Parallel-UVCDAT / Python for Diurnal Cycle Analysis

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Monday, 5 January 2015
Curt Covey, LLNL, Livermore, CA; and C. Doutriaux and D. Williams

The "average" diurnal cycle of temperature, precipitation, etc. is a link between long-term climate and the hour-by-hour evolution of weather. Modern space-based observations provide a picture in unprecedented detail: for example the Tropical Rainfall Measuring Mission (TRMM) has produced data at 0.25 degree latitude/longitude resolution over the most of the globe at 3-hour intervals for the past 15 years. Analysis of such data can be embarrassingly parallel. A separate avenue of parallel processing arises from comparison of these observations with output from the two-dozen climate models providing 3-hourly data to the Coupled Model Intercomparison Project (CMIP; see Taylor et al., BAMS 2012). Observations from TRMM and other space-based datasets are available on the same network, and in the same format, as CMIP data (Teixeira et al., BAMS in press). We will present a comparison of model simulations with each other and with TRMM data, enabled by parallel-processing features of the Ultrascale Visualization Climate Data Analysis Tools (UVCDAT) extension of Python (Williams et al. IEEE Computer 2013). Despite its unprecedented detail, the TRMM dataset is small enough to allow simple temporal parallelism in which each node of a high-performance computing cluster is responsible for a time-segment's entire data. For example, the Linux cluster sierra.llnl.gov has 12 processors and 24 GB shared memory per node, while a month of TRMM 3-hourly data on 576,000 latitude-longitude grid points takes up 31 days * 8 timepoints/day * 576000 grid points * 4 bytes/grid point = 0.6 GB, providing ample room for each of the 12 processors to share local memory without contention. This consideration makes the TRMM data ideal for “production mode” parallelism with Python/UVCDAT. The nearby figure shows a sample from the TRMM database processed serially. The figure gives the average local time of maximum precipitation for each day of July 2001 for each grid point with > 1.5 mm/day rainfall. Consistent with previous analyses (e.g. Dai et al., Climate Dynamics 2007, Fig. 5j) the figure includes late afternoon maxima in the summer monsoon regions of Africa and South/Southeast Asia, as well as variation of the times of maxima from west to east in the center of North America, indicating a propagation of rainfall from the Rockies to the Great Plains during the day. Also evident in our preliminary results--but not shown here--is the expected predominance of the diurnal (24 hour) Fourier harmonic, albeit with an additional semidiurnal (12 hour) harmonic that tends to broaden the precipitation-maximum peak. Despite these encouraging results, the noisiness apparent in the figure and others we have produced suggests that averaging over much longer time periods could enhance the visual coherence and statistical significance of our results. Also, other versions of the satellite observations that rely on more direct measurements of rainfall can be more accurate (albeit at the price of sparser spatial coverage; see Kikuchi and Wang, J Climate 2008). Together with the need to include model simulations, some of which are almost as high resolution as TRMM, these considerations argue for a systematic parallel processing of our diagnostic computations. We will present the results of this parallel processing. LLNL-ABS-657869: This work was performed under auspices of the Office of Science, US Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.