Re: Benchmark - 2nd run

[Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>]

:I have found the origin of the problems with 5.2. The new ACPI code
:generates more than 36000 interrupts per second on my hardware. (Asus
:P2B-DS). Disabling ACPI improves the 5.2 perfs. Curiously there is not
:a big  improvement for the lorenz test (now: 24.82 real / 24.80 user /
:0.00 sys) but the bytebench benchmark is much closer to the 5.1
:benchmark: 

    These numbers look a lot better.  In particular, the arithmatic
    tests do a better job reflecting the improvements in 
    GCC 3.x (FreeBSD-5) verses GCC 2.95.x (FreeBSD-4.x and DFly).

:Arithmetic Test (type = arithoh)        1055234.6 lps 
:Arithmetic Test (type = register)         91672.9 lps 
:Arithmetic Test (type = short)            88086.6 lps 
:Arithmetic Test (type = int)              91662.9 lps 
:Arithmetic Test (type = long)             91682.2 lps 
:Arithmetic Test (type = float)           105154.2 lps 
:Arithmetic Test (type = double)          105242.2 lps 

    The syscall overhead tests look reasonable for the added 
    overhead in FreeBSD-5 syscalls.  The pipe-throughput
    and context switching test reflects the great deal of
    work that has been done in the FreeBSD-5 pipe code
    to offset other overheads, though why they can't beat
    4.x now I just don't know.
 
:System Call Overhead Test                 82300.9 lps 
:Pipe Throughput Test                      78012.8 lps 
:Pipe-based Context Switching Test         23268.0 lps 

    The I/O read tests basically reflect bandwidth through
    (or mostly through) the buffer cache.  Write/Copy
    tests are more realistic and show a serious falloff
    relative to 4.x, and 5.x has a lot of work ahead of 
    it to improve those numbers.

:File Read  (10 seconds)                  181930.0 KBps
:File Write (10 seconds)                   12033.0 KBps
:File Copy  (10 seconds)                   11905.0 KBps
:File Read  (30 seconds)                  182251.0 KBps
:File Write (30 seconds)                   11933.0 KBps
:File Copy  (30 seconds)                   11811.0 KBps

    The C compiler test makes sense (gcc 3.x takes longer
    to compile things then gcc 2.95.x).  Shell script
    exec overheads look correct now too considering the
    added exec overhead and the dynamic /bin.  That gap 
    is only going to continue to widen.

:C Compiler Test                             241.2 lpm 
:Shell scripts (1 concurrent)                572.2 lpm 
:Shell scripts (2 concurrent)                324.9 lpm 
:Shell scripts (4 concurrent)                163.9 lpm 
:Shell scripts (8 concurrent)                 82.0 lpm 

    I don't know what is going on with sqrt().  The
    numbers are the same for 5.1 but 4.x gets 84K and
    DFly only gets 25.7K.  I don't understand what's going
    on there.

    The recursion test reflects the better compiler.  We'll
    soon have gcc-3 in our tree too (David Rhodus is working
    on it) :-).
	
					-Matt
					Matthew Dillon 
					<dillon@xxxxxxxxxxxxx>

:Dc: sqrt(2) to 99 decimal places          54949.6 lpm 
:Recursion Test--Tower of Hanoi             8382.6 lps 
:
:
:Jean-Marc
:
:-- 
:Jean-Marc Zucconi -- PGP Key: finger jmz@xxxxxxxxxxx [KeyID: 400B38E9]
: