High Performance Extreme Computing, Data Processing, and

High Performance Extreme
Computing, Data Processing,
and Visualization
Si Liu
Software Engineering Assembly 2014
Apr 7-11, 2014
Team members:
Greg Abram
John Cazes
Greg Foss
Si Liu
Don Cook
Craig Stair
Dave Gill
Jordan Powers
WRF nested runs
Memory requirement
IO efficiency and workflow
Data analysis and visualization
Conclusion and future work
Raytheon R&D project
•  Highly localized weather modeling
•  Weather simulation for several hours
–  extremely high spatial resolution
–  extremely high temporal resolution
•  High temporal resolution data collection for
animated demonstration
Domain of Interest
•  Around O'Hare International Airport,
Chicago, Illinois
–  Longitude: 87.9047W
–  Latitude: 41.9786N
•  Cover the cylinder area
–  Diameter: 120 nautical miles
About 222 km
–  Height: 70,000 feet
About 21 km
•  The expected resolution:
–  Horizontal: 167 meter or finer
–  Vertical: 300 feet or finer
About 91 meters
Weather Research and Forecasting Model
•  Open source community software developed and
supported by NCAR and collaborative partners
•  Parallel mesoscale weather model
•  Used for both research and operational forecasts
•  A large worldwide community of users (over 20,000 in
over 130 countries)
WRF Nested Runs
•  A fine-grid run based on the parent coarse-grid run
•  Cover only a portion of the parent domain
•  Lateral boundaries driven from the parent domain
•  Why nested runs:
–  High resolution model run over the large domain is
prohibitively expensive (memory, storage, computing)
–  High resolution for a very small domain with mismatched time
and spatial lateral boundary conditions
–  Other reasons
Two-way nested runs
Run parent and child (nested) domains in a single
simulation with information exchange between them
•  Parent simulation: 500m horizontal
•  Child (nested) simulation:167m horizontal
•  Still expensive (at least doubles the problem size we
finally solve in every single run)
One-way Nested Runs
•  Achieved with ndown.exe program
•  A finer-grid run is made as a subsequent run after the
coarser-grid run
–  Make a complete coarse-grid run (500m horizontal)
and collect output data
–  Create the initial and lateral boundary conditions for
the fine-grid run with the WRF ndown.exe program
–  Run the fine-grid simulation (167m horizontal) with the
input files generated in the previous step
Vertical refinement
•  Achieved through ndown.exe
•  Original plan:
100 vertical levels (parent domain)
à around 200/300/500 vertical levels (child domain)
•  Vertical refinement is yet limited in WRFV-3.4.* and 3.5.*.
•  Current implementation:
234 vertical levels (parent domain)
à 234 vertical levels (child domain)
Nested Domains Sketch Map
•  Outer domain
–  1345 x 1345 grid cells (500 m)
–  234 vertical level (300 feet)
•  Inner domain
–  1345x1345 grid cells (167 m)
–  234 vertical levels (300 feet )
•  Nested ratio
–  Horizontal 3:1
–  Vertical 1:1
WRF Workflow
•  Obtain the Global Forecast System (GFS) model data
•  Run geogrid.exe, ungrib.exe, and metgrid.exe in WRF
Preprocessing Systems (WPS)
•  Run real.exe to generate the initial and lateral boundary condition
files for the coarse-grid run
•  Make a coarse-grid run (only a few output files are necessary)
•  Re-run geogrid.exe and metgrid.exe for both parent and nested
•  Re-run real.exe for both parent and nested domains
•  Execute ndown.exe to generate the fine-grid initial and lateral
boundary conditions
•  Make the fine-grid run and produce output files frequently as
TACC Stampede System
•  Dell Linux Cluster
•  6,400+ Dell PowerEdge server nodes
–  2 Intel Xeon E5 (Sandy Bridge) processors
–  1 Intel Xeon Phi Coprocessor (MIC Architecture)
•  The aggregate peak performance
–  Xeon E5 processors: 2+ PF; Xeon Phi processors: 7+ PF
•  Login nodes, large-memory nodes, graphics nodes
•  Global parallel Lustre file system
WRF OOM Crisis
•  OOM is a normal problem to such high-resolution
simulations. WRF is no exception.
•  WRF has not supported vertical partitions yet.
Monitor WRF Memory Usage
•  Each task needs a huge amount of memory.
•  Task zero may require more memories than others.
•  TACC Stats: http://tacc-stats.tacc.utexas.edu/
How to Meet the Memory Requirement (1)
•  Original settings
•  Fewer MPI tasks per node
How to Meet the Memory Requirement (2)
•  SLURM reconfiguration is necessary.
•  One dedicated node for Task 0
32 GB
32 GB
32 GB
32 GB
•  One dedicated large-memory node for Task 0
1 TB
32 GB
32 GB
32 GB
IO Workload
•  Each output file is huge:
–  More than 400 million grid points
–  About 200 variables
–  11+ GB per file
•  Record the output every 3 model seconds:
–  Generate 20 files per model minutes, 1200 files per model hour
•  Serial IO and Parallel IO
–  Wasting a lot of computing resources
–  Slow down the whole system (so many IO requests)
IO Techniques
•  Restrict output variables
–  Modified Registry.EM_COMMON (WRF recompiled)
–  Cut the output file size by 30-50%
•  WRF checkpoint and restart mechanism
–  Make job complete within the wallclock limit
–  Reduce the risk of job failure and data loss
–  Validate the output data after every single run
•  Use local disk to temporarily keep the output
–  Local /tmp space (68-70GB available disk space)
•  WRF output files and WRF restart files partition
–  About 10 minutes à about 0.5s per output step
Merge split WRF output
•  Regroup split WRF output files
–  Task-based à Time-step-based
–  Several “tar/untar” work to reduce the Metadata Server workload
of the Lustre file system
•  Merge split WRF output
Advanced Regional Prediction System (ARPS)
Thousands of sequential jobs
Large-memory node
TACC Parametric Job Launcher Utility
utility for submitting multiple serial applications simultaneously
IO Workflow
Local Disc
IO Cost Summary
Traditional IO workflow Method
Our IO workflow
Time per step
10 minutes
0.4-0.5s on average (1024
Total Time
more than 8 days
About 8-10 minutes
Extra time for
data processing
8-10 hours depending on the
computing resources
Target data files 1201 wrf-out files, 11 GB each
13TB in total for one-hour
1201 wrf-out files, 11GB each
13TB in total for one-hour run
Extra space
256 tar ball files, 36GB each
10TB in total
Data Analysis and Visualization
•  WRF output uses pressure values to identify altitude
whereas visualization software requires coordinate
values in the height axis
•  Translate WRF output files to VTK files (Python)
–  converting pressure values into Z coordinates
–  irregular grid
•  Resulting VTK files are read into ParaView
•  Target variables are modeled as a surface
•  For a generalized aviation reference, an aviation map
provided by Raytheon is included for background in the
Achievements and Conclusions
•  We model a specific time frame and region to provide
meteorological data with high spatial and temporal
•  The resolution in time and space is well beyond what
has ever been attempted with previous WRF simulations.
•  The memory management techniques are portable to
other programs and other platforms.
•  The new-designed IO workflow is more convenient and
Future work
•  Compare with other experimental data for validation
•  A follow-on study will be performed over the other
domains of interests
•  Even-higher resolution simulations if necessary
•  Understand the WRF design and exceed the limit of WRF
•  Investigate optimized IO workflow with T3PIO library
•  Improving WRF performance with Xeon Phi
Special thanks to
•  Ming Chen (NCAR)
•  Bill Barth, Tommy Minyard, Todd Evans (TACC)
Si Liu
[email protected]
For more information: