Skip to content

timholy/ProfileView.jl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ProfileView.jl

Build Status codecov PkgEval

NOTE: Jupyter/IJulia and SVG support has migrated to the ProfileSVG package.

Introduction

This package contains tools for visualizing and interacting with profiling data collected with Julia's built-in sampling profiler. It can be helpful for getting a big-picture overview of the major bottlenecks in your code, and optionally highlights lines that trigger garbage collection as potential candidates for optimization.

This type of plot is known as a flame graph. The main logic is handled by the FlameGraphs package; this package is just a visualization front-end.

Compared to other flamegraph viewers, ProfileView adds interactivity features, such as:

  • zoom, pan for exploring large flamegraphs
  • right-clicking to take you to the source code for a particular statement
  • analyzing inference problems via code_warntype for specific, user-selected calls

These features are described in detail below.

Installation

Within Julia, use the package manager:

using Pkg
Pkg.add("ProfileView")

Tutorial: usage and visual interpretation

To demonstrate ProfileView, first we have to collect some profiling data. Here's a simple test function for demonstration:

function profile_test(n)
    for i = 1:n
        A = randn(100,100,20)
        m = maximum(A)
        Am = mapslices(sum, A; dims=2)
        B = A[:,:,5]
        Bsort = mapslices(sort, B; dims=1)
        b = rand(100)
        C = B.*b
    end
end

using ProfileView
@profview profile_test(1)  # run once to trigger compilation (ignore this one)
@profview profile_test(10)

@profview f(args...) is just shorthand for Profile.clear(); @profile f(args...); ProfileView.view(). (These commands require that you first say using Profile, the Julia profiling standard library.)

If you use ProfileView from VSCode you'll get an error UndefVarError: @profview not defined. This is because VSCode defines its own @profview, which conflicts with ProfileView's. Fix it by using ProfileView.@profview.

If you're following along, you may see something like this:

ProfileView

(Note that collected profiles can vary by Julia version and from run-to-run, so don't be alarmed if you get something different.) This plot is a visual representation of the call graph of the code that you just profiled. The "root" of the tree is at the bottom; if you move your mouse along the long horizontal bar at the bottom, you'll see a tooltip that's something like

boot.jl, eval: 330

This refers to one of Julia's own source files, base/boot.jl. eval is the name of the function being executed, and 330 is the line number of the file. This is the function that evaluated your profile_test(10) command that you typed at the REPL. (Indeed, to reduce the amount of internal "overhead" in the flamegraph, some of these internals are truncated; see the norepl option of FlameGraphs.flamegraph.) If you move your mouse upwards, you'll then see bars corresponding to the function(s) you ran with @profview (in this case, profile_test). Thus, the vertical axis represents nesting depth: bars lie on top of the bars that called them.

The horizontal axis represents the amount of time (more precisely, the number of backtraces) spent at each line. The row at which the single long bar breaks up into multiple different-colored bars corresponds to the execution of different lines from profile_test. The fact that they are all positioned on top of the lower peach-colored bar means that all of these lines are called by the same "parent" function. Within a block of code, they are sorted in order of increasing line number, to make it easier for you to compare to the source code.

From this visual representation, we can very quickly learn several things about this function:

  • On the right side, you see a stack of calls to functions in sort.jl. This is because sorting is implemented using recursion (functions that call themselves).

  • mapslices(sum, A; dims=2) is considerably more expensive (the corresponding bar is horizontally wider) than mapslices(sort, B; dims=1). This is because it has to process more data.

Color Coding

  • Red: is reserved for function calls that have to be resolved at run-time. Because run-time dispatch (aka, dynamic dispatch, run-time method lookup, or a virtual call) often has a significant impact on performance, ProfileView highlights the problematic call in red. It's worth noting that some red is unavoidable; for example, the REPL can't predict in advance the return types from what users type at the prompt, and so the bottom eval call is red. Red bars are problematic only when they account for a sizable fraction of the top of a call stack, as only in such cases are they likely to be the source of a significant performance bottleneck. In the image above, can see that mapslices relied on run-time dispatch; from the absence of pastel-colored bars above much of the red, we might guess that this made a substantial contribution to its total run time. (Your version of Julia may show different results.) See Solving type-inference problems below for tips on how to efficiently diagnose the nature of the problem.

  • Yellow: indicates a site of garbage collection, which can be triggered at a site of memory allocation. You may find that such bars lead you to lines whose performance can be improved by reducing the amount of temporary memory allocated by your program. One common example is to consider using @views(A[:, i] .* v) instead of A[:, i] .* v; the latter creates a new column-vector from A, whereas the former just creates a reference to it. Julia's memory profiler may provide much more information about the usage of memory in your program.

GUI features

Customizable defaults:

Some default settings can be changed and retained across settings through a LocalPreferences.toml file that is added to the active environment.

  • Default color theme: The default is :light. Alternatively :dark can be set. Use ProfileView.set_theme!(:dark) to change the default.

  • Default graph type: The default is :flame which displays from the bottom up. Alternatively :icicle displays from the top down. Use ProfileView.set_graphtype!(:icicle) to change the default.

Gtk Interface

  • Ctrl-q and Ctrl-w close the window. You can also use ProfileView.closeall() to close all windows opened by ProfileView.

  • Left-clicking on a bar will cause information about this line to be printed in the REPL. This can be a convenient way to "mark" lines for later investigation.

  • Right-clicking on a bar calls the edit() function to open the line in an editor. (On a trackpad, use a 2-fingered tap.)

  • CTRL-clicking and dragging will zoom in on a specific region of the image. You can also control the zoom level with CTRL-scroll (or CTRL-swipe up/down).

    CTRL-double-click to restore the full view.

  • You can pan the view by clicking and dragging, or by scrolling your mouse/trackpad (scroll=vertical, SHIFT-scroll=horizontal).

  • The toolbar at the top contains two icons to load and save profile data, respectively. Clicking the save icon will prompt you for a filename; you should use extension *.jlprof for any file you save. Launching ProfileView.view(nothing) opens a blank window, which you can populate with saved data by clicking on the "open" icon.

  • After clicking on a bar, you can type warntype_last and see the result of code_warntype for the call represented by that bar.

  • ProfileView.view(windowname="method1") allows you to name your window, which can help avoid confusion when opening several ProfileView windows simultaneously.

  • On Julia 1.8 ProfileView.view(expand_tasks=true) creates one tab per task. Expanding by thread is on by default and can be disabled with expand_threads=false.

NOTE: ProfileView does not support the old JLD-based *.jlprof files anymore. Use the format provided by FlameGraphs v0.2 and higher.

Solving type-inference problems

Cthulhu.jl is a powerful tool for diagnosing problems of type inference. Let's do a simple demo:

function profile_test_sort(n, len=100000)
    for i = 1:n
        list = []
        for _ in 1:len
            push!(list, rand())
        end
        sort!(list)
    end
end

julia> profile_test_sort(1)  # to force compilation

julia> @profview profile_test_sort(10)

Notice that there are lots of individual red bars (sort! is recursive) along the top row of the image. To determine the nature of the inference problem(s) in a red bar, left-click on it and then enter

julia> using Cthulhu

julia> descend_clicked()

You may see something like this:

ProfileView

You can see the source code of the running method, with "problematic" type-inference results highlighted in red. (By default, non-problematic type inference results are suppressed, but you can toggle their display with 'h'.)

For this example, you can see that objects extracted from v have type Any: that's because in profile_test_sort, we created list as list = [], which makes it a Vector{Any}; in this case, a better option might be list = Float64[]. Notice that the cause of the performance problem is quite far-removed from the place where it manifests, because it's only when the low-level operations required by sort! get underway that the consequence of our choice of container type become an issue. Often it's necessary to "chase" these performance issues backwards to a caller; for that, ascend_clicked() can be useful:

julia> ascend_clicked()
Choose a call for analysis (q to quit):
 >   partition!(::Vector{Any}, ::Int64, ::Int64, ::Int64, ::Base.Order.ForwardOrdering, ::Vector{Any}, ::Bool, ::Vector{Any}, ::Int64)
       #_sort!#25(::Vector{Any}, ::Int64, ::Bool, ::Bool, ::typeof(Base.Sort._sort!), ::Vector{Any}, ::Base.Sort.ScratchQuickSort{Missing, Missing, Base.Sort.Insertio
         kwcall(::NamedTuple{(:t, :offset, :swap, :rev), Tuple{Vector{Any}, Int64, Bool, Bool}}, ::typeof(Base.Sort._sort!), ::Vector{Any}, ::Base.Sort.ScratchQuickSo
           #_sort!#25(::Vector{Any}, ::Int64, ::Bool, ::Bool, ::typeof(Base.Sort._sort!), ::Vector{Any}, ::Base.Sort.ScratchQuickSort{Missing, Missing, Base.Sort.Inse
           #_sort!#25(::Nothing, ::Nothing, ::Bool, ::Bool, ::typeof(Base.Sort._sort!), ::Vector{Any}, ::Base.Sort.ScratchQuickSort{Missing, Missing, Base.Sort.Insert
             _sort!(::Vector{Any}, ::Base.Sort.ScratchQuickSort{Missing, Missing, Base.Sort.InsertionSortAlg}, ::Base.Order.ForwardOrdering, ::NamedTuple{(:scratch, :
               _sort!(::Vector{Any}, ::Base.Sort.StableCheckSorted{Base.Sort.ScratchQuickSort{Missing, Missing, Base.Sort.InsertionSortAlg}}, ::Base.Order.ForwardOrde
                 _sort!(::Vector{Any}, ::Base.Sort.IsUIntMappable{Base.Sort.Small{40, Base.Sort.InsertionSortAlg, Base.Sort.CheckSorted{Base.Sort.ComputeExtrema{Base.
                   _sort!(::Vector{Any}, ::Base.Sort.IEEEFloatOptimization{Base.Sort.IsUIntMappable{Base.Sort.Small{40, Base.Sort.InsertionSortAlg, Base.Sort.CheckSor
v                    _sort!(::Vector{Any}, ::Base.Sort.Small{10, Base.Sort.InsertionSortAlg, Base.Sort.IEEEFloatOptimization{Base.Sort.IsUIntMappable{Base.Sort.Small{

This is an interactive menu showing each "callee" above the "caller": use the up and down arrows to pick a call to descend into. If you scroll to the bottom you'll see the profile_test_sort call that triggered the whole cascade.

You can also see type-inference results without using Cthulhu: just enter

julia> warntype_clicked()

at the REPL. You'll see the result of Julia's code_warntype for the call you clicked on.

These commands all use ProfileView.clicked[], which stores a stackframe entry for the most recently clicked bar.

Command-line options

The view command has the following syntax:

function view([fcolor,] data = Profile.fetch(); lidict = nothing, C = false, fontsize = 14, kwargs...)

Here is the meaning of the different arguments:

  • fcolor optionally allows you to control the scheme used to select bar color. This can be quite extensively customized; see FlameGraphs for details.

  • data is the vector containing backtraces. You can use using Profile; data1 = copy(Profile.fetch()); Profile.clear() to store and examine results from multiple profile runs simultaneously.

  • lidict is a dictionary containing "line information." This is obtained together with data from using Profile; data, lidict = Profile.retrieve(). Computing lidict is the slow step in displaying profile data, so calling retrieve can speed up repeated visualizations of the same data.

  • C is a flag controlling whether lines corresponding to C and Fortran code are displayed. (Internally, ProfileView uses the information from C backtraces to learn about garbage-collection and to disambiguate the call graph).

  • fontsize controls the size of the font displayed as a tooltip.

  • expand_threads controls whether a page is created for each thread (requires julia 1.8, enabled by default)

  • expand_tasks controls whether a page is shown for each task (requires julia 1.8, off by default)

  • graphtype::Symbol = :default controls how the graph is shown. :flame displays from the bottom up, :icicle displays from the top down. The default is :flame which can be changed via e.g. ProfileView.set_graphtype!(:icicle).

These are the main options, but there are others; see FlameGraphs for more details.

Source locations & Revise (new in ProfileView 0.5.3)

Profiling and Revise are natural partners, as together they allow you to iteratively improve the performance of your code. If you use Revise and are tracking the source files (either as a package or with includet), the source locations (file and line number) reported by ProfileView will match the current code at the time the window is created.