CUDA Deep Dive

Motivation

CUDA 4.0 comes with some new exciting features such as:

• the ability to share GPUs across multiple threads;
• or use all GPUs in the system concurrently from a single host thread;
• and unified virtual addressing for faster multi GPU programming;

and many more.

The ability to access all the GPUs in a system is particularly nice on Amazon, since the large GPU enabled instances come with two Tesla M2050 Fermi boards, each capable of 1030 GFlops theoretical peak performance with 448 cores and 3GB of memory.

Getting started

Signing up to Amazon’s AWS is easy enough with a Credit Card, and once you are logged in, go to the EC2 tab of your console which should look something like this:

The EC2 console page.

Now press the Launch Instance button and in the Community AMIs tab set the Viewing option to Amazon Images and search for gpu and Select the CentOS 5.5 GPU HVM AMI and press Continue:

Choose the CentOS 5.5 GPU HVM AMI (bottom one).

Next we need to select the Instance Type and its important here to select the Cluster GPU type, and then press Continue:

Select the Cluster GPU Instance Type.

Next we need to Create a New Key Pair, by giving it a name like amazon-gpu and press Create & Download your Key Pair to download it to your local computer as a file called amazon-gpu.pem:

Create and download Key Pair.

We press Continue to go to the Firewall setting. Here we Create a new Security Group, give it a name and description, and then Create a new rule for ssh so that we can log into our instance once its up and running, and press Continue:

Create a new Security Group and a new ssh rule.

And finally we can review our settings and Launch it:

Review and Launch instance.

Back in our EC2 console we can go to our Instances and see our new AMI’s Status. It should be booting or running, rather than stopped as in the case below:

AMI Instance's Status and Description.

The Description tab will also contain the Public DNS which we can use together with the Key Pair we downloaded locally to ssh into our instance:

$ chmod 400 amazon-gpu.pem
$ ssh root@ec2-50-16-170-159.compute-1.amazonaws.com -i amazon-gpu.pem

__| __|_ ) CentOS
_| ( / v5.5
___|\___|___| HVMx64 GPU

Welcome to an EC2 Public Image
Please view /root/README


[root@ip-10-16-7-119 ~]#


Updating CUDA to 4.0

Now we need to update the CUDA driver and toolkit on our instance, so the first thing we do is to update the Linux Kernel and reboot the instance via the web console:

[root@ip-10-16-7-119 ~]# yum update kernel kernel-devel kernel-headers
Loaded plugins: fastestmirror
Determining fastest mirrors
* addons: mirror.cogentco.com
* base: mirror.umoss.org
* extras: mirror.symnds.com
* updates: mirror.umoss.org
addons | 951 B 00:00
base | 2.1 kB 00:00
base/primary_db | 2.2 MB 00:00
extras | 2.1 kB 00:00
extras/primary_db | 260 kB 00:00
updates | 1.9 kB 00:00
updates/primary_db | 635 kB 00:00
Setting up Update Process
Resolving Dependencies
--> Running transaction check
---> Package kernel.x86_64 0:2.6.18-238.12.1.el5 set to be installed
---> Package kernel-devel.x86_64 0:2.6.18-238.12.1.el5 set to be installed
---> Package kernel-headers.x86_64 0:2.6.18-238.12.1.el5 set to be updated
--> Finished Dependency Resolution

Dependencies Resolved

================================================================================
Package Arch Version Repository Size
================================================================================
Installing:
kernel x86_64 2.6.18-238.12.1.el5 updates 19 M
kernel-devel x86_64 2.6.18-238.12.1.el5 updates 5.5 M
Updating:
kernel-headers x86_64 2.6.18-238.12.1.el5 updates 1.2 M

Transaction Summary
================================================================================
Install 2 Package(s)
Upgrade 1 Package(s)

Complete!


I leave it as an exercise to figure out how to reboot the instance from the console, but once its back up and running, we can ssh back into it to download and install the CUDA 4.0 drivers, toolkit and SDK. For example:

[root@ip-10-16-7-119 ~]# wget http://developer.download.nvidia.com/compute/cuda
/4_0/toolkit/cudatoolkit_4.0.17_linux_64_rhel5.5.run
--2011-06-23 04:47:05-- http://developer.download.nvidia.com/compute/cuda/4_0/toolkit/cudatoolkit_4.0.17_linux_64_rhel5.5.run
Resolving developer.download.nvidia.com... 168.143.242.144, 168.143.242.203
Connecting to developer.download.nvidia.com|168.143.242.144|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 212338897 (203M) [application/octet-stream]
Saving to: `cudatoolkit_4.0.17_linux_64_rhel5.5.run'

100%[======================================>] 212,338,897 33.2M/s in 6.3s

2011-06-23 04:47:12 (32.0 MB/s) - `cudatoolkit_4.0.17_linux_64_rhel5.5.run' saved [212338897/212338897]

[root@ip-10-16-7-119 ~]# chmod +x cudatoolkit_4.0.17_linux_64_rhel5.5.run
[root@ip-10-16-7-119 ~]# ./cudatoolkit_4.0.17_linux_64_rhel5.5.run


will install the CUDA toolkit. Similarly install the drivers and SDK and finally check everything is working by typing:

[root@ip-10-16-7-119 ~]# nvidia-smi -a -q

GPU 0:0:3
Product Name : Tesla M2050
Display Mode : Disabled
Persistence Mode : Disabled
Driver Model
...
GPU 0:0:4
....


MultiGPU example

Once CUDA 4.0 is installed and working, we can test out the MultiGPU example that comes with the SDK installed earlier. Firstly we will need to install the C++ compiler:

[root@ip-10-16-7-119 simpleMultiGPU]# yum install gcc-c++

and then we need to set our LD_LIBRARY_PATH to include the CUDA libraries:

[root@ip-10-16-7-119 release]# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/lib

After that, we can go to the NVIDIA_GPU_Computing_SDK/C/ folder and type make. The binaries will be installed in the NVIDIA_GPU_Computing_SDK/C/bin/linux/release/ directory and if we go there, we can run the simpleMultiGPU example:

[root@ip-10-16-7-119 release]# ./simpleMultiGPU
[simpleMultiGPU] starting...
CUDA-capable device count: 2
Generating input data...

Press ENTER to exit...


MultiGPU Cluster Setup

Using the above setup and this video, it is also possible to configure an 8 node cluster of GPU instances as described here for high performance computing applications. I will try to do a MultiGPU and Open MPI example in another blog post so stay tuned.

The talk is divided into three parts, initially we introduce the structure of Mathematica, in particular its MathLink API and go into the basics idea of creating a simple C++ application which we can call from Mathematica. Then we discuss the API in a bit more details, especially receiving and sending arrays to and from Mathematica. Its here where we also discus how to receive and send complex numbers, which is handy when doing FFT for example. We then briefly discuss running MathLink applications on remote computers, which is specially useful if you share your CUDA enabled computer with others. Finally we go through some basic error and interruption handling in the MathLink API.

The second part then concentrates on the CUDA aspect of the MathLink application, in some sense the whole philosophy of the talk. If we create a CUDA application that can get and receive data from Mathematica, via the MathLink API, then we are done! In particular we give an overview of a simple example using the mathematica_cuda plugin, which lets you do just this. For a more universal solution, one that works under Windows, there is the excellent CMake module: FindCUDA together with my FindMathLink module which I wrote about previously. We then finish this part by going through a complete example: FFT via CUFFT and show how one goes about getting it working in Mathematica.

The last part, time permitting, is where we show some of the work we have been doing with sending computations to the GPU from Mathematica. In particular I will show some of the work I have been doing with image deconvolution of Confocal and Wide-field images. I am using the GPU to do my deconvolution experiments and using Mathematica to read in the images and analyze the results. Shoaib will present his work on calculating the vegetation index in multi and hyper-spectral satellite images.

I hope you find this overview helpful. We will put the slides up here when the tutorial is over, and if you plan to attend the conference it would be great to see you and get your feedback. Also if there is something specific you would like us to cover, you still have a few days to let us know.

I looked on the web if someone else had already written such a module for MathLink, and in the end found Erik Franken who sent me a version he had modified from a version by Jan Woetzel and others:

By this time I had a version on github which I wrote up. Feel free to download it from here.

Recently Markus van Almsick sent me a more advanced version which I will integrate into my version soon.

Anyways if you are interested, I made a gist for it. Its  in Java and uses JLink to communicate with Mathematica. It was never running for long as I suspect it violated some end user license, but basically one would send a Mathematica command to @mathematica and it would tweet you back your result evaluated by the MathKernel. I am hoping Wolfram might create a similar bot themselves for when you need to know the value of a special function quickly

The basic structure of the project follows that of the Nvidia’s Cuda SDK, in that the individual projects are in their own folder inside the projects folder. Right now I have the scalarProd example from Nvidia. I have also included Nvidia’s cuda utilities cutils and extended the make system to handle Mathematica template files.

Currently I have tested it only on 64-bit Linux, but hopefully I will see if I can get it working under Mac and Windows. I also plan to add more documentation in the project’s wiki on github, and hopefully  get some more useful examples implemented, perhaps FFT.

The initial template of my function is in the scalarProd.tm file:

which describes the ScalarProd[] function in Mathematica, and links it to the scalarProd() C method, which is where we receive the two arrays from Mathematica and use CUDA to calculate their scalar product and send the result back. This and the main() function for Linux and Mac, which is what I was using, are in the scalarProd.cu file. Note that Windows has a slightly different main() method.

and in the same scalarProd.cu we now include the scalarProd_kernel.cu kernel from CUDA’s SDK together with our scalarProd() C function:

Now we are ready to run Mathematica’s mprep pre-processor from MathLink to generate a scalarProdtm.cu file, and on this we run CUDA’s compiler nvcc and compile everything with the appropriate CUDA and MathLink libraries to generate our scalarProd binary, which we can now call from within Mathematica: