Parvis is exploring the use of  Pagoda to replace  NCO in the post-processing of ultra-large climate model output files.

Pagoda is written with the  Global Arrays toolkit.

PagodaTimings on large CAM HOMME files.

The dimension organization of variables affects io performance because it affects the size of read blocks assuming the entire file isn't read into memory at once. Currently, most data is stored as [time, level, lat, lon]. This graph shows that placing the vertical as the last dimension enables scalable data distribution on the larger dimension. pgra (using PNetCDF) on 1/8 degree data. Number of levels = 26. Problem occurs when going from 16 to 32 processors. (horizontal axis = processes, vertical = time). Pagoda timings on Eureka