loading page

Parmap: Analytics Engine Scalable for Climate Model Evaluation on Cloud and High-Performance Computing Platforms
  • Joseph Jacob,
  • Brian Wilson,
  • Huikyo Lee
Joseph Jacob
Jet Propulsion Laboratory, California Institute of Technology

Corresponding Author:joseph.jacob@jpl.nasa.gov

Author Profile
Brian Wilson
Jet Propulsion Laboratory, California Institute of Technology
Author Profile
Huikyo Lee
Jet Propulsion Laboratory, California Institute of Technology
Author Profile

Abstract

The need to better understand climate change has driven model simulations to greater fidelity with improved spatiotemporal resolution (e.g., < 10 km at sub-hourly cadence). For example, the 7 km GEOS-5 Nature Run (G5NR) with 30-minute outputs from 2005-07 at the NASA Center for Climate Simulation (NCCS) is ~4 PB and is not easily portable. The rise of these high-fidelity climate models coincides with the emergence of cloud computing as a viable platform for scientific analytics. NASA has adopted a cloud computing strategy using public providers like Amazon Web Services (AWS). However, it is not cost- or time- effective to move the High- Performance Computing (HPC)-based model computations and data to the cloud. Thus, there is a need for scalable model evaluation compatible with both the cloud and HPC platforms like NCCS. To fill this need we have extended the analytics component of the Apache Science Data Analytics Platform (SDAP) with a streamlined version that specifically targets high-resolution science data products and climate model outputs on a regular coordinate grid. Gridded inputs (as opposed to other data structures like point clouds or swath-based measurements supported by SDAP), enable offsets to particular grid cells to be directly computed, allow for processing on the original NetCDF or HDF granules, do not require a second tiled copy of the data, and accommodate a simpler technology stack since no geospatial database is required for lookups or tile storage. Our core module, Parmap, abstracts the map-reduce model so that users can select from a variety of map computational modes, including Spark, Dask, serverless AWS Lambda, PySparkling, and Python multiprocessing. Example analytics include area-averaged time series and time-averaged, correlation and climatological maps. Benchmarks compare favorably with the full SDAP implementation.