Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

Sign up
Appearance settings

Expose a profiler API #229

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Matyrobbrt wants to merge 31 commits into lucko:master
base: master
Choose a base branch
Loading
from TheModdingInquisition:spark-api

Conversation

@Matyrobbrt
Copy link

@Matyrobbrt Matyrobbrt commented Jul 17, 2022
edited
Loading

This PR exposes a profiler API in spark-api.

Changes made

  • Added a new subproject, spark-proto which all other projects depend on (including the API). This allows the API to access the java object of reports
  • Added a new subproject, api-test, which is a test Forge mod for testing the API.
  • Different classes in the me.lucko.spark.common.sampler package have been moved in the API, some of them as interfaces with hidden implementations using Spark
  • SamplerModule has been changed to use the implementation of the API

Maven setup

As configured in spark-api/build.gradle and spark-proto/build.gradle, I recommend the following maven structure:

  • base group me.lucko.spark
  • Proto in me.lucko.spark:spark-proto shading the proto library
  • API in me.lucko.spark:spark-api, depending on me.lucko.spark:spark-proto

API Usage

Example API usage:

private static void profile() throws Exception {
 final Spark api = SparkProvider.get();
 final Profiler profiler = api.profiler();
 // Infinite report
 {
 final Profiler.Sampler sampler = profiler.createSampler(ProfilerConfiguration.builder()
 .dumper(ThreadDumper.ALL)
 .grouper(ThreadGrouper.AS_ONE)
 .ignoreNative()
 .build(), (e, msg) -> System.err.println("Error creating sampler: " + e + ": " + msg));
 if (sampler == null)
 throw new AssertionError();
 sampler.start(); // Start the profiler
 Thread.sleep(1000 * 60 * 5); // Wait 5 minutes
 sampler.stop(); // Stop the profiler
 final ProfilerReport report = sampler.dumpReport(ReportConfiguration.builder()
 .comment("My comment")
 .sender("Me", UUID.randomUUID())
 .build());
 System.out.println("Report 1 uploaded to " + report.upload());
 report.saveToFile(Path.of("report1.sparkprofile")); // Save the report
 }
 // Report with set time
 {
 final Profiler.Sampler sampler = profiler.createSampler(ProfilerConfiguration.builder()
 .dumper(new SpecificThreadDumper(Thread.currentThread()))
 .grouper(ThreadGrouper.BY_NAME)
 .ignoreSleeping()
 .samplingInterval(12)
 .forceJavaSampler()
 .duration(Duration.ofMinutes(10))
 .build(), (e, msg) -> System.err.println("Error creating sampler: " + e + ": " + msg));
 if (sampler == null)
 throw new AssertionError();
 sampler.start(); // Start the profiler
 sampler.onCompleted(ReportConfiguration.builder()
 .separateParentCalls(true).build())
 .whenComplete((report, throwable) -> {
 if (throwable != null)
 throw new RuntimeException(throwable);
 System.out.println("Report 2 uploaded to " + report.upload());
 }); // Configure the dump after 10 minutes, when the sampler is done
 }
}

Effectively supersedes #131.

Copy link

+1. I would like to use Spark to profile the integrated server startup (i.e. before I can enter a command) and this would be very useful.

Matyrobbrt added 20 commits July 18, 2022 17:28
Copy link
Author

The PR is ready for review.

Copy link
Owner

lucko commented Aug 29, 2022
edited
Loading

Hi, thanks for the PR! I will try to have a deeper look at this soon.

I will be totally honest/upfront with you, I have a few reservations about merging this currently:

  • Systems that automatically trigger the profiler and upload the data are already causing a problem for me in terms of data storage (remember that spark data is uploaded centrally). I am concerned that this is only going to get worse if we add an API for it. I would at least like to get that under control (it is barely at the moment) & fully understand the $ costs before we add a feature to accelerate the data usage 😛
  • I would prefer to not have implementation classes in the API package/module. It should contain only interfaces and enums ideally.
  • If the API is tightly coupled to the implementation (as it seems to be at the moment), it makes it more difficult to make changes to spark-common and the platforms in the future, which may hinder improvements or new functionality down the line.

That's where I'm at atm :) please feel free to ping/chat to me on Discord (https://discord.gg/PAGT2fu) if you have any comments or suggestions, or of course you can reply here.

Copy link
Author

Moved discussion to Discord

Copy link

ArDiDEx commented May 8, 2023

hey, i'm curious if this is still planned to be added or has been cancelled

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

AltStyle によって変換されたページ (->オリジナル) /