console VS pprof

tokio-console is a debugger for Rust async programs that use Tokio. To get started, add the console-subscriber crate to your project and add the following line which will initialise the subscriber and allow tokio-console to connect to it:

You could try https://github.com/tokio-rs/console to debug and profile what happens with tokio tasks in your program.

The tokio-console CLI is a fun one. The console-subscriber supports shipping to a console server running elsewhere, apparently. That gives you a window into what's happening now.

Tokio console maybe? https://github.com/tokio-rs/console

If I add "console_subscriber::init()" line as https://github.com/tokio-rs/console recommends, tracing_subscriber cannot be initialized.

I'm building an HTTP load tester called pdc! I have run out of obvious (to me at least) places to look for performance gains. I'm achieving around 45,000 requests per second, per core. Right now I'm using hyper with a separate tokio runtime (in current thread mode) running on each core. So far having runtime on each core/NUMA node has really helped with cache coherency. Any recommendations for profiling beyond tokio console or tokio metrics (Convenient timing amirite!)?

It was just an accident and has been fixed https://github.com/tokio-rs/console/issues/270.

P.S. Tokio [now also has Tokio Console](https://github.com/tokio-rs/console) allowing you to conveniently troubleshoot your tasks if they are causing issues :)

You can opt-in to async runtime such as tokio, and you can use tokio-rs/console for it's top-like metric

The pprof format is not tied to Go. From my understanding, it's used within Google across multiple languages. The format is defined in the pprof repository[0], and the visualization tool is source-language agnostic. I've seen libraries in numerous languages (e.g. Python, Java) to publish profiles in pprof format. This is an indicator the pprof format has become de-facto. Grafana Pyroscope[1] is a tool that's capable of parsing the pprof format, agnostic to the source programming language, and has instructions for Go, Java, Python, Ruby, node.js, Rust, and .NET.

My understanding is that you're searching for a combination of the profiles, metrics, and tracing. Caddy supports all 3.

[0] https://github.com/google/pprof/blob/main/doc/README.md

[1] https://grafana.com/docs/pyroscope/latest/

metrics and tracing need to be manually enabled (for now, perhaps)

Because we couldn't identify the issue using the results we got from Callgrind, we reached for another profiler, gperftools. It's a sampling profiler and therefor it has a smaller impact on the application's performance in exchange for less accurate call statistics. After filtering out the unimportant parts and visualizing the rest with pprof, it was evident that something strange was happening with the send function. It took only 71 milliseconds with the previous implementation and more than 900 milliseconds with the new implementation of our Bolt server. It was very suspicious, but based on Callgrind, its cost was almost the same as before. We were confused as the two results seemed to conflict with each other.

github.com - google/pprof

The application performances in a specific code path (e.g. gdb, pprof, …).

pprof is https://github.com/google/pprof, it's a very useful tool in golang , and really really really convenient

Go also has pretty good out of the box profiling (pprof[0]) and third-party runtime debugging (delv[1]) that can be used both remotely and local.

These tools also have decent editor integration and can be use hand in hand:

https://blog.jetbrains.com/go/2019/04/03/profiling-go-applic...

https://blog.jetbrains.com/go/2020/03/03/how-to-find-gorouti...

[0] https://github.com/google/pprof

[1] https://github.com/go-delve/delve

And what we do in Google Cloud is that we still use the pprof. But it's a kind of forked version of the pprof because the visualization part is totally different. So we give that tool as the Cloud Profiler. So that is the product name. And then, the difference between the pprof and a Cloud Profiler is that Cloud Profiler provides the agent library for each famous programming language such as Java, Python, Node.js, and Go. And then what you need to do is to just write 5 to 10 lines of code in a new application. That launches the profile agent in your application as a subsidiary thread of the main thread. And then, that thread periodically collects the profile data of the application and then sends that data back to Google Cloud and the Cloud Profiler.

gprftools (https://github.com/gperftools/gperftools) can be easily plugged in using LD_PRELOAD and signal, and has nice go implemented visualization tool https://github.com/google/pprof.