Re: [boost] Re: Re: Re: Re: Profiling Library suggestion

From: christopher diggins (cdiggins_at_[hidden])
Date: 2005-02-11 20:12:08

----- Original Message ----- 
From: "Preston A. Elder" <prez_at_[hidden]>
To: <boost_at_[hidden]>
Sent: Friday, February 11, 2005 7:57 PM
Subject: [boost] Re: Re: Re: Re: Profiling Library suggestion

> On Fri, 11 Feb 2005 19:30:36 -0500, christopher diggins wrote:
>
>> I am confused. I don't see how measuring time intervals with either
>> xtime_get() or QueryPerformanceCounter or std::clock() would differ apart
>> from resolution and accuracy.
>
> xtime_get retrieves the current wall clock time.  This means that a
> difference between xtime_get values will tell you the actual amount of
> seconds passed, you could duplicate its results with your watch.
>
> std::clock() and I assume QueryPerformanceCounter() give you an amount of
> clock ticks.  This number is only incremented when the CPU works on your
> task.

Does "your task" here mean the current thread or does it mean the current 
process (or does it depend)?

> Say you have a very highly loaded system, 100% CPU usage, lots of threads
> (say, 1000), you name it.
>
> The std::clock() difference would be very different from the xtime_get
> difference.  Why?  Because under those circumstances, the kernel may only
> get to run your task infrequently, and may not give it as much CPU time as
> it would get on a system with no load.
>
> Therefore, std::clock() is saying 'this is how much of the kernel CPU
> timeslice I got', where as xtime_get is telling you what the current time
> is, regardless of how much or how little CPU time the kernel is allocating
> to your process.
>
> I don't know of any windows commandline utilities to track this, but If
> you, say, had an application on linux/unix and ran 'time' on it, like so:
>
> time tar xvfz ../mtxdrivers-rh9.0-v1.1.0-pro-beta.tar.gz
>
> You would get the following output:
>
> real    0m0.414s
> user    0m0.166s
> sys     0m0.104s
>
> The 'real' is how much the difference between the wall clock times at the
> start and end of the task.
>
> The user and sys are how much time in 'CPU seconds' (aka. how many
> CPU clock cycles) the application was given to perform its task ('user'
> being how much time the user-space portion of the code used, and 'sys'
> being how much time the kernel took to execute its part of it (eg. disk
> I/O, etc).
>
> As you can see, the real time is greater than the sum of the other two
> because the kernel did not give the application 1 whole CPU to use for the
> entire duration of the application.
>
> So xtime_get is equivalent to 'real', and std::clock() is equivalent to
> 'user + sys'.  Otherwise known as 'how long did this run for?' and 'how
> long did this take two run?', which are different questions :)
>
> Assuming the kernel gave the application 1 whole CPU to use from start to
> end of the application's invocation, they would be equal.  This rarely
> happens.

Thank you very much for the explanation. What advantage would 
wall-clock-time have for profiling tasks?

Christopher Diggins
Object Oriented Template Library (OOTL)
http://www.ootl.org