Re: a faster way to gettimeofday? rdtsc strangeness

[Jamie Lokier <lk@tantalophile.demon.co.uk>]

H. Peter Anvin wrote:
> > Can /proc/cpuinfo really be trusted in figuring out how long a cycle is?
> 
> It uses RDTSC, so yes.

Yes and no.  The kernel's measurement isn't as accurate as it could be.

I have some code which calibrates TSC cycles against gettimeofday().

Considering that gettimeofday() works by using the kernel's frequency
estimate that it shows in /proc/cpuinfo, the results from the code
should match /proc/cpuinfo very precisely ndeed.

However, over a million gettimeofday() calls, the frequency of the 100Hz
tick would tend to dominate the result, showing any error in the
kernel's estimate which appears in /proc/cpuinfo (which is derived from
measuring the length of a 100Hz tick in TSC cycles).

With 10^6 gettimeofday() calls, and linear regression through the set of
measurements when an interrupt didn't happen (checked by reading TSC
before and after the system call and filtering on the minimum), I get
superb results on a FreeBSD box:

	Kernel reports clock frequency:  746.154 MHz
	Measured clock frequency:        746.154 MHz

On various Linux boxes, however:

	Kernel reports clock frequency:  596.924 MHz
	Measured clock frequency:        596.929 MHz

	Kernel reports clock frequency:  866.708 MHz
	Measured clock frequency:        866.675 MHz

	Kernel reports clock frequency:  664.597 MHz
	Measured clock frequency:        664.582 MHz

	Kernel reports clock frequency:  366.601 MHz
	Measured clock frequency:        366.584 MHz

As you can see, the kernel consistently overestimates the CPU clock
frequency, or more precisely it overestimates the number of cycles in a
"100Hz" system clock tick.  This is something that could be improved in
the kernel code, as FreeBSD demonstrates.  Note that the clock frequency
probably varies by a similar order of magnitude to this error due to
thermal and power variation anyway.  But it would be nice if the
measurements matched as they do in FreeBSD!

I have one of those laptops that Alan mentioned which does change the
CPU clock rate.  It's a Toshiba 4070CDT, here running in fast then slow
mode:

	Kernel reports clock frequency:  366.601 MHz
	Measured clock frequency:        366.584 MHz

	Kernel reports clock frequency:  366.601 MHz
	Measured clock frequency:        184.503 MHz

It can switch between fast and slow modes while a program is running,
which occurs either when you press a special key combination, or the
mains power cuts off.

enjoy,
-- Jamie