Re: TRON for AdaOS

[Vinzent 'Gadget' Hoefler <nntp-2005-02@t-domaingrabbing.de>]

Ole-Hjalmar Kristensen wrote:

> My point is that the task switching overhead itself is very low on
> Linux.

Yes. But such things can change when you apply real-time requirements.
Sometimes it is very surprising to see the difference between an
average-case optimization and a worst-case optimization. ;-)

Well, I searched for it and the context switching times simply seem to
depend on the method used.

<URI:http://trecom.upv.es/articles/NTCC150-4.PDF> states:

|The communication method between the scheduler and the Linux process
|can be done by means of the RT-Linux standard RT-fifos, although a
|shared memory mechanism has also been implemented in order to improve
|efficiency. This method allows the scheduler to improve context
|switches by three milliseconds below the accomplished with the FIFO
|communication.

So indeed here they talking about milliseconds, but with the shared
memory approach this seems to change considerably:

|Some measurements about context switching times in RT-Linux have been
|made in the execution of the previous program. [...] In the case of
|interrupts, this time is very variable, but it can be up to 20
|microseconds in the worst case. [...] In the case of system calls, the
|context switching time is much lower: 8 µs. The scheduler requires
|about 4 microseconds.

So that doesn't sound too bad. ;)


Vinzent.

-- 
worst case: The wrong assumption there actually is one.