Awesome
eio-trace
eio-trace can be used to record and display traces of programs using the Eio library (requires Eio 0.14 or later).
Installation
To install eio-trace:
opam install eio-trace
macOS
These instructions will walk you through installing the system dependencies necessary in order to install eio-trace's dependencies from OPAM.
These instructions make use of Homebrew, adapt them to MacPorts, Fink, etc., if needed.
First ensure everything is up-to-date:
brew update
brew upgrade
opam update
Then install the required packages:
brew install bash cairo gtk+3 m4 pkg-config libffi
Consider restarting your terminal session at this stage.
You can now install eio-trace's dependencies.
If installing conf-gtk3
fails with a libffi
version error, run the following before trying again:
export PKG_CONFIG_PATH=/usr/local/Cellar/libffi/3.4.4/lib/pkgconfig
Update the version number to match the currently installed version of libffi.
Tracing a program
To run an Eio program and display the trace:
eio-trace run -- myprog ...
You might like to start by tracing the example that comes with eio-trace:
dune build && eio-trace run -- ./_build/default/examples/net/main.exe
<p align='center'>
<img src="./doc/net.svg"/>
</p>
Scrolling with the mouse or touchpad will zoom in or out of the diagram.
If the program is still running when the window appears, you can press F5 to update to the latest events.
To record a trace:
eio-trace record -f trace.fxt -- myprog ...
This runs myprog ...
with the OCAML_RUNTIME_EVENTS_START
environment variable set, which causes it to record events to a ring buffer.
eio-trace saves these events to trace.fxt
.
The trace can be viewed using generic tools such as Perfetto, but eio-trace's own visualisation is more useful, as it takes advantage of Eio's structured concurrency:
eio-trace show trace.fxt
To convert a trace to SVG format:
eio-trace render trace.fxt
You can also use --format=png
for PNG output.
To calculate time spent in GC:
eio-trace gc-stats trace.fxt
./trace.fxt:
Ring GC/s App/s Total/s %GC
0 0.273 2.258 2.530 10.78
1 0.304 0.855 1.159 26.22
2 0.301 0.859 1.160 25.98
3 0.302 0.859 1.161 26.02
4 0.301 0.858 1.159 25.96
5 0.300 0.866 1.165 25.72
6 0.297 0.873 1.170 25.36
All 2.077 7.427 9.504 21.86
Note: all times are wall-clock and so include time spent blocking.
Reading traces
Eio fibers are shown as horizontal bars. Green regions show when the fiber is running. Lighter-green regions highlight particular operations. Dark grey regions indicate that the fiber is suspended (allowing other fibers to run). Only one fiber in a domain can be running at a time.
When a new fiber is created, a vertical green line links the parent fiber to the child.
Cancellation contexts (including switches) are shown as brackets.
Switches can be named using Switch.run ~name
. Unnamed switches are shown as "switch".
A context being cancelled is indicated by a vertical red line.
Domain-wide events, such as garbage collection and waiting for events, are shown as coloured regions behind all the fibers in the domain. GC periods are shown in shades of red when running, or yellow when waiting. Yellow is also used when the domain is waiting for events outside of GC. When there are multiple domains, they are shown stacked vertically:
<p align='center'> <img src="./doc/gc.svg"/> </p>In the above trace, the upper domain performed GC while suspended (the red "minor" region in the top right, inside the "suspend-domain" region). This is possible because each domain has a "backup" thread that handles GC while the domain is suspended.
For minor GCs:
- The domain initiating the GC enters a "stw_leader" (stop-the-world) phase and waits for the other domains to stop.
- One by one, the other domains stop and enter "stw_api_barrier" until all domains have stopped.
- All domains perform a minor GC, clearing their minor heaps.
- They then enter a "minor_leave_barrier" phase, waiting until all domains have finished.
- Each domain returns to running application code (including GC finalizers).
Phases that usually involve sleeping are shown with a yellow background, but sometimes they do perform work (the trace events don't give us enough information to know in all cases).
Controls
- Right click for a menu
- F5 : reload the trace file
- s : start-time (elapsed time at left edge of window since trace start)
- d : duration (duration to show in the currently visible area of the window)
d
is useful for getting two windows to use the same scale, so that they can be compared easily.
Missed events
If eio-trace does not read a domain's ring buffer quickly enough then some events will be lost and you will see something like this:
+Warning: ring 0 lost 94001 events
This will likely result in misleading traces. There are several ways to fix this:
- Use e.g.
eio-trace -F 1000
to make eio-trace read the ring more frequently. - Set e.g.
OCAMLRUNPARAM=e=20
to increase the size of the rings. - If events are being lost at startup, consider adding a short sleep to the start of your program so that eio-trace has time to attach it it.
Limitations
- OCaml 5.1 can deadlock when tracing multiple domains. This was fixed in OCaml 5.2.
- Events are reported per-domain, but not per-systhread. Events generated in systhreads will get mixed up and cannot be shown correctly. They will either appear attached to whatever fiber happens to be running, or shown as domain-level events.
- The rendering is not optimised yet and may become quite slow on larger traces.
Real-world example
The OCaml 5 Performance Problems blog post shows many ways of examining the behaviour of an Eio program using eio-trace and other tools.