Tracing CLN Performance

CLN includes a simple opentracing exporter that allows tracing the execution of the node in real-time, without incurring a performance penalty when not listening for traces. Quoting the Wikipedia entry on Tracing:

In software engineering, tracing involves a specialized use of logging to record information about a program's execution. This information is typically used by programmers for debugging purposes, and additionally, depending on the type and detail of information contained in a trace log, by experienced system administrators or technical-support personnel and by software monitoring tools to diagnose common problems with software.

The tracing system in CLN is implemented using USDTs (no, not that kind of USDT). As such it emits events into the kernel, from where an exporter can receive them. If no exporter is configured then the kernel will replace the call-sites of the probe with a NOP, thus causing only minimal overhead when not tracing.

Compiling with tracing support

CLN will build with tracing support if the necessary headers (sys/sdt.h) are present during the compilation. For debian and ubuntu based systems that is easily achieved by installing systemtap-sdt-dev:

# apt-get install -y systemtap-sdt-dev

Don't forget to run ./configure and make to recompile after installing the dependencies. config.vars should contain the following line after running ./configure:

HAVE_USDT=1

If you have a compiled binary you can verify whether it was compiled with USDT support with the following command:

$ readelf -S lightningd/lightningd | grep -i sdt

Alternatively you can list the tracepoints in the binary with the following:

$ sudo bpftrace -l 'U:lightningd/lightningd:*'
usdt:lightningd/lightningd:lightningd:span_emit
usdt:lightningd/lightningd:lightningd:span_end
usdt:lightningd/lightningd:lightningd:span_resume
usdt:lightningd/lightningd:lightningd:span_start
usdt:lightningd/lightningd:lightningd:span_suspend

Exporters

The simplest way to get started with eBPF in general (which the tracing is built upon) is the bpftrace command that we've already seen above when checking if the binary was built with tracing support.

$ sudo bpftrace -l 'U:lightningd/lightningd:*'
usdt:lightningd/lightningd:lightningd:span_emit
usdt:lightningd/lightningd:lightningd:span_end
usdt:lightningd/lightningd:lightningd:span_resume
usdt:lightningd/lightningd:lightningd:span_start
usdt:lightningd/lightningd:lightningd:span_suspend

There is a sample exporter that can be used to instrument a single
binary, batch the spans it receives and submit them as a batch to an
otelcol or tempo instance in contrib/cln-tracer
using the zipkin format for spans and traces.

Notice that due to a [limitation][bpftracer305] in the way the eBPF
script is handled you'll at most get the first 495 bytes of the
payload. This is due to the 512 byte limitation for eBPF programs out
of the box.

Tracing Overhead

While we try to make tracing as lightweight as possible it isn't
free. To quantify how much time is spent actually maintaining the
tracing context we built a small test in common/test/run-trace which
creates a sequence of 25'000 traces, with a total of 1'000'000 spans,
without any additional operations.

The following run with [hyperfine][hf] shows the runtime for that test
without an exporter attached:


$ hyperfine common/test/run-trace
Benchmark 1: common/test/run-trace
  Time (mean ± σ):     368.4 ms ±   4.8 ms    [User: 142.7 ms, System: 225.1 ms]
  Range (min … max):   358.3 ms … 377.2 ms    10 runs

While the following is the same run, but with an exporter attached:

$ ❯ hyperfine common/test/run-trace
Benchmark 1: common/test/run-trace
  Time (mean ± σ):     561.5 ms ±  14.1 ms    [User: 164.5 ms, System: 396.5 ms]
  Range (min … max):   546.5 ms … 598.9 ms    10 runs

So depending on whether an exporter is attached, creating and emitting
span without and with an exporter takes around 370ns and 560ns
respectively.