[j-nsp] Qfabric

[Joel Jaeggli]

On 2/25/11 5:42 AM, Saku Ytti wrote:
> On (2011-02-24 17:15 -0800), Joel Jaeggli wrote:
>  
>> that activity can be simple as front-running large orders (which take
>> longer to fill) with small ones, an elaborate algorithm is not
>> necessarily a requirement. I'm kind of down on the market utility of
>> such activity but it's not presently illegal.
> 
> I suppose it still may be somewhat more complex task than my trading
> software, in terms of how many instructions it needs from CPU, and thus
> necessarily higher latency, as it is competing with other stuff, for
> example kernel house-keeping.

So, I haven't worked for a company that had those sorts of customers for
a while (like two years). but generally it's a probability game and
strictly marginal improvements in even highly jittered execution times
are likely to have demonstrable impact on their success rate.

network infrastructure is a very small portion of the capital outlay for
one of these operations so refreshing equipment to take advantage of
small opportunities is easy to justify...

That said, once you've gone cut-through, shed all the stateful devices
in the path and reduced the rtt from 40km to 150meters you are correct
that focus on other portions of the system is going to dominate.

> Are these software being ran in regular PC, or has someone tapped this
> market and is producing FPGA based fixed delay trading solutions at
> nanoseconds jitter scale?
> I wonder how much jitter in normal PC could be reduced for running
> simplistic software, even with esoteric stabbing of the kernel, sub 1ms
> jitter on software solution seems unachievable.
>