using the correct tool to time my program

using the correct tool to time my program

[Mark Farnell]

I have a parallel program, which calls some message passing functions
which block.  Then I attempted to use times() (with struct tms) to
time the program.

the struct tms gives me the "program time" and the "system time" as
well as the child process time (if any).  However does it also include
the time spent when my program is blocked (i.e. network time).

This is important because although I use gprof to separate the CPU
time spent by my program vs libraries, when I calculate the total time
spent, I really need to include the time spent waiting for message
from network (network time).

Therefore is times() really appropriate for my purpose?  If not, what
other functions can I use instead? (such as gettimeofday?)

Thanks!

Mark

Re: using the correct tool to time my program

[Glynn Clements]

Mark Farnell wrote:

> I have a parallel program, which calls some message passing functions
> which block.  Then I attempted to use times() (with struct tms) to
> time the program.
> 
> the struct tms gives me the "program time" and the "system time" as
> well as the child process time (if any).  However does it also include
> the time spent when my program is blocked (i.e. network time).

No. times() measures CPU usage; any time spent idle isn't measured. 
Ditto for getrusage().

> This is important because although I use gprof to separate the CPU
> time spent by my program vs libraries, when I calculate the total time
> spent, I really need to include the time spent waiting for message
> from network (network time).
> 
> Therefore is times() really appropriate for my purpose?  If not, what
> other functions can I use instead? (such as gettimeofday?)

If you need to measure real time (aka "wall-clock time"), use
gettimeofday() or clock_gettime().

-- 
Glynn Clements <glynn@gclements.plus.com>