Science on the TeraGrid Daniel S. Katz [email protected]Director of Science, TeraGrid GIG Senior Computational Researcher, Computation Institute, University of Chicago & Argonne National Laboratory Affiliate Faculty, Center for Computation & Technology, LSU Adjunct Associate Professor, Electrical and Computer Engineering Department, LSU
34
Embed
Science on the TeraGrid Daniel S. Katz [email protected] Director of Science, TeraGrid GIG Senior Computational Researcher, Computation Institute, University.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Senior Computational Researcher, Computation Institute, University of Chicago & Argonne National
Laboratory
Affiliate Faculty, Center for Computation & Technology, LSU
Adjunct Associate Professor, Electrical and Computer Engineering Department,
LSU
Outline
•Introduction to TeraGrid
•Current Science
•Future Science
What is the TeraGrid• World’s largest distributed cyberinfrastructure for open scientific
research, supported by US NSF
• Integrated high performance computers (>2 PF HPC & >27000 HTC CPUs), data resources (>3 PB disk, >60 PB tape, data collections), visualization, experimental facilities (VMs, GPUs, FPGAs), network at 11 Resource Provider sites
• Allocated to US researchers and their collaborators through national peer-review process
• DEEP: provide powerful computational resources to enable research that can’t otherwise be accomplished
• WIDE: grow the community of computational science and make the resources easily accessible
• OPEN: connect with new resources and institutions
• Integration: Single: portal, sign-on, help desk, allocations process, advanced user support, EOT, campus champions
Governance
•11 Resource Providers (RPs) funded under separate agreements with NSF– Different start and end dates– Different goals– Different agreements– Different funding models
•1 Coordinating Body – Grid Integration Group (GIG)– University of Chicago/Argonne National Laboratory– Subcontracts to all RPs and six other universities– 7-8 Area Directors– Working groups with members from many RPs
•Terminal: ssh, gsissh•Portal: TeraGrid user portal, Gateways– Once logged in to
portal, click on “Login”
•Also, SSO from command-line
Science Gateways
•A natural extension of Internet & Web 2.0• Idea resonates with Scientists
– Researchers can imagine scientific capabilities provided through familiar interface• Mostly web portal or web or client-server program
•Designed by communities; provide interfaces understood by those communities– Also provide access to greater capabilities (back end)– Without user understand details of capabilities– Scientists know they can undertake more complex analyses
and that’s all they want to focus on– TeraGrid provides tools to help developer
•Seamless access doesn’t come for free– Hinges on very capable developer
Slide courtesy of Nancy Wilkins-Diehr
TeraGrid -> XD Future
•Current RP agreements end in March 2011– Except track 2 centers (current and future)
•TeraGrid XD (eXtreme Digital) starts in April 2011– Era of potential interoperation with OSG and others– New types of science applications?
•Current TG GIG continues through July 2011– Allows four months of overlap in coordination– Probable overlap between GIG and XD members
•Blue Waters (track 1) production in 2011
12
Outline
•Introduction to TeraGrid
•Current Science
•Future Science
TG App: Predicting storms• Hurricanes and tornadoes cause massive
loss of life and damage to property• TeraGrid supported spring 2007 NOAA and
University of Oklahoma Hazardous Weather Testbed– Major Goal: assess how well ensemble
forecasting predicts thunderstorms, including the supercells that spawn tornadoes
– Nightly reservation at PSC, spawning jobs at NCSA as needed for details
– Input, output, and intermediate data transfers– Delivers “better than real time” prediction– Used 675,000 CPU hours for the season– Used 312 TB on HPSS storage at PSC
Slide courtesy of Dennis Gannon, ex-IU, and LEAD Collaboration
TG App: SCEC-PSHA
• Part of SCEC (Tom Jordan, USC)
• Using the large scale simulation data, estimate probablistic seismic hazard (PSHA) curves for sites in southern California (probability that ground motion will exceed some threshold over a given time period)
• Used by hospitals, power plants, schools, etc. as part of their risk assessment
• For each location, need a Cybershake run followed by roughly 840,000 parallel short jobs (420,000 rupture forecasts, 420,000 extraction of peak ground motion)– Parallelize across locations, not individual
workflows
• Completed 40 locations to date, targeting 200 in 2009, and 2000 in 2010
Managing these requires effective grid workflow tools for job submission, data management and error recovery, using Pegasus (ISI) and DAGman (Wisconsin)
15Information/image courtesy of Phil Maechling
TG App: GridChem
Slide courtesy of Joohyun Kim
TG Apps: Genius and Materials
HemeLB on LONILAMMPS on TeraGrid
Fully-atomistic simulations of clay-polymer nanocomposites
Slide courtesy of Steven Manos and Peter Coveney
Why cross-site / distributed runs?
1.Rapid turnaround, conglomeration of idle processors to run a single large job
2.Run big compute & big memory jobs not possible on a single machine
Modeling blood flow before (during?) surgery
Outline
•Introduction to TeraGrid
•Current Science
•Future Science
ENZO• ENZO simulated cosmological
structure formation• Big current production simulation:
– 4096x4096x4096 non-adaptive mesh, 16 fields per mesh point
– 64 billion dark matter particles – About 4000 MPI processes,
1-8 OpenMP threads per process– Reads 5 TB input data– Writes 8 TB data files
• Restart reads latest 8 TB file
– All I/O uses HDF5, each MPI process reading/writing their own data– Over few months for simulation, >100 data files written, >20 read for
restarts– 24 hour batch runs
• 5-10 data files output per run• Needs ~100 TB free disk space at start of run
– (adaptive case is different, but I/O is roughly similar)
Slide courtesy of Robert Harkness
ENZO Calculation Stages1. Generate initial conditions for density, matter velocity field, dark matter particle
position and velocity (using parallel HDF5)– Using NICS Kraken w/ 4K MPI processes, though TACC Ranger is reasonable alternative– ~5 TB initial data created in 10 0.5-TB files
2. Decompose initial conditions for # of MPI tasks in simulation (using sequential HDF5)– Decomposition of mesh into "tiles" needed for MPI tasks requires strided reads in large data cube
• Very costly on NICS Kraken, but can be done more efficiently on TACC Ranger• If done on Ranger, then 2 TB (4 512-GB files) must be transmitted from NICS to TACC, and after running MPI
decomposition task (with 4K MPI tasks), 8K files (2 TB) must be returned to NICS
– Dark matter particle sort onto "tiles" is most efficient on NICS Kraken because it has a superior interconnect• Sort usually run in 8 slices using 4K MPI tasks
3. Evolve time (using sequential HDF5) – Dump data files during run– Archive data files (8 TB every couple of hours -> 1100 MB/sec but NICS HPSS only reaches 300
MB/sec)
4. Derive data products – Capture 5-6 fields from each data file (~256 GB each)– Send to ANL or SDSC for data analysis or viz– Archive output of data analysis of viz (back at NICS)
Overall run produces >1 PB data, >100 TB to be archived, >100 TB free disk space
Slide courtesy of Robert Harkness
Science Portals 2.0• Workspace customized by user for specific projects• Pluggable into iGoogle or other compliant (and open source) containers• Integrates with user workspace to provide a complete and dynamic view of the user’s
science, alongside other aspects of their lives (e.g. weather, news)• Integrates with social networks (e.g. FriendConnect, MySpace) to support collaborative
science• TG User Portal provides this same information, but this view is more dynamic and more
reusable, and can be more flexibly integrated into the user’s workspace• Gadgets suitable for use on mobile devices
Technology Detail• Gadgets are HTML/JavaScript embedded in XML• Gadgets conform to specs that are supported by many containers (iGoogle, Shindig, Orkut, MySpace)
Technology Detail• Gadgets are HTML/JavaScript embedded in XML• Gadgets conform to specs that are supported by many containers (iGoogle, Shindig, Orkut, MySpace)
Resource Load gadget shows current view of load on available resourcesResource Load gadget shows current view of load on available resources
Job Status gadget shows current view of job queues by site, user, statusJob Status gadget shows current view of job queues by site, user, status
File Transfer gadget for lightweight access to data stores, simple global file searching, and reliable transfers
File Transfer gadget for lightweight access to data stores, simple global file searching, and reliable transfers
Domain science gadgets complement general purpose gadgets to encompass the full range of scientists’ interests
Domain science gadgets complement general purpose gadgets to encompass the full range of scientists’ interests
Slide courtesy of Wenjun Wu, Thomas Uram, Michael Papka
Done with Andre Merzky, Katerina Stamou, Shantenu Jha
SAGA-based DAG ExecutionPreserving Performance
Done with Andre Merzky, Katerina Stamou, Shantenu Jha
32
Credit: Yaakoub El-Khamra, Shantenu Jha
• Suppose you want to perform a history match on a 1 million grid cell problem, with a thousand ensemble members
• The entire system will have a few billion degrees of freedom
• This will increase the need for autonomy, fault tolerance, self healing etc...
App: EnKF for Oil Reservoir Sim.
• Ensemble Kalman Filters – recursive filters that can be used to handle large, noisy data; the data are sent through the Kalman filter to obtain the true state of the data
• Used here to model oil field, w/ data from sensors, to build a model that represents reality
• Using EnKF, scientific problem solved using “multiple models” (ensemble members)
• Physical models represented as ensembles that vary in size from large MPI-style jobs to long-running single processor tasks
• Varying parameters sometimes also lead to varying systems of equations and entirely new scenarios, increasing both computational and memory requirements
• Each model must converge before the next stage can begin, hence dynamically load-balancing to ensure that all models complete as close to each other as possible is a desired aim
• In the general case the number of jobs required varies between stages.
EnKF Results: Scaling-Out
•Using more machines decreases the TTC and variation between experiments
•Using BQP (batch queue prediction) decreases the TTC & variation between experiments further
•Lowest time to completion achieved when using BQP and all available resources
Credit: Yaakoub El-Khamra, Shantenu Jha
Khamra & Jha, GMAC, ICAC’09
R=Ranger, Q=Queen Bee, A=Abe
TeraGrid: Both Operations and Research
•Operations – Facilities/services on which researchers rely– Infrastructure on which other providers build
AND
•R&D – Learning how to do distributed, collaborative science on a
global, federated infrastructure– Learning how to run multi-institution shared infrastructure