Re: [PATCH 1/6] new timeofday core subsystem for -mm (v.B3)

[john stultz <johnstul@us.ibm.com>]

On Tue, 2005-06-21 at 17:57 -0700, john stultz wrote:
> > I don't need any practical numbers, I can already see from the code, that 
> > it's much worse and unless you eliminate the 64bit calculations from the 
> > fast path, I don't know what you are trying to optimize.
> 
> That's exactly what I'm trying to optimize. By precalculating some of
> the 64 bit manipulations, we can remove them from the fast path.
> 
> Ok, from my initial tests on my i686 laptop (@600Mhz), using the
> cheapest timesource available (the TSC), the unoptimized B3 version of
> the code I sent out shows a 17% performance hit in gettimeofday(). That
> ratio will be even smaller if you use a more expensive timesource. So
> starting there, let me see how much I can shave off.

Just a quick update: With a bit of quick optimizing the current design,
removing most of the 64bit math from the gettimeofday fast path, I've
managed to cut it down to within 2% of the current code using the c3tsc
vs tsc. The c3tsc timesource adds a bit of overhead that the mainline
code doesn't do, so I'm going to see how a more comparable TSC vs TSC
test goes.

thanks
-john