enso/docs/profiler/profiling.md

Summary

This document provides a guide on how to achieve and maintain high-performance evaluation of Enso (IDE, Engine, Language Server, and other services). It is divided into the following chapters:

• Profiling
  Discovering performance bottlenecks and understanding their causes.
• Performance goals
  Desired performance and comparison to other products.
• Performance improvements plan
  Strategies on improving performance of the selected components.
• Continuous performance monitoring
  Maintenance of high performance of all components and early discovery of new performance issues.

Profiling

To understand, improve, and maintain performance of the IDE, we need to be able to reliably measure the performance of actions it performs.

Measurables

Measurements we make will fall into one of two categories:

• Objective measurements refer to intervals directly observable by the user, e.g. time from opening the app until the loading spinner gets out of the way.
• Internal measurements provide insight into how time is spent while performing the objectives, e.g. time spent compiling ide.wasm.

Objectives to measure

The following objectives should be measured by the framework; it is part of this task to identify the most important user-visible measurements. Internal measurements will be defined as-needed to understand how time is spent performing objectives.

• App startup
  Time between double-clicking the app icon and having functional hello screen of recent projects. Sub-processes:

  • Hello Screen initialization
    Time between double-clicking the app icon and displaying the loading spinner:
    • App window initialization
      Time between double-clicking the app and displaying the user log-in prompt.

• Project startup
  Time from selecting a project from the Hello Screen to a working graph view of the project. See also engine startup for non-IDE details on a similar topic. Sub-processes:

  • Graph Editor initialization
    Time needed to display fully-functional graph editor (not colored). Profiling should include:
    • Time needed for GUI Backend to receive and process the required information from the Engine.
    • FRP events propagation. This should mainly focus on the public Graph Editor FRP API (used by GUI backend):
      • Time needed to propagate FRP events through the system.
      • Count of FRP events of a particular type, e.g. how many times a text refresh was requested.
    • Time needed to create each node component (model, display object, etc.).
    • FPS statistics (explained later).
  • Type information propagation
    Time needed to color the graph. Profiling should include:
    • Count of FRP events of a particular type. For example, how many times a text refresh was requested.
    • Time needed to update each particular node.
    • FPS statistics (explained later).

• Working with a project
  This section should contain various objectives, including:

  • Creating/moving/editing/removing a node.
  • Connecting nodes.
  • Editing expressions (incl. typing a letter).
  • Entering nodes.
  • Opening/changing/closing a visualization.
  • Saving the project.
  • Switching projects.
  • Opening node searcher.
  • Switching selection in node searcher.

• Idle performance
  The profiling statistics for opened project that does not receive any user-interactions.

Also, please browse the performance-related issues to check if they are not describing processes missing from the above list. In particular, please check the IDE Performance Issues issue.

Measurement relationships

The data model for measurements is hierarchical, to enable use ranging from high-level overviews to fine-grained analysis of details.

Each of the trees comprising the hierarchy is rooted at an objective measurement, because we only measure intervals that contribute to a user-observable delay.

The parent-child relationship of the hierarchy indicates that the child was started by the parent. Usually, but not always, a task's parent cannot complete until the child does.

Beside start and end times, the framework logs when tasks are paused or resumed (e.g. before and after an .await); this provides enough information to define a current profiler: The current profiler is the most-recently started-or-resumed profiler that is not currently paused and has not yet ended.

Some profiling data such as FRP events and per-frame statistics is implicitly attributed to the current profiler [these terms will be elaborated on in a future revision of this document].

Collecting the data

Interactively

Users may capture profiling information at any time during an interactive session of the application (usage in debug scenes is not yet supported). The procedure to collect data is:

• Compile Enso with the desired profiling level enabled. The default (which is used in release builds) is to record only objective measurements, which includes information about performance, but little detail for identifying causes. A greater level of detail can be obtained by passing the profiling-level argument to the build script (see "Development" subheading); for maximum detail, use --profiling-level=debug. For example:

  ./run ide build --profiling-level=detail
  

• To profile in Electron:
  • Start Enso with the save-profile option, e.g.:

    ./distribution/bin/enso --save-profile=your_profile.json
    

  • Perform any activity the profiler should record (the profiler is always running).
  • Press Ctrl+Alt+Q to save profile (to the path specified earlier) and quit.
• To profile in Chrome:
  • Open Enso in Chrome.
  • Perform any activity the profiler should record (the profiler is always running).
  • Press Ctrl+Alt+P to end profile recording.
  • Open the Browser Console by pressing Ctrl+Alt+J (do not do this before profiling; having the console open can degrade performance of WASM applications.)
  • Near the end of the Console output, there will be a JSON object printed; it will usually be large, and truncated. Click the Copy button below it.
  • Paste the log into a .json file.

Batch mode

A batch mode supports recording profiling information for predefined workflows without user-interaction. It can be run as follows:

# First build with full profiling information.
./run ide build --profiling-level=debug
# Now run the generated binary in batch mode.
./distribution/bin/enso --entry-point profile --workflow create_node --save-profile create_node.json

The profile entry point selects batch mode; when it is used, the save-profile and workflow arguments are mandatory. save-profile specifies an output filename; workflow selects from a list of defined batch-mode workflows to measure. Supported workflows include:

• new_project: This workflow includes starting the application, opening an empty project, and closing the application.
• collapse_nodes: Measures selecting and collapsing a group of nodes.
• create_node: Measures adding a new node to the graph.
• enter_collapsed_node: Measures entering a collapsed node.
• open_visualization: Measures enabling a visualization for a node.
• open_project_orders: Measures opening the project called Orders.

To see a full list of currently-implemented workflows, run the enso executable with the arguments --entry-point=profile, and --workflow=help. Additional workflows will be implemented in the future.

The profiling environment

An important consideration when profiling is controlling the environment. In particular, setting a fixed CPU frequency is important for measurement consistency (with Turbo Boost enabled I have been seeing up to a 2x difference between consecutive runs; with it disabled, it is only around 20%). Because achieving this depends on hardware, it is up to the user to do so; it is recommended to create a script that automatically disables turbo boost before running benchmarks rather that do so manually, since a manual approach is error-prone. On Linux, frequency scaling can be controlled by writing to /sys values (which values depends on processor); on AWS instances, it can be set with a kernel command line (https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/processor_state_control.html#baseline-perf). In any environment where frequency cannot be disabled (or when running with it enabled to obtain realistic measurements of reference hardware), high variance will make it more difficult to detect performance changes.

Examining the data

Run-graphs

A "run graph" tool provides a high-level overview of the data in the profile. The upper part of the view is an icicle plot[^1] showing data that is similar to a callgraph¹, but rather than show function scopes it shows profiler scopes (these often coincide). The lower part of the view, sharing the (horizontal) time axis, is a graph of async tasks, showing their active and inactive intervals. [^1]: https://www.cs.middlebury.edu/~candrews/showcase/infovis_techniques_s16/icicle_plots/icicleplots.html

Invocation:

./distribution/bin/enso --entry-point=profiling-run-graph --load-profile=profile.json

It takes as input the profiling data representing a single run (averaged data would not be suitable for graphical representation).

Command-line tools

• intervals (cargo run -p enso-profiler-data --bin intervals < profile.json): This tool produces a table showing which profilers spent the most time running.
• measurements (cargo run -p enso-profiler-data --bin measurements < profile.json): This tool produces a textual representation of all the data in a profile, including the hierarchy and lifetimes of async profilers, sync profilers, and metadata. It supports fine-grained examination of profile contents, and profiler-based debugging.

Format conversion

• devtools (cargo run -p enso-profiler-data --bin devtools < input.json > output.json): This tool converts a profile from Enso's profile format to the format used by Chrome Dev Tools, for compatibility with third-party tools.

Examining frontend-backend timing breakdowns of tasks

Some additional tools support examining how different processes are contributing to the latency of different tasks, and examining the communication between processes (processes in enso-profiler-data, and message_beanpoles in enso-profiler-enso-data).

Averaging profile data

Given multiple recorded profiles of the same workflow, it is possible to compute averages of specific measurements; this is essential for reducing variance to support A/B testing (see [#future-extensions]), above. However, there is no way to produce a complete "averaged profile" (suitable for visualization) from multiple individual profiles, for various reasons. The average of multiple profiles is hard to even define: A particular measurement may repeat a different number of times in one profile from another. Even ignoring that, an "average" profile wouldn't necessarily represent a set of timings that could occur in a real profile: For example, a property may hold in all profiles that Task A and Task B never occur at the same time; yet if they occur in different orders in individual profiles, their intervals could overlap in an averaged profile.

Epic 1: Profiler Implementation

Because some measurements need to start before WASM is initialized, the profiling framework should provide an implementation with two interfaces – for JavaScript and Rust. Some API syntax will differ, but both APIs should support the same functionality; when the data is emitted, profiling entries from JavaScript and Rust will be merged into one log. There should be no difference in the behavior between calling APIs in any of the languages.

To minimize performance impact of profiling, instrumentation should do not require any more time-consuming operations than obtaining a timestamp (with performance.now) and pushing an entry to a buffer. These entries will be emitted at process exit.

Profiling levels

Fine-grained profiling (like per-frame rendering profiling) is useful for understanding the performance characteristics, but it can cause significant performance overhead. That is why the API should define a few profiling levels:

• objective
  Measurements that correspond directly to aspects of the user experience. An objective can contain other objectives, e.g. GUI initialization (which might be defined as: time from opening the app until the app is ready to receive user input) contains time until the loading spinner finishes.
• task
  Coarse-grained tasks, such as app window initialization, GUI downloading, or WASM compilation. A task can contain other tasks e.g. GUI initialization contains GUI downloading.
• details
  All processes which can be removed in compile-time for the official stable release for users. We might provide some users with special releases with enabled detailed profiling, however, it should be possible to debug and understand most of user-provided logs with disabled details view.
• debug
  All processes which should be removed in compile-time by default even during app development. It applies to every heavy-usage of the profiling framework, such as per-frame rendering profiling.

When a profiling level is disabled, all instrumentation associated with that level is eliminated at compile time. In order to avoid passing real profiler values down the call stack when they are not needed by enabled profiling, each type of profiler will conditionally compile to either a real object or a zero-sized type with no-op implementations. This conditional compilation is controlled with the profiling-level family of Cargo feature flags, and integrated with the build script, e.g. ./run watch --profiling-level=task.

Rust API

The primary interface is a profile attribute macro. When applied to a function definition, it inserts instrumentation to measure the function's entry and exit times. The measurement is logged under the name of the function, with file/line information attached.

// The argument to the macro identifies the profiling level at which this measurement is enabled.
#[profile(Detail)]
fn compute_four() -> u32 {
    2 + 2
}

Low-level API

A lower-level API is available for special purposes. The #[profile] API above is preferred, especially for non-async measurements.

// Sometimes a block that isn't a whole function may need to be measured; in that case, a RAII
// interface is available.
//
// A trait `profiler::Parent<T>` supports accepting as an argument any object representing a
// profiler that could be a parent of a profiler type T. This includes profilers of the same or
// higher profiling level, and their corresponding RAII objects (described below).
//
// Usage note:
// When the RAII guard is implicitly dropped at the end of a block, it should be created as the
// first line of the block; it measures one full block, and the line containing `start_X!` acts as a
// title for the block.
fn big_computation(prof: impl profiler::Parent<profiler::Task>) {
    match some_condition {
        // When task-level profiling is enabled, the `start_task!` line will create a measurement
        // named "condition_a". The measurement will extend from the line to the end of its
        // enclosing scope, when the __profiler object is dropped.
        Condition::A => {
            let _profiler = profiler::start_task!(prof, "condition_a");

            // If detail-level profiling is also enabled, "small_computation" will log a
            // child of the "condition_a" task.
            let foo = small_computation(_profiler);
            ...
        }
    }
}

// `await` in a context profiled by the low-level API should be done through a wrapper that logs the
// beginning and end of the await interval. This is necessary to have accurate information about
// what tasks are running when.
//
// If this is not done, tools interpreting profiling data will be able to detect the oversight and
// identify the profiler at fault, as long as the callee, or any async task that runs during the
// await period, is correctly profiled.
//
// However if the above detectable case does not occur, and during the await period an unprofiled
// async task starts an implicitly-parented profiled task, then some profiler(s) that should be
// root-level may be interpreted as children of this profiler.
async fn awaiting(prof: impl profiler::Parent<profiler::Task>) {
    profiler::await_!(some_future(), prof);
}

// Sometimes we need multiple measurements from the same start time; to support this pattern, there
// is an interface to create a child profiler that starts at the same time as its parent.
fn same_start_child(prof: impl profiler::Parent<profiler::Task>) {
    let _profiler = profiler::task_with_same_start!(prof, "same_start_child");
    ...
}

// The hierarchy of measurements starts at a const Objective object, APP_LIFETIME.
//
// Since APP_LIFETIME is not a true objective measurement (i.e. we aren't, in general, trying to
// minimize its length), its direct children must be objective measurements, and form the logical
// roots of the measurement hierarchy.
fn a_root_objective() {
    let _profiler = profiler::start_objective!(profiler::APP_LIFETIME, "a_root_objective");
    ...
}

// Some measurements need to start when the app starts (the "time origin" in Web terminology);
// APP_LIFETIME can be used with with_same_start.
fn starts_at_beginning() {
    let _profiler = profiler::objective_with_same_start!(
        profiler::APP_LIFETIME,
        "starts_at_beginning"
    );
    ...
}

// Rarely, a single block may need to be divided into multiple measurements.
// This can be accomplished by creating sub-blocks, and using the same RAII interface as above.
fn multipart_fn(prof: impl profiler::Parent<profiler::Task>) {
    let foo;
    let bar;
    {
        let _profiler = profiler::start_task!(prof, "first_part");

        foo = calculate_foo();
        bar = calculate_bar();
    }
    {
        let _profiler = profiler::start_task!(prof, "second_part");

        do_something(foo, bar);
    }
}

// Occasionally, we need to measure an interval that doesn't start and end in any one scope.
//
// This can be accomplished by storing the RAII object returned by `start_X!` and ensuring it is
// dropped at the time the measurement should end.
//
// In this example, the measurement lasts for the life of the containing `Baz`.
struct Baz {
    ..,
    // profiler::Started<profiler::Objective> holds a profiler::Objective, and a start time.
    // When it is dropped, the end time will be set and the measurement will be logged.
    profiler: profiler::Started<profiler::Objective>,
}
impl Baz {
    fn new(parent: impl profiler::Parent<profiler::Objective>) -> Self {
        let profiler = profiler::start_objective!(parent, "doing Baz");
        Baz { .., profiler }
    }
}

JavaScript API

The API in JavaScript supports the same functionality as the Rust low-level API above, but is used a bit differently because we can't support RAII in JS.

// Higher-level API, for scoped measurements: Measure a block by passing it as a closure.
await profiler.objective.measure(profiler.APP_LIFETIME, "init-gui", async (subprof) => {
    ...
});

// Lower-level API: Start and end explicitly.
const prof = profiler.objective.start(profiler.APP_LIFETIME, "a root objective");
const subprof = profiler.task.with_same_start(prof, "a subtask");
subprof.end();
prof.end();

Profiling data

File format

After profiling data is captured, it is written to a file. The data model of this file is an event log that reflects the raw data collected by the profiling framework. It is a low-level format, considered internal to the profiler crates.

Some properties of the format are important to document. Its data can be divided into two categories:

• The interval data is information about the profilers themselves: their lifetimes, state changes, and relationships.
• Metadata provides information about the application, usually associated with specific profilers.

The definitions of the interval data will not change often; metadata is expected to change as more data-logging is supported. It is desirable that when a changeset alters a metadata definition, we can still tell whether that changeset introduces a performance regression in the interval data; more generally, we can check performance against a baseline even if metadata changes have happened since the baseline was recorded. To ensure this:

• The definitions of interval data should be as stable as possible.
• The format is designed and implemented such that when a metadata type's definition changes, we can still read at least the intervals (We will also still be able to read metadata of other types, although this is not essential).

In the initial implementation, there is no explicit versioning to detect changes to the datatype definitions; we rely on the self-describing format (JSON) to recognize changes. Because of this:

• Developers altering metadata definitions should try not to introduce an important change to the semantics of a field without any concurrent structural change. (E.g. If a field's meaning has changed, change its name too.)
• Developers interpreting profiling data should be aware that tools may misinterpret some metadata if producer and consumer are subtly incompatible versions.

Measurement API

The profiler framework includes a data library which interprets the low-level event log data, and makes the information available through a higher-level API.

Open questions

There are open questions that should be answered as part of implementation of this design idea. Please note, that you should not answer them by yourself. Instead, you talk with other team members who have broad domain-specific knowledge in these areas, please. Questions:

• What will the interface to profiling data be?
• What tools will be implemented to support working with profile files? (Need to support: comparing mean timings between profiles; including variance between sets of profiles)

Epic 1: Performance Goals and Reference Hardware

To know where we stand in terms of performance, we need to have references. This means we have to look at how other software behaves and compare how slow/fast Enso is for objective measures.

Hardware

The tests should be conducted on the following hardware types:

• The newest MacBook Pro.
• An "average client laptop" without dedicated GPU, which should be a "common" hardware used in enterprises nowadays.
• A "minimal-spec laptop".

Open questions

There are several open questions that should be answered as part of implementation of this design idea. Please note, that you should not answer them by yourself. Instead, you talk with other team members who have broad domain-specific knowledge in these areas, please. Questions:

• What is the exact hardware specification that we should consider "average client laptop" and "minimal-spec laptop" described above?
• What is the time needed for other software to perform the objective-level actions and which software we should compare Enso to (e.g. Alteryx, Knime, Tableau, ...).

Epic 2: Performance improvements planning

Once we have determined our performance goals and created the tools to measure them, we can focus on improving the most significant performance bottlenecks. The process would be as follows:

• We will compare the current results with results of other software, and we will describe our performance goals.
• We will analyze the profiling information and check which processes need improvements in order for the goals to be matched.
• We will create user-stories (tasks) for improving these components, and we will plan their implementation. All tasks will become sub-tasks of one performance-related epic.

Epic 3: Continuous performance monitoring

As a last step, we should take what we have learned, and ensure that we do not introduce regressions in the future. We should add a CI step that fails if we introduce new performance regressions.

To achieve this we should add a command line interface that makes a reference report, tests the application, and returns whether the test was worse in any of the measurements than the reference report (including some margin of error to avoid spurious failures).

We then add a reference performance report to the build system and compare new PRs against this report. If they are introducing regression, CI will fail and the performance of the PR will need to be checked/improved.

Moreover, the CI should generate time reports over time, in a similar fashion to the performance reports in Rust.

In order to allow high sensitivity without excess false positives and with a minimum of required testing resources, we should probably have a way of tagging issues that may affect specific objective-level measures. For example, the most expensive measurement to take is time to initialize the app and open a project; looking at the most recent PRs, most could safely be assumed to have no impact on startup time. (Of course, it is important to err on the side of safety here--but I think the chance of an issue having a completely unforeseeable performance impact is lower than the overhead associated with running this test for every PR; and if we run it too often, we'll get used to seeing false positives, which reduces the value of testing).

Future extensions

A/B-testing tool

Additional profiling information

The time profiling is a good start for understanding application performance metrics, however, it is not enough to analyze causes of bottlenecks, especially in async-first applications, like graphical user interfaces.

The profiling toolkit should also collect per-frame statistics for the duration of each of the processes and report their summary, including:

• Graph Editor public API (used by GUI backend) FRP events count, e.g. how many times the Graph Editor was asked to create a node or modify an expression.
• Min, max, and average rendering statistics:
  • WASM memory usage.
  • GPU memory usage.
  • Sprites count.
  • Shaders count.
  • ... other statistics currently collected by the GUI performance monitor implemented in gui/src/rust/ensogl/lib/core/src/debug/stats.rs (you can open it by CMD/CTRL+ALT+<TILDE>, where <TILDE> is the button below the escape key and because of a bug, can be different from tilde on some keyboard layouts).

Another important type of measurement is cold-start performance: the amount of time it takes to reach important startup milestones on a freshly-booted machine, when the app's resources are not yet cached to memory and the processor has not recently run the process (so no branch prediction data etc. is cached yet). Because it requires many repeated measurements to control for high variance (and a reboot for each measurement), this data will take a lot of time to collect; it is not feasible to run it in regular CI tests. However, it reflects important aspects of real-world usage, so it is important data to check on a regular basis. (To be determined: how frequently.)

Production monitoring

It's important to know how performance benchmarks compare to actual user experience. In production builds, we could collect objective-level measurements, aggregate them into statistical reports, and submit them through the anonymized-data facilities. The details of this are outside the initial scope of this design.

More advanced run graphs

Run-graphs could support rendering callgraph data collected by V8 alongside the measurements generated by our instrumentation.

Run-graphs could support interactive functionality, such as collapsing the lowest-level profiling data until it is clicked.

---

1. https://en.wikipedia.org/wiki/Call_graph ↩︎