KGPU - Augmenting Linux with the CUDA GPU

by deutronium • 14 years ago

77 32 14 years ago

Where is this useful? The bus speed across to the GPU is so slow, I thought it was only meaningful for near-autonomous operations.

jws2• 14 years ago

README suggests RAID processing, file system encryption, and AES.

GIFtheory• 14 years ago

Agreed. I'm also struggling to see what kind of massively parallel operations need to be done in kernel space in the first place.

jws2• 14 years ago

Maybe it doesn't have to saturate the GPU to be a win. If you can just banish some cache busting, streaming work, like raid processing, to a tiny sliver of the GPU it could be a win.

Tuna-Fish2• 14 years ago

Right now, the bandwidth to modern GPUs is actually pretty decent (16GB/s bidirectional), but the latency is still horrid. This means that you need rather large operations for offloading to pay off. I think doing raid-5 or full disk encryption with large blocks might just barely be worth it.

However, with AMD and Intel integrated GPUs, this is about to change. AMD is doing a lot of work on HSA, which can be summarized as "GPU and CPU share same memory, and can communicate by passing pointers". I can see this kind of work being really useful in the near future.

patrickgzill• 14 years ago

I had thought in the past that storing the index of a database (not the data, just index) on the card and using that to handle complex queries, might be interesting. Not sure if that has a practical, real-world use though.

dfc5• 14 years ago

More and more CPUs have AES instructions and my old lenovo ideapad has a crypto coprocessor. Do you think the GPU offload will be worth it when the sytem has hw accelerated crypto?

caf• 14 years ago

Not so much RAID5, which is just an XOR operation that is as good as free on a modern CPU, but RAID6 where a more computation-intensive Reed-Solomon code is used.

dkhenry• 14 years ago

One situation where you might see immediate results is in High volume routing. I think this project [1]. They were using the GPU to saturate multiple 10GbE interfaces with a commodity processor.

1.http://shader.kaist.edu/packetshader/

rincebrain• 14 years ago

I actually spoke to them about doing a similar thing, and they said they were using a GPU because they were pushing lots of sub-1500 MTU packets, and that commodity h/w could probably saturate multiple 10GbE interfaces that were just doing large packets for high-throughput.

odranoelson• 14 years ago

I wonder whether this won't conflict with other user applications using the GPU.

mtgx• 14 years ago

Would this be possible with OpenCL, too?

gnu8• 14 years ago

Why is this cuda and not OpenCL? There's no reason to legitimize nvidia's proprietary nonsense, it just enables their bad behavior.

iso-8859-1• 14 years ago

CUDA is older and more popular: http://www.google.com/trends/explore#cat=0-5&q=opencl%2C...

snogglethorpe• 14 years ago

Nonetheless, it's a bad idea to favor proprietary languages that lock you into a particular company's products.

CUDA should have died when OpenCL came about.

Of course since such lockin is to Nvidia's benefit, it's understandable why they keep promoting their proprietary solution...

fool2• 14 years ago

The GPU code seems to be relatively isolated to memory operations in gpuops.cu. Not altogether sure but a quick review suggests such mapping is supported by OpenCL. So one could rewrite that module and the whole thing would work without CUDA. Of course, in terms of compilers it is going to be a while before users can move away from nvcc for nvidia graphics card support.

However, as with the parent I'd really like to see a generic OpenCL vectorization kernel module. Strong suspicion that this work was directly or indirectly underwritten by Nvidia so I guess someone (Intel?) needs to step up and fund similar academic projects.

fool2• 14 years ago

Suspicion confirmed:

"KGPU is a project of the Flux Research Group at the University of Utah. It is supported by NVIDIA through a graduate fellowship awarded to Weibin Sun."

http://code.google.com/p/kgpu/

hyperbovine• 14 years ago

What sort of vitriol is this? How dare they invent a new technology and build an API to it.

DannyBee2• 14 years ago

Didn't you know all good API's are designed from scratch in open committees, instead of being standardized after multiple competing implementations?

wavesum• 14 years ago

I don't understand :)

Could someone explain what this kind of technology means in practice? Does this mean I can GPU accelerate my old code?

greggman• 14 years ago

You do NOT want to do this. Unlike the CPU which can be preempted current GPUs can not. If you give them 2-30 minutes of work to do they will not return until that work is done. Windows gets around this by resetting the GPU if it doesn't respond for more than a few seconds. Linux/osx no such luck, at least not yet.

But, once reset the state of the GPU is often unknown. Not a good thing if you are embedding GPU code into your kernel.

KGPU - Augmenting Linux with the CUDA GPU

Related Stories