What is wrong with my profiling?

[HenriDeh]

I am profiling my Deep Learning algorithm and I am getting this weird shape. That long line on the left does not make sense, is it precompiling or something ? The name of the top bar is task.jl,poptaskref: line 660. Maybe this is related to the fact that I am using CuArrays for my computing. But I don’t get why it is not on top of eval.

By the way I wanted to ask. If on the flame graph on the right I am seeing a lot of typeinfer.jl and abstractinterpretation.jl bars, does it mean I have a problem with type stability ?

[Tamas_Papp]

1. did you make a small run to make sure everything is compiled? otherwise it will profile the compiler, which may not be the objective.

2. you may be running into

Profiling code reveals that most time is used calling 3 functions in task.jl

I profiled my call to a function and got the following plot: [profile_make_JuMP_model] Those first 3 long horizontal lines are calls to task_done_hook, then wait, then poptaskref (in ascending order). This happens even for the 2nd or 3rd call to my function, and I don’t understand what they could be. I stopped specifying a lot of the types used / returned in my functions after a while since I was losing time doing so properly - could this be the cause? How could I find out?

[HenriDeh]

1. Yes it was indeed the compiling.
2. I am running this on Juno so this may be the reason I get that weird bar. Next time I’ll try from the linux shell

Thank you

[pfitzseb]

You could also give Juno’s build in profile viewer (Juno.@profiler foo(...)) a go, which should filter out extraneous frames when you’re running with only one thread.

[HenriDeh]

Ha yes it works well with the juno profiler