Just was getting some thoughts out about using GPGPUs (nVidia’s Tesla and Fermi products, ATI’s FireStream) for handling some of the menial cloud-based operations (metadata calc, metadata location servicing, offload, etc.) that we have as part of any type of cloud file system.
So, what are your thoughts? The capacity to further collapse x86 commodity platforms by utilizing the extreme amount of processing capability present within these GPGPUs and coupling optimized OpenCL or CUDA routines into the standard CFS processing stack.
Just something to think about.
Related articles by Zemanta
- AMD revs up Stream SDK (theregister.co.uk)
- NVIDIA announces Fermi products for 2010 ship (insidehpc.com)
- Micro-burst: Master Node Topology (flickerdown.com)
![Reblog this post [with Zemanta]](http://img.zemanta.com/reblog_e.png?x-id=a636da71-a04b-4296-8dad-8e97cffb9a7e)
[...] This post was mentioned on Twitter by davegraham, MAG. MAG said: RT @davegraham: Blog Post:: Video: Cloud and GPUs http://bit.ly/4y2jR1 [...]
The HPC world accesses GPUs on a large scale today on clusters using jobs schedulers like PBS/Torque/Grid Engine. One current barrier to using GPUs with current-generation clouds is that using GPUs from a guest OS on a type-1 hypervisor (e.g., Xen) is still in the research stage. OS-level virtualization may be more feasible today, as there is only one kernel to interact with the GPU, but OS-level virtualization isn't commonly used for clouds.
Today
The challenge with this type of configuration is getting the data in and out of the massive number of GPU's in a timely fashion.
They work well where you have to perform some repetitive task many times on a piece of data so examples such as cracking WPA keys where you want to use a brute force against the keys works well.
Using them to transcode video can be an advantage if the transcoding application can keep the GPU's working.
Trying to use them as a general purpose CPU does not work as well because of the limited memory and IO capability on the GPU. We are still currently limited by the PCIe bus. With 240 processors on the end of relatively limited IO bus this starts to be a problem.