Parallel Ordinary Kriging using OpenCL and OpenMP

Description

This is an implementation of the Ordinary Kriging algorithm using parallel technologies.

Compiling

We provide a CMake script in order to make compiling the code straightforward. You just need a C++11 compiler with OpenMP support and the OpenCL SDK installed. We will provide detailed information in the future.

Running

After compiling the code you should see ParallelOK executable in the build folder. To use it simple do

ParallelOK --input [XYZ File] --output [Output File] --lags-count [N] --grid-size [N]

• --lags-count [N]: The number of lags when generated the empirical semivariogram.
• --grid-size [N]: Creates a NxN grid to make predictions.

Optional Arguments

• --profile: Will print detailed information about steps runtimes
• --platform [ID]: Select the OpenCL platform with ID to run OpenCL
• --num-devices [N]: Number of devices to use, omit to use all available devices.
• --run-serial: If present will run a serial version of the Ordinary Kriging. This option forces the program to run in serial mode even if --platform was provided.

XYZ File

The XYZ File is a simple point cloud format where each line represents a point in 3D space. For the kriging algorithm, each z value is considered a response value for a random variable at location (x,y).

Example of a XYZ file

-113.24195 40.01692 -2.07465272
-71.99837 38.22831 -0.58196695
-81.89808 28.22333 -1.54724761
-85.5524 49.05601 0.99257163

Future Work

• Structure the code as a library to incorporate in other applications
• Review the matrix-vector multiplication code for GPUs

Contact

Please, feel free to file an issue directly in this repository.

Implementation Results

To show that our implementation can scale across multiple compute devices we show here two graphs. The first is the amount of time spent by the prediction step with the increase in the number of compute devices. The second graph is showing that our implementation can achieve a quasi-linear speedup with the increase in the number of devices being used.

The experiments were performed on a machine with 64 AMD Opteron(TM) Processor 627 cores grouped into 8 physical chips with 128GB of RAM memory and 4 NVIDIA Tesla M2090 GPUs each with 2096GB of GDDR5 RAM 384-bit RAM memory.

Our results were published in the CCIS 2016. The link to the full paper can be found below.

Cite this work

@article{krigging,
  author = {Sabino, Thales Luis and Tavares, Gisele and Goliatt, Leonardo and Lobosco, Marcelo and Chaves, Filipe and Santos, Rodrigo},
  year   = {2017},
  month  = {08},
  pages  = {},
  title  = {A parallel implementation of the ordinary kriging algorithm for heterogeneous computing environments}
}

A parallel implementation of the ordinary kriging algorithm for heterogeneous computing environments