Implement High-Resolution Timer Using clock_gettime

Topic Overview

Nanosecond Timing in Userspace

clock_gettime(CLOCK_MONOTONIC) is what glibc, LLVM's benchmark harness, and the Linux kernel's own microbenchmarks use for sub-millisecond timing. Unlike gettimeofday, monotonic clocks never jump backward when NTP adjusts wall time. Pairing start and end reads around a tight loop is the foundation of every custom benchmark harness.

Building a Small Timer API

Read timespec at region start and end, subtract, convert to nanoseconds or microseconds. Warm up caches before the timed region so cold-start effects do not dominate.

cCopy
Loading…

• Never time a single iteration of a nanosecond-scale loop; batch many iterations
• Pin thread to a core when comparing micro-optimizations on laptops
• CLOCK_PROCESS_CPUTIME_ID measures CPU time, not sleep time
• Print both mean and min across several runs

Working Through the Exercise

Keep the relevant documentation open while you implement. When your output disagrees with the reference, trace one failing case by hand before changing random lines.

• Trace first: walk one failing input step by step
• Read the manual: skim the official doc for the mechanism named in this task
• Compare baselines: save measurements from the prior step so speedups are honest

Why for this exercise

You will implement now_ns() and wrap a benchmark region with warmup plus repeated samples. This exercise asks you to report elapsed time for a known workload and verify results are stable across runs.

Your Task

A home-made timer built on clock_gettime has three classic bugs. The first is getting the timespec subtraction wrong when the nanosecond field wraps (you must borrow a second). The second is using CLOCK_REALTIME, which can jump backwards when NTP adjusts the clock. The third is timing something so short that the overhead of reading the clock dominates. Given recorded timespec pairs, compute intervals correctly, detect those problems, and derive per-operation costs.

Input

cCopy
Loading…

Rules

• Validation. A timespec is valid if tv_sec >= 0 and 0 <= tv_nsec <= 999999999, otherwise print NAME: invalid timespec (tv_nsec must be 0..999999999).
• Subtraction. sec = S2 - S1, nsec = NS2 - NS1; if nsec < 0, borrow one second (sec--, nsec += 1000000000).
• Negative interval:
  • on realtime: clock went backwards by X (wall-clock adjustment: use CLOCK_MONOTONIC);
  • on monotonic: end is before start (a monotonic clock never goes back: check the order).
• Human-readable durations: N ns below 1 µs, N.NNN us below 1 ms, N.NNN ms below 1 s, and N.NNNNNNNNN s otherwise (truncated).
• Calibration. The overhead is the minimum of the non-negative deltas (the mean is shown too, rounded half up). After calibrating, each interval also prints the value minus the overhead (not below 0), with a warning if the interval is under 100 × the overhead. Intervals remember this corrected value.
• per NAME N: the stored (corrected) nanoseconds / N, 2 decimals, rounded half up. The most recent interval with that name is used.
• Errors: clock: monotonic or realtime, calibrate: need back-to-back deltas, per NAME: no such interval, and per NAME: need at least 1 operation.

Output

cCopy
Loading…

Example

Input:

cCopy
Loading…

Output:

cCopy
Loading…

Requirements

1. Subtract in two fields with an explicit borrow. Do not convert to double, because a double cannot hold nanoseconds since the epoch exactly.
2. Use the minimum back-to-back delta as the timer overhead. Larger values include interrupts and noise.

Hidden tests cover borrows at exact second boundaries, a long interval of hours, intervals before and after calibration, a zero-length interval, per-operation costs below one nanosecond, and invalid input.