Visualization Infrastructure and Services at the MPCDF

Visualization Infrastructure and Services at the MPCDF Markus Rampp & Klaus Reuter Max Planck Computing and Data Facility (MPCDF) (visualization@mpcdf.mpg.de) Interdisciplinary Cluster Workshop on Visualization Garching, Nov 4, 2015 slide 1

Outline Topics overview remote visualization services hardware & software infrastructure project support challenges and outlook MPCDF Visualization Team people involved (part-time, main focus is HPC) Elena Erastova (visualization projects) Klaus Reuter (software and hardware coordination, consulting, projects, training) Markus Rampp (consulting, projects, training) slide 2

Visualization infrastructure for the Max-Planck-Society MPCDF visualization services: provide central software and hardware infrastructure for remote visualization target: interactive data exploration & analysis, presentation support for adaptation and instrumentation of simulation codes guidance for selection, adoption and usage of analysis & visualization software dedicated support for individual (particularly demanding) visualization projects Main conceptual challenges: broad range of disciplines in MPG: Plasmaphysics, Astrophysics,..., comp. Biology many different scientific contexts variety of simulation codes: home-grown, commercial, open-source, third-party,... non-standardized, heterogeneous data structures and formats legacy analysis pipelines,... massive datasets from HPC simulations: (massive: amount of raw data, memory requirements, complexity) multidimensional (3D + time), multi-variate data unusual grids: meshless data, special curvilinear coordinates,... slide 3

Visualization infrastructure for the Max-Planck-Society Status consulting & dedicated project support since 2008 MPG visualization cluster operational since Sep. 2010 open to all MPG scientists and collaboration partners many projects supported (some highlights by K. Reuter) broad userbase (beyond Garching campus) Rationale for centralized visualization in the MPG: a necessity for a HPC centre rather than an optional service huge amounts of output data produced by HPC simulations transfer of raw data for local analysis & visualization no more possible even dumping the RAM is becoming prohibitive due to I/O constraints in-situ visualization (not covered here) visualization requires HPC-like resources (specialized hardware, housing,... ) requires substantial expertise on methods, software,... sustainability Technological prerequisites efficient and transparent remote rendering solution via WAN: VirtualGL/TurboVNC slide 4

Central visualization infrastructure: technical prerequisites traditional X over ssh (e.g. ssh -X) 3D data are transfered to the client fails to deliver interactive frame rates uses X-server/graphics card of the client not suited for 3D applications server-side rendering only (compressed) image stream is transferred delivers interactive frame rates with moderate WAN bandwidth uses X-server/graphics card(s) of the server generic solution (OpenGL) mature software solutions/products: VirtualGL/TurboVNC (Open Source, ex SUN) (original illustrations by L. Scheck, by courtesy of LRZ) slide 5

Remote-visualization cluster Focus: slide 6 enable our (geographically dispersed) scientific users to perform complex visualization tasks without special technical prerequisites (software, hardware) remote visualization Hardware overview (HP cluster) 5 standard visualization nodes each equipped with: 2 Intel quadcore CPUs: 8 cores, 144 GB RAM 2 NVidia FX 5800 graphics cards 1 high-end visualization node: 4 Intel hexacore CPUs, 24 cores, 256 GB RAM 2 NVidia FX 5800 graphics cards 1 login node: viz00.rzg.mpg.de dedicated disk system (GPFS, 30 TB) GPFS filesystem /ptmp of HPC system Hydra mounted 2 graphics workplaces (active stereo) in MPCDF offices Software stack SLES 11 (MPCDF standard cluster setup), VizStack middleware (GPUs, X-servers,... ) web-based reservation system (HP, MPCDF) remote rendering solution: VirtualGL/TurboVNC (free clients for Linux, MS Windows, MAC)

User interface Remote desktop (via TurboVNC) a standard desktop in a separate window application agnostic desktop icons for main applications preconfigured according to session properties (number of GPUs, CPUs) linux terminal: vglrun <command> slide 7

Software for Visualization Software for interactive data visualization and analysis VisIt: main workhorse for 3D analysis Paraview: main workhorse for 3D analysis VAPOR: large-scale data (requires preprocessing) Voreen: volume rendering Tools and libraries GNU R, IDL, MATLAB, gnuplot,... VTK, HDF5, SILO,... mencoder: scripts for x264 encoding of movies Special-purpose software Splotch: a (non-interactive), parallel ray tracer for SPH data. VMD (Visual Molecular Dynamics): a molecular graphics software. POV-Ray: a freeware multi-platform ray-tracing package. Blender: an open source, cross platform suite of tools for 3D creation. slide 8

Application support Documentation slide 9 http://www.rzg.mpg.de/services/visualisation/ Training courses (http://www.rzg.mpg.de/services/visualisation/scientificdata/presentations) K. Reuter: RZG-Services zur Visualisierung wissenschaftlicher Datensätze, DV-Treffen der Max-Planck-Institute, Göttingen, Sep 15, 2010 K. Reuter: Scientific Visualisation Services at RZG, Seventh GOTiT High Level Course, Garching, Oct 19, 2010 M. Rampp: Introduction to VisIt, LRZ course on Visualisation of Large Data Sets on Supercomputers, 2010 2011 M. Rampp: Introduction to VisIT, 11th Summer school on scientific visualization, CINECA Bologna/Italy, Jun 13, 2012 M. Rampp: Visualization of HPC simulation data: overview and tutorial, ISSS-12, Prague (2015) overview talks at Max-Planck-Institutes: MPA, Garching (2009), FHI, Berlin (2011), MPI f. Biophysics, Frankfurt (2014),... Project support dedicated support for visualization projects at different levels: from basic first level support to comprehensive visualization and analysis tasks requires (considerable) insight to scientific domain several completed and ongoing projects, in close collab. with the users/scientists: http://www.rzg.mpg.de/services/visualisation/scientificdata/projects contact: visualization@rzg.mpg.de

MPG/MPCDF reference applications Projects with MPCDF support (in close collab. with research groups) application domains: Plasmaphysics: MHD turbulence simulations for nuclear fusion research (IPP) Stellar astrophysics: Supernova simulations, NS mergers (MPA) Cosmology: Structure and star formation (MPE) Molecular dynamics: Materials research for plasma-wall-interaction (IPP), DFT (FHI) CFD: DNS simulations of turbulent Taylor-Couette flows (MPI-DS) data structures/grids: regular: cartesian, polar (2D, 3D), block-structured ( Yin-Yan ) irregular: (mapped) point clouds data sizes, dimensions: up to 2048 3 (cartesian), 1000 180 360 (polar), 2048 769 1153 (cylindrical) up to 10 6 (particles in 3D), 10 7 (nodes in 3D unstructured mesh) all: multi-variable (scalar, vector), time-dependent see also: http://www.rzg.mpg.de/services/visualisation/scientificdata/projects Presentation by K. Reuter slide 10

MPG/MPCDF reference applications slide 11

Challenges & outlook Technological hitting the limits of general-purpose software tools (VisIT, Paraview): interactivity, memory demands: O(1000 3 ) data use GPUs in HPC system, e.g. MPG Hydra with Nvidia K20x GPUs enables in-situ visualization: a big buzz or something interesting to watch? basic technique: implement library calls in simulation code (APIs for C, FORTRAN) mediates callbacks to visualization tool supported by Paraview ( catalyst ), VisIT ( libsim ) slide 16

Challenges & outlook Organizational dedicated project support is of key importance (but not scalable) beyond scientific data analysis and insights ever increasing standards and expectations for public understanding of science are we the right people to direct professional animations (TV documentaries)? do we need real scientific data for this at all? credits? Max-Planck Visualization Award highly efficient and innovative algorithms often don t make it into usable software slide 17

Challenges & outlook Organizational. dedicated project support is of key importance (but not scalable). beyond scientific data analysis and insights ever increasing standards and expectations for public understanding of science are we the right people to direct professional animations (TV documentaries)? do we need real scientific data for this at all? credits? y Max-Planck Visualization Award. highly efficient and innovative algorithms often don t make it into usable software Outlook. remote visualization on HPC system Hydra (to replace visualization cluster in early 2016) slide 18