1. PyStormTracker: A Parallel Object-Oriented Cyclone Tracker in Python
Albert Yau School of Marine and Atmospheric Sciences Stony Brook University
Kevin Paul and John Dennis Application Scalability and Performance (ASAP) Group Computational and Information Systems Laboratory National Center for Atmospheric Research
11/17/2018

2. Outline Scientific Background Motivation and Objectives Implementation
Sample Output
Benchmarks
Summary and Future Work
11/17/2018

3. What are cyclones?
A figure from a nineteenth century geography text showing storm frequency distribution (Hinman, 1888), adapted from Chang et al. (2002)
11/17/2018

4. Scientific Background
Cyclones are synoptic scale (~1000km) moving low pressure systems
Extra-tropical cyclones are responsible for much of the day-to-day weather variability and extreme weather events at mid-latitudes
Evolution of an extra-tropical cyclone as in the Norwegian Cyclone Model. (Source: weather.gov)
11/17/2018

5. Motivation
Tracking extra-tropical cyclones is challenging (Neu et al., 2013):
Cyclones vary greatly in size and shape, move with different velocities, split and merge over the lifetime
Most current cyclone trackers are written in FORTRAN, the codes are not readily available
Hard to modify, optimize, or compare existing cyclone trackers
11/17/2018

6. Objective 1
Implement a user and developer friendly cyclone tracker in Python using an object-oriented approach
Flexible and extensible by inheriting or replacing the classes and methods
Use existing scientific computing libraries such as NumPy and SciPy
The source codes will be publicly available at GitHub
11/17/2018

7. Motivation Emergence of high resolution and multi-ensemble model data
CESM Large Ensemble project: 40 ensemble members from for CMIP5 historical and RCP8.5 scenarios
Global Ensemble Forecast System (GEFS): up to 16 days of global forecasts from 21 ensemble members
GEFS ensemble members
Source: schumacher.atmos.colostate.edu
11/17/2018

8. Objective 2 To be able to process data from different datasets
GRIB: most commonly used in reanalysis and weather models: NCEP/NCAR Reanalysis and GEFS
netCDF: a more recent standard commonly used in climate models: CMIP5 archive of more than 30 climate models
PyNIO is chosen for file I/O
Utilize parallel computing technique to accelerate processing of large amount of data
MPI4Py as an interface to the MPI message passing library
11/17/2018

9. Implementation: Task Parallelism
grid
grid
centers
tracks
tracks
(meta data only)
Task 1
grid
centers
tracks
Task 2
grid
centers
tracks
(tree reduction)
Task 3
grid
centers
tracks
Task 4
grid
centers
tracks …
Task n
grid
centers
tracks
11/17/2018

10. Implementation: Classes
stormtracker.py
RectGrid class:
Rectilinear grid derived from Grid class
Does not read data until actually required
Split into smaller objects by time
Tracks class:
Input a list of Center
Match centers to existing Tracks using nearest neighborhood method
If matched, append to exisiting track; else append center as a new track
Merge two Tracks objects
stormtracker.py:
The main controller for generating Grid, Center and Track objects
Manage parallel processes
from detector import RectGrid, Center
from linker import Tracks
If __name__ == “__main__”:
grid = RectGrid(filepath=“…”)
centers = grid.detect()
tracks = Tracks()
tracks.append_center(centers)
Center class:
Detect centers from Grid data using scipy.ndimage.filters
Calculate great circle distances between centers
detector.py
linker.py
class Grid(object):
class Center(object):
def __init__(self, time, lat, lon, var): def abs_dist(self, center):
class RectGrid(Grid):
def __init__(self, filepath, trange): def get_time(self): def get_var(self, trange):
def split(self, num):
class Tracks(object):
def __init__(self): def match_center(self, centers): def append_center(self, centers): def merge_tracks(self, tracks):
The code repository will be available at:
11/17/2018

11. Sample Output CESM Large Ensemble #31:
6-hourly PSL
1-degree resolution
192*288 grid
Year 1990, 769 tracks
Tracks deeper than 975hPa, moved 1000km and longer than 2 days
11/17/2018

12. Benchmark: CESM 1 year – 1460 time steps
Number of MPI Tasks
Time
Detector
Linker
Total 1
569.55 2
517.25
297.34
814.58 4
261.58
160.38
421.96 8
136.39
88.17
224.56 16
77.21
48.20
125.40 32
46.99
30.73
77.72 64
30.44
21.43
51.86
128
20.80
17.59
38.39
256
17.67
15.86
33.53
512
15.77
15.17
30.93
1024
15.63
15.20
30.83
11/17/2018

13. Benchmark: CESM 15 years – 22360 time steps
Number of MPI Tasks
Time
Detector
Linker
Total 16
749.45 32
537.75
475.40 64
272.01
337.67
609.69
128
143.80
264.06
407.86
256
80.64
226.78
307.42
512
47.35
209.60
256.95
1024
30.71
200.80
231.50
2048
23.16
197.29
220.45
4096
19.33
196.02
215.36
11/17/2018

14. Summary and Future Work
The cyclone tracker is implemented using an object-oriented approach, which is flexible and easily extended:
For example: implement a CubedSphereGrid class; track tropical cyclones
11/17/2018

15. Summary and Future Work
We achieved reasonably good speedup in the detector, while the bottleneck is in the linker:
Reduce the communication cost by preselect tracks before merging
Instead of serializing (pickling) the objects, use buffered send/receive in MPI
11/17/2018

16. Summary and Future Work
Verify the quality of tracks:
Compare the tracks with other well established cyclone trackers
NCEP/NCAR Reanalysis:
Count of cyclone tracks deeper than 975hPa, moved at least 1000km and existed longer than two days in the north Pacific winter, compared with Hodges (1999) tracker
11/17/2018

17. Thank you! Questions? Email: albert.yau@stonybrook.edu
The code repository will be available at:
11/17/2018