How to measure time accurately.

How to measure time accurately.

[krishna]

Hi All,

Can any one tell me how to measure time accurately for a block of C code 
in device drivers.
For example, If I want to measure the time duration of firmware download.

Regards,
Krishna Chaitanya

Re: How to measure time accurately.

[Lee Revell]

On Mon, 2005-03-28 at 08:58 +0530, krishna wrote:
> Hi All,
> 
> Can any one tell me how to measure time accurately for a block of C code 
> in device drivers.
> For example, If I want to measure the time duration of firmware download.

rdtsc()

Re: How to measure time accurately.

[Chris Friesen]

krishna wrote:
> Hi All,
> 
> Can any one tell me how to measure time accurately for a block of C code 
> in device drivers.
> For example, If I want to measure the time duration of firmware download.

Most cpus have some way of getting at a counter or decrementer of 
various frequencies.  Usually it requires low-level hardware knowledge 
and often it needs assembly code.


On ppc you'd use the mftbu/mftbl instructions, as suggested by Lee on 
x86 you'd use the rdtsc instruction.

Chris

Re: How to measure time accurately.

[Jesper Juhl]

On Mon, 28 Mar 2005, Chris Friesen wrote:

> Date: Mon, 28 Mar 2005 23:07:14 -0600
> From: Chris Friesen <cfriesen@nortel.com>
> To: krishna <krishna.c@globaledgesoft.com>
> Cc: Linux Kernel <linux-kernel@vger.kernel.org>
> Subject: Re: How to measure time accurately.
> 
> krishna wrote:
> > Hi All,
> > 
> > Can any one tell me how to measure time accurately for a block of C code in
> > device drivers.
> > For example, If I want to measure the time duration of firmware download.
> 
> Most cpus have some way of getting at a counter or decrementer of various
> frequencies.  Usually it requires low-level hardware knowledge and often it
> needs assembly code.
> 
> 
> On ppc you'd use the mftbu/mftbl instructions, as suggested by Lee on x86
> you'd use the rdtsc instruction.
> 

In some cases you can simply count jiffies - depending on how accurate you 
need to time things I'd say that often something like this is adequate :


unsigned long start, time_spent;

start = jiffies;
/* do stuff */
time_spent = jiffies - start;
printk("stuff took %d jiffies (%d seconds)\n", time_spent, time_spent/HZ);


-- 
Jesper Juhl

Re: How to measure time accurately.

[Jan Engelhardt]

>In some cases you can simply count jiffies - depending on how accurate you 
>need to time things I'd say that often something like this is adequate :

These "some cases" exclude this one:
If interrupts are disabled, a jiffy might be missed. Take care.

If you are on UP and want to measure within
- a tick [when using preempt]
- kernel space [no preempt]
disabled interrupts should usually not be the case.


Jan Engelhardt
-- 
No TOFU for me, please.

Re: How to measure time accurately.

[Peter Chubb]

>>>>> "Chris" == Chris Friesen <cfriesen@nortel.com> writes:

Chris> krishna wrote:
>> Hi All,
>> 
>> Can any one tell me how to measure time accurately for a block of C
>> code in device drivers.  For example, If I want to measure the time
>> duration of firmware download.

Chris> Most cpus have some way of getting at a counter or decrementer
Chris> of various frequencies.  Usually it requires low-level hardware
Chris> knowledge and often it needs assembly code.

As a device driver is inside the linux kernel (unless you're writein a
user-mode device driver :-)) you can use the getcycles() macro that's
defined for most architectures.  It provides a snapshot of the
cycle-counter.

Caveats:
	1.  If you're running with power management, the  cycle
	    counter ticks at a  variable rate.
	2.  If you're on a multiprocessor, the cycle counters of
	    different processors need not be synchronised.
-- 
Dr Peter Chubb  http://www.gelato.unsw.edu.au  peterc AT gelato.unsw.edu.au
The technical we do immediately,  the political takes *forever*

Re: How to measure time accurately.

[Chris Friesen]

Peter Chubb wrote:
>>>>>>"Chris" == Chris Friesen <cfriesen@nortel.com> writes:

> Chris> Most cpus have some way of getting at a counter or decrementer
> Chris> of various frequencies.  Usually it requires low-level hardware
> Chris> knowledge and often it needs assembly code.
> 
> As a device driver is inside the linux kernel (unless you're writein a
> user-mode device driver :-)) you can use the getcycles() macro that's
> defined for most architectures.  It provides a snapshot of the
> cycle-counter.

For ppc this only gives 32-bit values, which overflow every 129 seconds 
on my G5.  Depending on how long you're trying to time, this could be a 
problem.

Chris

Re: How to measure time accurately.

[Jan Engelhardt]

> For ppc this only gives 32-bit values, which overflow every 129 seconds on my
> G5.  Depending on how long you're trying to time, this could be a problem.

Just take an extra measure to "record" overflows (2^32-1 => 0) and you're set.



Jan Engelhardt
-- 
No TOFU for me, please.