DragonFly kernel List (threaded) for 2004-01
Re: Benchmark - 2nd run
:Finally to test if there is a problem with the compiler I used the lorenz
:benchmark (it simply does raw math computations, and does not use much
:Lorenz on 5.1
:4.9binary: 15.77 real 15.73 user 0.01 sys
:5.1binary: 16.52 real 16.51 user 0.00 sys
:5.2binary: 16.36 real 16.32 user 0.02 sys
:Lorenz on 5.2
:4.9binary: 19.35 real 18.12 user 1.14 sys
:5.1binary: 27.65 real 26.05 user 1.48 sys
:5.2binary: 29.61 real 27.60 user 1.87 sys
:The 5.1 results indicate that there is no problem with the compiler
:generated code, but I don't know how to interpret the 5.2 results: I
:would expect also a bad behavior with the 4.9 binary (all are static)
:Then I booted a uniprocessor kernel and reran the latest test:
:Lorenz on 5.2 UP
:4.9binary: 15.80 real 15.71 user 0.02 sys
:5.1binary: 22.57 real 22.45 user 0.03 sys
:5.2binary: 24.16 real 24.05 user 0.03 sys
I think your Lorenz results are the most interesting of the
lot. It definitely shows that there's something lala going
on in 5.2-SMP-land. The system overhead should be near 0
and it's way up there at almost 2 seconds!
Check 'hz' on 5.2. 'sysctl kern.clockrate'. It kinda looks
like its going wild on 5.2. If it is not set to 100 you can
force it with a /boot/loader.conf setting 'kern.hz=100'.
If it isn't that it's got to be something else simple-stupid.
The unusual system times for a user-cpu-bound program would
explain the horrible results on the other benchmarks.