[ntp:hackers] nVidia CUDA GPU - wow - NTP never had it this good before.
tglassey at earthlink.net
Mon Jan 25 18:31:43 UTC 2010
On 1/24/2010 11:24 PM, Terje Mathisen wrote:
> tglassey wrote:
>> On 1/23/2010 11:05 PM, Hal Murray wrote:
>>>> So folks - I have a functional CUDA based NTP server and it's
>>>> performance is incredible.
>> Sorry forgot to add this for the TELOS board link - check this baby out
>> 1.3G 240Core board as a compute co-processor. And less than $1200 too.
>> Drop three or four of these into a DB Server and you got rock and
> Todd, I'm absolutely flabbergasted!
Me too - read my offlist response :-)
> It is really amazing both that you had the idea to do this, and
> actually made it work. :-)
Making it work wasn't hard - the real issue was how to separate the
thread controls in NTP. My model uses four layers of threads each
protecting the inner thread with the actual time-keeping thread running
in the most internal thread.
> NTP is such a low overhead protocol that a high-performance server
> could instead run the entire server protocol stack inside the hw
> interrupt driver, at which point the cost (for the Nvidia setup) of
> copying packets to/from the GPU seems like way more overhead. :-(
Actually no - it allows for real power in Autokey and additional crypto
services which can now be added which would kill NTP in normal use
models because of the thread-blocking for high overhead uses.
> No virus found in this incoming message.
> Checked by AVG - www.avg.com
> Version: 8.5.432 / Virus Database: 271.1.1/2643 - Release Date: 01/24/10 19:33:00
More information about the hackers