New alpha spinlock code passes regression test
Hi,
just wanted to report that 7.0.2 with my patch and a small change to
os.h to tell s_lock to be a long instead of an msemaphore.
Could someone tell us how to get more useful numbers out of the
benchmarks? On a DS20 the benchmark is basically meaningless...
Thanks,
Arrigo
Cool --- did you try the parallel regress tests, or just sequential?
My experience is that it takes quite a few iterations of the parallel
tests before you should have much confidence that there aren't locking
bugs lurking. But if it comes through that, send in the patch and
we'll gratefully accept it!
regards, tom lane
PS: also please note my just-posted call for testers of the fmgr
changes...
Tom Lane wrote:
Cool --- did you try the parallel regress tests, or just sequential?
My experience is that it takes quite a few iterations of the parallel
tests before you should have much confidence that there aren't locking
bugs lurking. But if it comes through that, send in the patch and
we'll gratefully accept it!
I've run the regression tests several times (also the parallel ones and
bigtest). It failed geometry, but it always does due to different
rounding. Passed everything else, so it looks ok.
Now I'm trying to figure out how I can do a sensible timing test on the
two versions to see whether it is any faster. The postgres benchmark
finishes in such a ridiculously short time, that that isn't telling me
anything.
Adriaan