Subj : Re: Lock Free -- where to start
To   : comp.programming.threads
From : chris noonan
Date : Fri Oct 14 2005 10:52 am

Dave Butenhof wrote:
> > One context switch takes place every timeslice (say 20
> > milliseconds) whether there are two threads or 1000.
> > This frequency (50 Hz) of context switching reduces
> > throughput slightly but not greatly. How do I know?
> > Because the operating system timeslice is determined
> > as a tradeoff between throughput and latency.
> It's a tradeoff based on arbitrary data that doesn't apply to ANY real
> application except occasionally and purely by accident. In many
> application workloads the impact can indeed be "great". On average, no
> application is average...

I don't understand. If thread-switching is driven by I/0,
then it doesn't matter what the timeslice is. If the
application is CPU-heavy, then the OS designers say
something like "a 20 ms timeslice loses us 3% of the
CPU and that's acceptable". It doesn't matter what
real applications do. Admittedly Microsoft (e.g.) may not
have bothered updating their timeslices as processors
have become faster. Admittedly applications may differ
in their latency requirements. I don't think it's
a matter worth fussing over.

Chris

.