophysics and Multiscale Fluid Dynamics


Co-design projects

§  Interdisciplinary projects involving several (or all) actors §  address the following priorities for optimal exploitation of new and

emerging supercomputing platforms: §  Refactoring of key application codes and re-engineering of their algorithms.

§  Development of numerical libraries and their integration into application codes.

§  Incorporations of innovative sub-systems (e.g. I/O, data streams, etc.) into existing simulation codes

§  Development of programming environments or components, performance tuning and analysis/monitoring tools.


Supported projects

Physics and Astrophysics: §  Particles and fields; PI: Laurent Villard (EPFL)

§  DIAPHANE Radiative Transport in astrophysical simulations; PI: Lucio Mayer (University of Zurich)

Material Science

§  ANSWERS: nano-device simulations; PI: Mathieu Luisier (ETH Zurich)

§  Electronic Structure Calculations: electronic structure calculations; PI: Jürg Hutter (University of Zurich)

§  ENVIRON A Library for Electronic-structure Simulations; PI: Stefan Goedecker (University of Basel)


Supported projects (continued)

Life Science §  Angiogenesis in Health and Disease: In-vivo and in-silico; PI:

Petros Koumoutsakos (ETH Zurich)

§  Coupled Cardiac Simulations: HPC Framework for Coupled Cardiac Simulations; PI: Rolf Krause (University of Lugano), Alfio Quarteroni (EPFL)

§  Genomic Data Processing; PI: Ioannis Xenarios (University of Lausanne)

§  HPC-ABGEM: agent-based general ecosystems models; PI: Christoph Zollikofer (University of Zurich)


Supported projects (continued)

Geo-physics and Climate

§  GeoPC: hybrid parallel smoothers for multigrid preconditioners; PI: Paul Tackley (ETH Zurich)

§  GeoScale: A framework for multi-scale seismic modelling and inversion; PI: Andreas Fichtner (ETH Zurich)

§  Grid Tools: Towards a library for hardware oblivious implementation of stencil based codes; PI: Oliver Fuhrer (Meteosuisse)

Computer Science

§  Heterogeneous Compiler Platform for Advanced Scientific Codes; PI: Torsten Hoefler (ETH Zurich)

§  Multiscale applications: Optimal deployment of multiscale applications on a HPC infrastructure; PI: Bastien Chopard (UNIGE)


PASC conference

§  PASC organizes yearly a Platform of Advanced Scientific Computing Conference.

§  The goal is to bring together research groups within diverse scientific domains to foster interdisciplinary collaborations and to strengthen (HPC) knowledge exchange.

§  PASC15 – June 1-2, 2015 hosted by ETH Zurich: http://www.pasc15.org/



Thank you for your attention.

Best Practices Supporting Science at CSCS

Documents