Rob Clark wrote:
>>> Rob Clark wrote:
>>> It would be great to use the GPU for ZFS or to offload some of the OS load
>>> ...
>>> http://developer.nvidia.com/object/nexus.html
>>>
>>> If SunStudio had some GPU support and ...
>>>
>>>
>
>
>> asura wrote:
>> Forgive if I'm mistaken, but I don't think the CPU is the bottleneck
>> when it comes to ZFS.
>>
>
> It is on my system, with many drives and higher level correction the
> performance stinks. Using a '300 would be overkill if I only used it
> for a single task; but one thing at a time, I'm not asking for a full
> port overnight, else I'd ask that OpenSolaris ran on my cell phone.
>
>
>
>> asura wrote:
>> Secondly, and more importantly for SS to generate
>> PTX would probably be quite an effort not to mention
>>
> So lets get started ...
>
>
>
>> bfriesen wrote:
>> GPUs are great for some things but it is difficult to imagine a GPU
>> being of assistance in the zfs implementation due to way too much
>> latency,
>>
>
> http://www.nvidia.com/content/PDF/fermi_white_papers/NVIDIA_Fermi_Compute_Architecture_Whitepaper.pdf
>
> They claim to have addressed most of the 2nd generation complaints.
>
> These (cards) are no longer "Video Graphics Adapters" (where the ancient
> term "VGA" comes from) that have been hacked to use the onboard GPU
> to solve simple equations. The newest cards can chose between 'shared
> memory' or 'cache memory' configurations on the fly. With many 10's of
> thousands of threads and context switching under 25 microseconds these
> cards are no slouch.
>
> These cards are expected to be under a $1000 in the next few years.
>
> If you write lousy code that takes poor advantage of the card you may
> get a speed up of 2 or 3 times. That seems a fair return on your money,
> time, and effort.
>
> If you write wonderful code that is especially designed to take advantage
> of the way the card operates _some_ (FEW) programs were able to run
> up to 1000 times faster on the last (2nd) generation; the new (300 series)
> cards are some 2 to over 10 times faster (depending on what you are doing).
>
>
> If the entire Operating System ran on the Graphics Card then you could
> run your programs on your computer with a lot lower latency than you
> are worried about ...
>
> While we _do_ want the Operating System to run / work / operate well,
> that "purpose" of the Computer is not to run the Operating System but
> to run the "Application". It is only us "OS Geeks" who get more enjoyment
> out of a 'New OS' than someone else might get out of high frame rates
> in Maya. So that is one solution ...
>
>
Hi Rob,
A few unspoke things..
1) GPGPU are good for certain kinds of code and probably very poorly
suited for others. There is research showing some nonintuitive
performance gains, but none of that is in any production compiler afaik
2) CUDA + OpenSolaris is not publicly available. If you work for a
company that's interested contact me off list
3) On a conservative estimate the budget to make this happen would be in
the hundreds of thousands range
a. We would have to get PathScale to build the OpenSolaris kernel
(currently already in progress)
b. We must decide on our front-end. Right now we're in-between a
CUDA front-end provided by NVIDIA and switching over to pragmas like HMPP
c. Finish rewriting the CUDA runtime
d. Invent a way for kernel space applications to efficiently and
safely offload to the accelerator
e. Rewrite portions of the code (Possibly zfs to take advantage of (b)
I could imagine if you moved some of this to useland like fuse does the
problem may become a bit less complex, but that's a tradeoff. In the
end you may find that having your compiler highly optimize the
compression code is good enough. I'd have to know which cpu you're
targeting, the compression algorithm and a few other things to
conclusively tell you how much better we could do.
(reply-to is set to my business email)
Thanks,
./Christopher
