TECA at 750,000 Cores: Lessons Learned from a Hero Run on Mira

Wednesday, May 14, 2014 - 07:00
Add to Calendar

The TECA (Toolkit for Extreme Climate Analysis) software, developed under the CASCADE project, enables petascale-class climate data analytics. TECA can be configured to extract extreme weather events, such as Tropical Cyclones, Atmospheric Rivers, Extra-Tropical Cyclones, etc from massive datasets. In this work, we present lessons learnt from the application of TECA to a large fraction of the CMIP-5 archive. Our end-to-end data analysis workflow consists of the following steps: - downloading 60TBs of CMIP-5 data via ESGF - storing and pre-processing data at NERSC - transferring 6TB sized dataset to ALCF over ESnet - running TECA at full concurrency on Mira We will present a performance data pertaining to end-to-end network bandwidth and server utilization. We will present TECA MPI initialization, Parallel I/O, and communication optimizations that enabled the full-scale Mira run. We also present preliminary results on the characterization of extra-tropical cyclone activity in CMIP-5 (historical period), and comparison to numerous reanalysis products (NCEP2, NCEP-CFSR, JRA55, ERA_Interim). Preliminary investigations seem to suggest a strong dependence of extra-tropical counts on model resolution. We are exploring the applicability of scaling relationships to account for resolution, followed by comparisons with RCP8.5 results to determine changes in ETC activity.

Presentation file(s):