The Fastest Linux Timestamps

As the cost of stamping the start and end times of an OpenTelemetry span rose to 46-49ns, timestamps themselves turned out to be a non-negligible overhead in ultra-low-latency pipelines.

The core background is x86's invariant TSC and how Linux's vDSO works. A typical clock_gettime() is handled via the vDSO without entering the kernel, but even so it involves checking a sequence lock, reading the TSC, converting cycles to nanoseconds, and normalizing to seconds.

The article dissects the vDSO's do_hres() path, explaining the structure where the kernel updates a reference time and correction values in a data page, and userspace reads it to compute the current time. It notes that this process requires lfence + rdtsc or rdtscp, and that reading the TSC itself is not completely free either.

It then proposes a simpler path suited for tracing purposes.

• Since a monotonic clock only needs a difference, not an absolute time, the cycles difference can be converted directly to nanoseconds.
• Instead of reading the wall clock and monotonic clock separately twice, redundant computation can be eliminated by computing only the values that are needed.
• This approach cut timestamp processing time by 30% on x86 Linux, while maintaining the same precision as the standard system clock.

However, the author clearly states that this optimization is overkill for most programs. It's a technique that only matters for special cases like tracing in ultra-low-latency environments.