Deep Learning @15 Petaflops/second: Semi-supervised pattern detection for 15 Terabytes of climate data (Invited)

Thursday, December 14, 2017 - 10:50
Add to Calendar

Anthropogenically-forced climate changes in the number and character of extreme storms have the potential to significantly impact human and natural systems. Current high-performance computing enables multidecadal simulations with global climate models at resolutions of 25km or finer. Such high-resolution simulations are demonstrably superior in simulating extreme storms such as tropical cyclones than the coarser simulations available in the Coupled Model Intercomparison Project (CMIP5) and provide the capability to more credibly project future changes in extreme storm statistics and properties. The identification and tracking of storms in the voluminous model output is very challenging as it is impractical to manually identify storms due to the enormous size of the datasets, and therefore automated procedures are used. Traditionally, these procedures are based on a multi-variate set of physical conditions based on known properties of the class of storms in question.

In recent years, we have successfully demonstrated that Deep Learning produces state of the art results for pattern detection in climate data. We have developed supervised and semi-supervised convolutional architectures for detecting and localizing tropical cyclones, extra-tropical cyclones and atmospheric rivers in simulation data. One of the primary challenges in the applicability of Deep Learning to climate data is in the expensive training phase. Typical networks may take days to converge on 10GB-sized datasets, while the climate science community has ready access to O(10 TB)-O(PB) sized datasets.

In this work, we present the most scalable implementation of Deep Learning to date. We successfully scale a unified, semi-supervised convolutional architecture on all of the Cori Phase II supercomputer at NERSC. We use IntelCaffe, MKL and MLSL libraries. We have optimized single node MKL libraries to obtain 1-4 TF on single KNL nodes. We have developed a novel hybrid parameter update strategy to improve scaling to 9600 KNL nodes (600,000 cores). We obtain 15PF performance over the course of the training run; setting a new watermark for the HPC and Deep Learning communities. This talk will share insights on how to obtain this extreme level of performance, current gaps/challenges and implications for the climate science community.

Link for More Information: