Profiling code slows runtime (CpuTime getrusage)

[svigerske]

Issue created by migration from Trac.

Original creator: bdavis

Original creation time: 2018-02-28 20:13:17

Assignee: ipopt-team

Version: 3.12

I am running Ipopt on MacOS Sierra built with clang. I am using the Matlab Interface on 2017a.
My problem order is low (2 variables, no constraints), but I need to run many times with different input.

I ran a profile with Instruments and found that a large amount of time was spent in the getrusage system call. This happens in IpUtils.cpp CpuTime() and SysTime().

My hack solution (attached) was to add a compiler define around this to bypass the calling of getrusage and simply return zero for the profile times. This improved my runtime by about 33%.

I wish there was a more elegant solution to this (an option?) that could be configured at runtime.

[svigerske]

Comment by bdavis created at 2018-02-28 20:16:19

I was unable to attach my file as it was flagged as spam :-(

[svigerske]

With Ipopt 3.14, detailed timing statistics are no longer recorded by default. If no timelimit is set, then there are calls to CpuTime(), SysTime(), WallTime() only at the begin and end of the Ipopt run (so it can report the overall runtime). If a timelimit is set, then there are additional calls at every iteration, but still this is much less than the previous behavior that recorded the time spend in various components of Ipopt and function evaluations.
There are options to reenable the detailed timing.