Benchmark: NetBSD 2.0 [allegedly] beats FreeBSD 5.3 in server performance

[Robert Watson]

On Thu, 6 Jan 2005, stheg olloydson wrote:

> it was said by Phil Brennan:
> 
> >What about the context switch time? Are there any plans to improve
> >this, and also to reduce the number of context switches needed?
> 
> See Robert Watson's reply to this thread. An unfortunate number of
> problems exist in threading and scheduling. Most are well-understood and
> are being worked on and 5.4 should see measurable improvement. 
> Personally, I am more concerned with network and scheduler perfomance. 
> I know the former is being addressed, but I don't hear anything about
> how work on SCHED_ULE is progressing. 

FWIW, one of the reasons that there hasn't been as much interest in
SCHED_ULE lately is likely that several of the features previously only
present in SCHED_ULE are now also present in SCHED_4BSD -- for example,
making more effective uses of IPIs in reducing latency during
inter-process communication across processors.  While SCHED_ULE does
contain a number of interesting things not present in SCHED_4BSD, the 4BSD
scheduler has hardly gone un-improved in that time.  However, Jeff
Robserson does seem to have picked up recently on both VFS SMP locking and
ULE.  The scheduler tracing and visualization tools he committed a couple
of weeks ago are really quite neat tools.

Robert N M Watson