KGPU - Augmenting Linux with GPUs
** Important note:
A new branch 'k32' is created for KGPU compilation on 3.x kernels,
I tested 3.2.16. If you want to try KGPU on recent kernels, definitely
checkout that branch.
I don't have time to modify everything to comply with the latest kernel,
so k32 branch has gaes and raid6 services disabled. Just leaves an
example service for hubbyists to borrow code to start with their own
service development.
What is it?
Treating the GPU as a computing co-processor. To enable the
data-parallel computation inside the Linux kernel. Using SIMD (or
SIMT in CUDA) style code to accelerate Linux kernel
functionality.
Make the Linux kernel really parallelized: which is not only
processing multiple requests concurrently, but can also partition
a single large requested computation into tiles and do them on
GPU cores.
GPU can give the OS kernel dedicated cores that can be fully
occupied by the kernel. But the multicore CPUs should not be
occupied by the kernel because other tasks also need them.
KGPU is not an OS running on GPU, which is almost impossible
because of the limited functionality of current GPU
architectures. KGPU tries to enable vector computing for the
kernel.
*To access the code, using git to clone:
git@github.com:wbsun/kgpu.git or goto
https://github.com/wbsun/kgpu .*
As for copyright license, we use GPLv2.
News
* RAID6 PQ computing function added as a service, gpq module
for its kernel part to replace the global raid6_call
algorithm with GPU one, it can beat the fastest SSE version
with 16 disks and >= 1MB data on my machine. Try it with a
RAID6 on dm driver.
* Scripts to run and stop kgpu.
* Simple build system.
* dm-crypt can use gaes_ecb or gaes_ctr directly.
Try it?
Hardware:
We use GTX480. You don't need such high-end video
card, but you should have a NVIDIA card that support CUDA
computing capability 2.0 or higher. If you don't have more
than 1G video memory, change KGPU_BUF_SIZE in kgpu/kgpu.h to
make sure KGPU_BUF_SIZE*2 < Size of Your Video
Memory - (x) where the max of x is a value that you need try
some times to figure out. Or simply leave x = 64M or 128M.
Notice a new change: we enabled a new feature to allow
KGPU remapping any kernel pages into CUDA page-locked
memory, the remapping also need video memory on the GPU
side, so now there are two GPU buffers with the same size,
which is KGPU_BUF_SIZE. So KGPU_BUF_SIZE should be <=
video memory size/2.
Software:
We compile the CUDA code with nvcc in CUDA 4.0. The OS
kernel is vanilla Linux 2.6.39.4. You MUST use a 64bit linux
kernel compiled targeting at x86_64!
Make and Run it:
Check out the code from Github or download the
archive from Google Code and extract files into say kgpu
directory:
cd kgpu && make all
Now all outputs are in build directory. To run it:
cd build && sudo ./runkgpu
This only starts KGPU module, helper and loads AES ciphers.
To use modified eCryptfs and dm-crypt, in the build directory:
sudo insmod ./ecryptfs.ko && sudo insmod ./dm-crypt
NOTE: DO NOT USE THIS ECRYPTS FOR IMPORTANT DATA!!!
THIS IS NOT COMPATIBLE WITH THE VANILLA ECRYPTFS.
SAME CARE SHOULD BE TAKEN WITH DM-CRYPT.
To stop it:
Umount your eCryptfs partition, delete dm-crypt mappers and:
sudo rmmod ecryptfs && sudo rmmod dm-crypt
Stop "helper" program by Ctrl-C
sudo ./stopkgpu (in build/)
Weibin Sun, Xing Lin
{wbsun, xinglin}@cs.utah.edu