Output wall clock time per iteration?

[ghost]

Hello,

I would like to know if there is an option to output the elapsed wall clock time per iteration? If there is no such function, i would like to modify the source code myself to add it. Can you point me to the file which i would have to edit?

[blegat]

It is done in #170

[ghost]

Nice, also thanks for your fast answer :)

Do i have to wait until the commit is merged or can i possibly use the branch with this commit right now? I am also thinking of modifying the source code until this feature is added.

Currently, when i try to use the branch with the commit, I get:

[blegat]

You need use the master branch of MathOptInterface and this branch of JuMP: jump-dev/JuMP.jl#2003 if you want to use the branch of #170

[ghost]

Thank you. It is working, but not for my use case (with PowerModels.jl). I think this issue can be closed for now, if you don't want to leave it open for further discussions.

[jchen975]

Thank you. It is working, but not for my use case (with PowerModels.jl). I think this issue can be closed for now, if you don't want to leave it open for further discussions.

Hi @snapdeb, were you able to resolve this issue on PowerModels.jl? I'm doing a large number of run_ac_pf() in parallel and I would like to know each run's number of iterations and each iteration's elapsed time. It would be a pain to parse the output file because of the From worker x outputs and all the other information. Any advice is appreciated!

[ghost]

Hi @jchen975,

The following screenshot shows the current state. The Ipopt#master branch is needed since the wall clock time per iteration is included there. Also i chose the PowerModels#master branch since it has fever unsatisfiable requirements (since it is more up-to-date).

The problem is that Ipopt#master depends on MathOptInterface v0.9.0, while PowerModels depends on MathOptInterface v0.8.0. Thus we need to wait until PowerModels receives a new update, or assist in the development if at all possible.

[jchen975]

@lschmelting Hey, thanks for your reply and opening the issue there. I suppose before that is updated, knowing the number of iterations in each run_ac_pf will suffice. Do you know if there is a function for that? I'm still pretty new to using JuliaOpt, but does BarrierIterations() have anything to do with what I want?

Alternatively, I would have to parse the Ipopt output, but I really don't want to do that if there's something I can use directly.

[ghost]

When you call JuMP you should use the default print_level=5, which if I remember correctly, includes the number of iterations and some info for each iteration. Some additonal timing statistics are shown when using print_timing_statistics="yes". Additionally, you can specify an output_file.

JuMP.with_optimizer(Ipopt.Optimizer, print_level=5, print_timing_statistics="yes", output_file="C:\\Users\\lschmelting\\workspace\\Ipopt_Output")

[jchen975]

When you call JuMP you should use the default print_level=5, which if I remember correctly, includes the number of iterations and some info for each iteration. Some additonal timing statistics are shown when using print_timing_statistics="yes". Additionally, you can specify an output_file.

JuMP.with_optimizer(Ipopt.Optimizer, print_level=5, print_timing_statistics="yes", output_file="C:\\Users\\lschmelting\\workspace\\Ipopt_Output")

Yes, and actually setting print_level=3 prints the number of iterations without further details, which I think is what I need for now. However, specifying an output file still requires further inspection or parsing to find the itr # information. Is there a better way to do that directly from PowerModels/JuMP/Ipopt?

[ghost]

I don't know of a better way, maybe someone else does.

[jchen975]

This short python script does the job. It's actually easier than I expected, but it would be nice to have a function to get #itr instead of having to parse the output file.

filepath = "./test.log"  # change this to the directory containing the output.log
key = "Number of Iterations...."

itr = []
with open(filepath) as fp:
	s = fp.readline()
	while s:
		l = [ss.strip() for ss in s.split(":")]
		if len(l) >= 2 and l[-2] == key:
			itr.append(int(l[-1]))
		s = fp.readline()

[mlubin]

Unless you can find somewhere to extract this information from Ipopt's API, there's not really much we can do on the Julia side.

[ghost]

wall_clock_time_per_iteration.txt

@jchen975: PowerModels is again usable with the Ipopt master branch: add Ipopt#master PowerModels#master.

I don't see a wall clock time per iteration output, though.
I use Ipopt like this: ipopt_solver = JuMP.with_optimizer(Ipopt.Optimizer, print_level=5, print_timing_statistics="yes", output_file="C:\\Users\\lschmelting\\workspace\\test").

Any hints?

[jchen975]

@lschmelting Sorry, I have no clue either. @ccoffrin could you help with this?

[ccoffrin]

I am not aware of an Ipopt feature that will print the time-per iteration, any insights @snapdeb, @blegat?

Another option is to inquire about this on the coin-or mailing list, e.g. see https://www.coin-or.org/. Once the required feature at the C++ level is clear we can see how to make it accessible from MOI/JuMP.

For what it is worth PowerModels#master is now tested with Ipopt.jl v0.6, which includes the udates from #170.

[dpo]

In the mean time, you could pass a callback and measure time yourself.