A different performance measure

Hi All,

After some discussions earlier this week about other ways context switch
times can be measured I thought I ought to go and see how fast VSTa really
is. I modified the perf1 code to use sched_yield() instead of the
messaging engine (sched_yield() should be available in the next full release
of VSTa). The code also runs in the "real-time" scheduler queue, which
whilst it gets a higher priority than the normal timeshare mode seems slow
for some reason (so I'll look into this as well now).

Running with my DX/4-100 test system and a kernel built under gcc 1.42, this
is now showing 15 us context switches :-) (measured over 1M switch ops).
The same system with a DX/2-66 (not the same system I usually mention for a
DX/2) shows 21 us. FWIW with this DX/2 system I get a perf1 figure of 75
us.

As I was after the best figures though I recompiled under gcc 2.5.7 and
scored 19.7 us on the DX/2 and 14.7 us on the DX/4 for the sched_yield()
switch times.

As I think has been mentioned there are lies, damn lies and benchmarks :-)

                        Regards,
                        Dave
Received on Fri Jan 20 06:35:39 1995