If I run the time command around a multi-threaded application, how is it affected? In this case, what do the user and sys times represent? Is it the sum of all times used by the various threads, or...?

The times reported for a process will be the time consumed by that process alone ... but this will be somewhat-less than elapsed time, and this is where/how the effect of multitasking can be seen.

When the dispatcher hands control to a process/thread, and then takes it away again a few microseconds later, it keeps an accounting of how much time the process consumed. When the process makes system-calls and subsequently returns from those calls, the elapsed-time spent servicing a system-call is also accounted for.

All of these times are intended to be used only for comparative purposes. They are not exact, and they can vary due to a myriad of possible factors.

I'm not quite clear on this -- because I used time a program using implicit parallelization (a schemik script) and I get times like so:

real 0m0.776s
user 0m1.310s
sys 0m0.020s

This is why I am wondering if "user" is the sum of the times, since presumably otherwise user cannot be greater than actual elapsed time, right?

"This is why I am wondering if "user" is the sum of the times, since presumably otherwise user cannot be greater than actual elapsed time, right?"

An application's time can be greater than real time if the application is multi-threaded and you time the application on a multi-CPU machine. However an application's time cannot be greater than real time multiplied by the number of CPUs.

I have done such multi-CPU timings on IBM mainframes where the application time exceeded real time but I have never tried to do it on an x86.

------------------------
Steve Stites