Re: [patch, v2.6.22-rc6] sys_time() speedup

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 26 Jun 2007 17:26:29 +0200 Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > > > the patch improves the sysbench OLTP macrobenchmark significantly:
> > > 
> > > Has that any real practical relevance?
> > 
> > Interesting question. [...]
> 
> i'm missing the <sarcastic> tag i guess ;-)
> 
> <sarcastic> Oh my, does database macro-performance have any relevance to 
> Linux bread and butter markets in general. Boggle, it is a really 
> difficult question i suspect. </sarcastic>
> 
> If we ignore those few million database and web server Linux boxes on
> the market and concentrate purely on the few m68k boxes that are still
> in existance, _then_ we might be doubtful about this question ;-)

On my machine, time(2) doesn't do any syscall at all - it uses the vsyscall
page.  I'd be surprised if a database uses sys_time() either.

> > [...] The patch adds a new test-n-branch to gettimeofday() so if 
> > gettimeofday() is used much more frequently than time(), we lose.
> 
> given that the cost to sys_gettimeofday() is less than a cycle (we test
> a value already in a register, with an unlikely hint), and the benefit
> to sys_time() is around 6000 cycles (or more), sys_gettimeofday() would
> have to be used thousands of times more frequently than sys_time() -
> which it clearly isnt. As a test i just triggered a really X-intense
> workload and for that gettimeofday-dominated landscape there was still 1
> sys_time() call for every 50 gettimeofday calls - so it's a small win
> even for this X workload.

So something in X is somehow calling sys_time()?  How come, and is that an
outlier?  How generalisable is this observation?