Science Gateway Group Overview Marlon Pierce, Suresh Marru, Raminder Singh Presentation to PTI CREST Lab May 2nd, 2012
Download ReportTranscript Science Gateway Group Overview Marlon Pierce, Suresh Marru, Raminder Singh Presentation to PTI CREST Lab May 2nd, 2012
Science Gateway Group Overview Marlon Pierce, Suresh Marru, Raminder Singh Presentation to PTI CREST Lab May 2nd, 2012 IU Science Gateway Group Overview • IU Science Gateway Group members – Marlon Pierce: Group Lead – Suresh Marru: Principal Software Architect – Raminder Singh, Yu Ma, Jun Wang, Lahiru Gunathilake, (Chathura Herath): Senior team members – Five interns and research assistants • NSF SDCI funding of Open Gateway Computing Environments project – TACC (M. Dahan), SDSC (N. Wilkins-Diehr), SDSU (M. Thomas), NCSA (S. Pamidighantam), UIUC (S. Wang), Purdue (C. Song), UTHSCSA (E. Brookes) – XSEDE Extended Collaboration Support Services Science Gateway Group Focus Areas • Open source, open community software for cyberinfrastructure – Apache Rave: http://rave.apache.org – (portal software) – Apache Airavata: http://www.airavata.org (workflow software) • Extended collaborations with application scientists – Astronomy, astrophysics, biophysics, chemistry, nuclear physics, bio informatics… Science Gateways outreach & democratize Research and Education Developers Lowering the barrier for using complex end-to-end computational technologies Researchers Democratize Cyberinfrastructure Empower Facilitate Research & Education Community Possibilities for Collaboration • Scientific workflows exhibit a number of distributed execution patterns. – Not just DAGs. – Workflow start as an abstraction, but need system, application, library level interactions. – We are trying to generalize our execution framework over a number of applications. – This is parallel, complementary to work in CREST. • Collaborations can be mediated through Apache, students’ independent study, targeted papers. Apache Airavata Open Community Software for Scientific Workflows airavara.org Apache Airavata • Science Gateway software framework to – Compose, manage, execute, and monitor computational workflows on Grids and Clouds – Web service abstractions to legacy command linedriven scientific applications – Modular software framework to use as individual components or as an integrated solution. • More Information – Airavata Web Site: airavata.org – Developer Mailing Lists: [email protected] Apache Airavata High Level Overview A Classic Scientific Workflow • Workflows are composite applications built out of independent parts. – Executables wrapped as network accessible services • Example: codes A, B, and C need to be executed in a specific sequence. – A, B, C: codes compiled and executable on a cluster, supercomputer, etc. through schedulers. • A, B, and C do not need to be co-located • A, B, and C may be sequential or parallel • A, B and C may have data or control dependencies – Data may need to be staged in and out • Some variations on ABC: – – – – Conditional execution branches, interactivity Dynamic execution resource binding Iterations (Do-while, For-Each) over all or parts of the sequence Triggers, events, data streams Linked Environment for Atmospheric Discovery Storms Forming Forecast Model Streaming Observations Data Mining Instrument Steering Refine forecast grid Envisioned by a multi-disciplinary team from OU, IU, NCSA, Unidata, UAH, Howard, Millersville, Colorado State, RENCI On-Demand Grid Computing Open Grid/Gateway Computing Environments LEAD GridChem OGCE TeraGrid User Portal Re-engineer, Generalize, Build, Test and Release OGCE NMI & SDCI Funding Atmospheric Science LEAD, OLAM Molecular Chemistry GridChem, ParamChem, OREChem Bio Physics OVP/ RST/ MIG Bio Informatics BioVLAB, mCpG Astronomy ODI, DES-SimWG Nuclear Physics XSEDE ECSS Gateway Projects Ultrascan LCCI Projects in the pipe QuakeSim, VLAB, Einstein Gateway 11 Realizing the Universe for the Dark Energy Survey (DES) Using XSEDE Support Principal Investigators: Prof. August Evrard (University of Michigan), Prof. Andrey Kravtsov (University of Chicago) Background & Explanation: The Dark Energy Survey (DES) is an upcoming international experiment that aims to constrain the properties of dark energy and dark matter in the universe using a deep, 5000-square degree survey of cosmic structure traced by galaxies. Fig. 1 The density of dark matter in a thin radial slice as seen by a synthetic observer located in the 8 billion light-year computational volume. Image courtesy Matthew Becker, University of Chicago. Fig. 2: A synthetic 2x3 arcmin DES sky image showing galaxies, stars, and observational artifacts. Courtesy Huan Lin, FNAL. To support this science, the DES Simulation Working Group is generating expectations for galaxy yields in various cosmologies. Analysis of these simulated catalogs offers a quality assurance capability for cosmological and astrophysical analysis of upcoming DES telescope data. Our large, multi-staged computations are a natural fit for workflow control atop XSEDE resources. DES Application Component Description CAMB Code for Anisotropies in the Microwave Background is a serial FORTRAN code that computes the power spectrum of dark matter (simulation initial conditions). Output is a small ASCII file describing the power spectrum. 2LPTic Second-order Lagrangian Perturbation Theory initial conditions code is an MPI-based C code that computes the initial conditions for the simulation from parameters and an input power spectrum generated by CAMB. Output is a set of binary files that vary in size from ~80-250 GB depending on the simulation resolution. LGadget MPI based C code that uses a TreePM algorithm to evolve a gravitational N-body system. The outputs of this step are system state snapshot files, as well as light-cone files, and some properties of the matter distribution, including the power spectrum at various timesteps. The total output from LGadget depends on resolution and the number of system snapshots stored, and approaches 10~TB for large DES simulation boxes. ADDGALS Creates a synthetic galaxy catalog for science analysis Case Study: Dark Energy Survey • • • • Long running code: Based on simulation box size L-gadget can run for 3 to 5 days using more than 1024 cores on TACC Ranger. Do-While Construct: Restart service support is needed to work around queue time restrictions. Do-while construct was developed to address the need. Data size and file transfer challenges: Lgadget produces 10~TB for large DES simulation boxes in system scratch so data need to moved to persistent storage ASAP File system issues: More than 10,000 lightcone files are doing continues file I/O. Ranger has one Luster metadata server to serve 300 I/O nodes. Sometime metadata server can’t fine these lightcone files, which make simulations to stop. We have wasted ~50k SU this month struggling with I/O issues and to get recommendation to use MPI I/O Figure: Processing steps to build a synthetic galaxy catalog. Xbaya workflow currently controls the top-most element (Nbody simulations) which consists of methods to sample a cosmological power spectrum (ps), generating an initial set of particles (ic) and evolving the particles forward in time with Gadget (N-body). The remaining methods are run manually on distributed resources. Case Study: ParamChem • ParamChem researchers try to optimize the geometry of new molecules which may or may not converge with in a given time or number of steps. • Factors that include the mathematical convergence issues in solutions for partial integro-differential equations to potential shallowness of an energy landscape. • The intermediate outputs from model iterations can be used to determine convergence. Complex graph executions with support for long running and interactive executions to address non-deterministic convergence problems. Case Study: LCCI • The Leadership Class Configuration Interaction (LCCI) project is targeted to accurately predict properties of nuclei for astrophysical and fusion energy processes. – James Vary’s group at Iowa State • One of the PetaScale Apps – Use DOE INCITE and NSF Blue Waters awarded resources – Currently using 55 million processor hours on ORNL Cray XK6 machine and Argonne Blue Gene/P. • Workflow Goals – – – – Democratizing science: Reduce the learning curve associated with running simulations Controlled access: avoid inappropriate use of super computing resources Reproducibility of results Avoiding waste: needless duplication; minimize erroneous use of codes and erroneous exchanging of intermediate files – Sustainability: Ensure long-term preservation of applications, configurations and results – Provenance: Provide the ability to track down the provenance of results as well as reuse previously completed results where applicable without recalculating – Parametric sweeps: Allow components to run over a range of dataset such that applications may produce richer simulations Example Workflow: Nuclear Physics Courtesy of collaboration with Prof. James Vary and team, Iowa State Next Generation Workflow Systems • Scientific workflow systems and compiled workflow languages have focused on modeling, scheduling, data movement, dynamic service creation and monitoring of workflows. • Building on these foundations we extend to a interactive and flexible workflow systems. – interactive ways of steering the workflow execution – interpreted workflow execution model – high level instruction set supporting diverse execution patterns – flexibility to execute individual workflow activity and wait for further analysis. Various State Changes can tap into lower layers Running Workflow Ready Node Wai ng Node Load from provenance Running Node Failed Node Finished Node Finished Workflow Ready Workflow Paused Workflow Stopped Workflow Failed Workflow Uncertainties in Workflows • Mathematical uncertainty: – PDE’s may not converge for certain conditions – Statistical techniques lead to nondeterministic results, propagation of uncertainties. – CLoser observation at computational output ensure acceptability of results. • Domain uncertainty: – Optimization execution patterns: Scenarios of running against range of parameter values in an attempt to find the most appropriate input set. – Initial execution providing estimate of the accuracy of the inputs and facilitating further refinement. – Outputs are diverse and nondeterministic • Resource uncertainty: – Failures in distributed systems are norm than an exception – Transient failures can be retried if computation is side-effect free/Idempotent. – Persistent failures require migration Next Steps • Workflow start as an abstraction, but need system, application, library level interactions. – We are trying to generalize our execution framework over a number of applications. – This is parallel, complementary to work in CREST. • Collaborations can be mediated through Apache, students’ independent study, targeted papers. Backup Slides Apache Rave Apache Rave • Open Community Software for Enterprise Social Networking, Shareable Web Components, and Science Gateways • Founding members: • • • • Mitre Software SURFnet Hippo Software Indiana University • More information • Project Website: http://rave.apache.org/ • Mailing List: [email protected] 1 Gadget Dashboard View Gadget Store View Extending Rave for Science Gateways • Rave is designed to be extended. – Good design (interfaces, easily pluggable implementations) and code organization are required. – It helps to have a diverse, distributed developer community • How can you work on it if we can’t work on it? • Rave is also packaged so that you can extend it without touching the source tree. • GCE11 paper presented 3 case studies for Science Gateways Apache Software Foundation and Cyberinfrastructure Why Apache for Gateway Software? • Apache Software Foundation is a neutral playing field – 501(c)(3) non-profit organization. – Designed to encourage competitors to collaborate on foundational software. – Includes a legal cell for legal issues. • Foundation itself is sustainable – Incorporated in 1999 – Multiple sponsors (Yahoo, Microsoft, Google, AMD, Facebook, IBM, …) • Proven governance models – Projects are run by Program Management Committees. – New projects must go through incubation. • Provides the social infrastructure for building communities. • Opportunities to collaborate with other Apache projects outside the usual CI world. The Apache Way • Projects start as incubators with 1 champion and several mentors. – Making good choices is very important • Graduation ultimately is judged by the Apache community. – +1/-1 votes on the incubator list • Good, open engineering practices required – DEV mailing list design discussions, issue tracking – Jira contributions – Important decisions are voted on • Properly packaged code – – – – Build out of the box Releases are signed Licenses, disclaimers, notices, change logs, etc. Releases are voted • Developer diversity – Three or more unconnected developers – Price is giving up sole ownership, replace with meritocracy Apache and Science Gateways • Apache rewards projects for cross-pollination. – Connecting with complementary Apache projects strengthens both sides. – New requirements, new development methods • Apache methods foster sustainability – Building communities of developers, not just users – Key merit criterion • Apache methods provide governance – Incubators learn best practices from mentors – Releases are peer-reviewed Apache Contributions Aren’t Just Software • Apache committers and PMC members aren’t just code writers. • Successful communities also include – – – – Important users Project evangelists Content providers: documentation, tutorials Testers, requirements providers, and constructive complainers • Using Jira and mailing lists – Anything else that needs doing. Case Study: LEAD • To create an Integrated, Scalable Geosciences Framework, LEAD among things resulted in a developing a flexible Scientific workflow system. • The initial goal was to realize WOORDS: Workflow Orchestration for On-Demand Real-Time Dynamically-Adaptive System. • The system enables execution of legacy scientific codes and facilities sophisticated coupling while interacting with data and provenance sub-systems. Case Study: One Degree Imager • • • • • A single investigation requires multiple night observations Each night takes hours of observations with multiple exposures An exposure is divided into 64 Orthogonal Transfer Array (OTAs) Each OTA is an 8x8 collection of 512x512 pixel CCD images. Reducing these data sets require workflow planning taking advantage of system architectures. • Currently we take advantage of threaded parallelism at node level branching out to multiple node executions. Pipeline Parallelization TOP Campaign ... FTR Night/Filter FTR ... EXP Exposures EXP ... OTAs OTA OTA Illustrating Interactivity Asynchronous refinements Applica on Steering Job Level Interac ons Orchestra on level Interac ons Parametric Sweeps Provenance Workflow Steering Mathema cal Domain Job launch, gliding Resource Uncertain es Checkpoint/ Restart Model Refinement Execution Patterns Parametric Sweeps Start Start Level 0 4 instances X 4 à 16 outputs Level 1 2instances X (4x4)à 32 outputs Level 2 1 instance X (32x32)à 1024outputs B A C Pruned Computation Dot vs Cartisian Start! ! ! ! !!!!4x4!instances!!16!outputs! !! ! ! Start !! ! !! Level% 0! ! Level 1 ! ! ! 2x16 instances! 32 outputs Level 2 ! instances! 256 outputs 1x256 A! B! C! Interactivity Contd. • Deviations during workflow execution that do not affect the structure of the workflow – dynamic change workflow inputs, workflow rerun. interpreted workflow execution model. – dynamic change in point of execution, workflow smart rerun. – Fault handling and exception models. • Deviations that change the workflow DAG during runtime – Reconfiguration of activity. – Dynamic addition of activities to the workflow. – Dynamic remove or replace of activity to the workflow