Description of problem:
The default PMtimer is not scalable and performance numbers with Oracle timed
& non-timed stats do not give the customer a good out of the box experience.
With GH TSC will be the default timer and the code is in RHEL5, so the
change/discussion is for the Opterons with the TSCnow patch.
(The modified kernel in reference here is cpuid(0) instruction added to the
RDTSCP code path to simulate the GH latencies).
With modified kernel to simulate GH latencies:
HPET TSC TSC vs HPET
w/ No timed Stats 36251.6 41675.40 1.15
W/ timed Stats 34419.2 40342.80 1.17
%diff 5.3% 3.3%
w/ no timed stats: TSC is 15% faster than HPET
w/ timed stats: TSC is 17% faster than HPET
Oracle 10G HPET PMtimer TSC TSC vs PMtimer TSC vs HPET
w/ notimed stat 37454 39495 43947 1.11 1.17
timed stats 32968 33995 42096 1.24 1.27
% diff 13.6% 16.18% 4.4%
w/ notimed Stats - 11% faster than pmtimer, and 17% faster than HPET
w/ timed Stats on - 24% faster than pmtimer, and 27% faster than HPET
*** This bug has been marked as a duplicate of 218814 ***