[R-sig-hpc] A few questions on using R and CUDA

[Nafis Sadat]

Hi everyone:

My name is Nafis Sadat, and I am a software engineer at the University of Washington. I've been doing some research on the side recently on looking at R and GPU capabilities, and after pinging Dr. Dirk Eddelbuettel (after finding out his work on the 'gcbd' package), he suggested that I can send out my question to the r-sig-hpc mailing list hoping for some answers!

I work in the geospatial sciences team in my institute, where we try to model diseases at the pixel level around the world using spatial modeling methods. The models that our researchers use are written using the R packages INLA (http://www.r-inla.org/) and TMB (https://github.com/kaskr/adcomp/wiki); all of the code base is entirely in R. One issue that we simply can't avoid is the large dimension of the data (so, if we were modeling at a finer mesh for example, we would have more parameters to optimize, and that leads to longer model runtimes).

A year ago, there was no work done in the team to try and optimize this part of the process, but then I worked with another software engineer to compile R from source using the BLAS/LAPACK provided by Intel MKL, and that sped up things pretty decently when it came to a lot of matrix algebra. In fact, both the INLA and TMB packages are built by linking against BLAS (and I have confirmed that the MKL is indeed being used when I run the model by doing a verbose trace). Fast forward to now, and I came across this page: https://developer.nvidia.com/cublas.

I have a machine lying around at home which has an Nvidia GPU and the line where Nvidia claims that "cuBLAS performs up to 35X faster than the latest version of the MKL BLAS on common benchmarks". I have not looked into super details about the backend of this (and only recently I have been coming across the .cu extensions...). In an ideal world, if I could simply build R by linking to the CUDA BLAS instead of the MKL BLAS and have everything else run as-is, then that would be the dream, but after fudging around earlier today where I tried to symlink 'libnvblas.so' to 'libblas.so', that build of R was quite a failure.

If anyone here would happen to have any thoughts or suggestions on this, I would absolutely love your inputs. I have not found any answers on this on Google, and I just posted a question on Stack Overflow, so I'm hoping that I can learn more about this!

Thanks everyone for your patience in reading through this!

Nafis




Nafis Sadat
Software Engineer
Institute for Health Metrics and Evaluation | University of Washington
2301 5th Avenue, Suite 600 | Seattle, WA 98121| USA
Tel: +1-206-897-3726<tel:+1%20206-897-3726> | Fax: +1-206-897-2899<tel:+1%20206-897-2899> | Campus Mailbox: 358210
sadatnfs using uw.edu<mailto:sadatnfs using uw.edu> | http://www.healthdata.org<http://www.healthdata.org/>



	[[alternative HTML version deleted]]