Benchmark - 2nd run

Matthew Dillon dillon at apollo.backplane.com
Fri Jan 16 14:22:32 PST 2004


:Would anyone attempt to explain the major factors in the measured
:decrease in performance from 4.9 to 5.2 and why these numbers
:don't always seem to match the user experience?

    Act of god?  Who knows.  Certainly there is something funny
    going on between 5.1 and 5.2, probably something stupid simple
    and easy to fix once found.  The benchmark numbers are better then
    what Jean got in his first set of tests but there is still something
    screwy going on.  e.g. the sqrt() test should have returned the same 
    numbers between 4.9 and DFly.  All I can think of is that these
    tests are hitting edge cases with the L1/L2 cache.

    I'm quite happy with the 4.9 vs DFly numbers, despite the weird
    results.  It shows that we haven't made any horrendous implementation
    mistakes. 

    Also, remember that Jean ran these tests on a fairly old machine.
    5.x depends heavily on cmpxchg and needs a larger L1/L2 cache footprint
    to operate efficiently, so it ought to fair much better on newer machines.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>





More information about the Kernel mailing list