These days ASICs are finding their way into HFT. I know of at least a few where the trading algo runs along the TCP stack in custom silicon. Hard to compete with that with Linux or Windows.
As a practical matter, prior to the acquisition, I would have chosen QNX as the best real-time kernel available for little money. These days, a custom Linux kernel is the best an average person can do (imo).
you can do sub millis within thread, but if you do any context switching, your stuck with os granularity, which as the poster you responded to said, is around 10-15ms depending on which flavor you're running. windows isn't an rtos, sorry.
from what i can tell most posters don't work in the industry and have no idea what they are saying. fact of the matter is that both windows and linux are used. prior to 2007/08 i think windows was the main OS used for some reason.
C# is not genrally slower than C++. C++ just offers higher potential for optimisation. So, specific algorithms like hashmaps with string keys can be implemented substantially faster in C++. Copying long C++ strings using Intel MKL is an order of magnitude faster than C#. The thing is in many trading algorithms these specific scenarios will contribute single percentage improvement, which is immaterial. As Big mentioned, garbage collection effectively freezes a NET application till it's finished. This freeze may last hundreds of miliseconds or even seconds in some cases. So, one has to use a number of workarounds to minimise the effect of garbage collection. Another issue is if you have to use a specific "native" library, calls from "native" code to NET and vice versa carry a performance penalty. Even with safety checks disabled using a simple "native" operation in NET code may become the single most important bottle neck.
Can anyone name a market data source that works with linux? To me, this would be the biggest problem as no data provider seems to cater to this market.
Interactive Brokers offer data via Java API which works on most platforms including Linux. As a data source it's not perfect but it does the job.
If you were going to build a High Frequency Trading system on a custom linux solution, I think you would need a data source that was a lot more advanced. You'd need full time & sales as well as level 2 depth. This is my point. The options appear to be really limited.