https://github.com/wbsun/kgpu

Currently GPUstore. Technical paper: http://www.cs.utah.edu/~wbsun/gpustore.pdf
https://github.com/wbsun/kgpu

Last synced: about 19 hours ago
JSON representation

Currently GPUstore. Technical paper: http://www.cs.utah.edu/~wbsun/gpustore.pdf

• Host: GitHub
• URL: https://github.com/wbsun/kgpu
• Owner: wbsun
• License: other
• Created: 2011-03-28T19:24:09.000Z (about 13 years ago)
• Default Branch: master
• Last Pushed: 2013-12-02T23:53:16.000Z (over 10 years ago)
• Last Synced: 2024-03-05T21:33:30.287Z (4 months ago)
• Language: C
• Homepage: https://code.google.com/p/kgpu/
• Size: 1.22 MB
• Stars: 214
• Watchers: 33
• Forks: 34
• Open Issues: 5
• Metadata Files:
  • Readme: README

Lists

• awesome-stars - kgpu

README

        

     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:

     [email protected]: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