Accelerating climate kernels

Project reference: 1718

Climate research is a major user of supercomputing facilities, however the spatial resolution of climate models are much coarser than similar models for weather forecasting, and thus they no longer scale as the computers become bigger. Instead climate research is highly dependent on increasing the time a simulation covers, into hundred of thousands of years, with a time resolution of six hours. Thus for climate research to improve the field needs faster computers, not bigger computers.

This means that accelerators are very interesting for climate research, GPGPUs, Xeon Phis, and FPGAs.

The project identifies two of computational kernels that should be ported to run on accelerators:
⦁ Ocean density solver
⦁ Barotropic solver

Global projections of surface temperature monthly mean for January 2099.

Project Mentor: Mads R. B. Kristensen

Site Co-ordinator: Brian Vinter

Learning Outcomes:

Learn the performance and programming characteristic of different accelerators particularly GPGPUs and Xeon Phis.

Identifying code that are suitable for a specific accelerator.

Student Prerequisites (compulsory): 

Basic linux skills

Elementary knowledge of parallel computing

Student Prerequisites (desirable): 

Experience with OpenCL/CUDA and OpenMP

Training Materials:

The memory layout of the 2nd generation Intel® Xeon Phi™ processors code-named Knights Landing (KNL):

MCDRAM as High-Bandwidth Memory (HBM) in Knights Landing Processors: Developer’s Guide

OpenCL developing guide:
http://developer.amd.com/tools/heterogeneous-computing/amd-accelerated-parallel-processing-app-sdk

Workplan:

Week 1: Training
Week 2-3: Study of numerical solvers for climate simulations
Week 4-7: Identify and port code that will benefit from accelerators
Week 8: Finalise Report

Final Product Description: 

The final product is accelerated numerical solvers for climate simulations

Adapting the Project: Increasing the Difficulty:

Include more numerical solvers and accelerators

Resources:

The student will need access to machines with:
⦁ GPGPU
⦁ Xeon Phis
Which we will provide.

Organisation:

Niels Bohr Institute University of Copenhagen