ovni/doc/user/runtime/tracing.md

3.8 KiB
Raw Blame History

Tracing a new program

Read carefully this document before using libovni to instrument a new component. There are a few rules you must follow to ensure the runtime trace is correct.

Trace processes and threads

  • Call ovni_version_check() once before calling any ovni function.

  • Call ovni_proc_init() when a new process begins the execution.

  • Call ovni_thread_init() when a new thread begins the execution (including the main process thread).

  • Call ovni_thread_require() with the required model version before emitting events for that model.

  • Call ovni_flush() and ovni_thread_free() when it finishes (in that order).

  • Call ovni_proc_fini() when a process ends, after all threads have finished.

You can use ovni_ev_emit() to record a new event. If you need more than 16 bytes of payload, use ovni_ev_jumbo_emit(). See the trace specification for more details.

Compile and link with libovni. When you run your program, a new directory ovni will be created in the current directory $PWD/ovni which contains the execution trace.

You can change the trace directory by defining the OVNI_TRACEDIR environment variable. The envar accepts a trace directory name, a relative path to the trace directory, or its absolute path. In the first case, the trace directory will be created in the current directory $PWD.

Rules

Follow these rules to avoid losing events:

  1. No event may be emitted until the process is initialized with ovni_proc_init() and the thread with ovni_thread_init().

  2. When a thread ends the execution, it must call ovni_flush() to write the events in the buffer to disk.

  3. All threads must have flushed its buffers before calling ovni_proc_fini().

Select a fast directory

During the execution of your program, a per-thread buffer is kept where the new events are being recorded. When this buffer is full, it is written to disk and emptied, an operation known as flush. This may take a while depending on the underliying filesystem.

Keep in mind that the thread will be blocked until the flush ends, so if your filesystem is slow it would interrupt the execution of your program for a long time. It is advisable to use the fastest filesystem available (see the tmpfs(5) and df(1) manual pages).

You can select the trace directory where the buffers will be flushed during the execution by setting the environment variable OVNI_TMPDIR. The last directory will be created if doesn't exist. In that case, as soon as a process calls ovni_proc_fini(), the traces of all its threads will be moved to the final directory at $PWD/ovni. Example:

OVNI_TMPDIR=$(mktemp -u /dev/shm/ovni.XXXXXX) srun ./your-app

To test the different filesystem speeds, you can use hyperfine and dd. Take a closer look at the max time:

$ hyperfine 'dd if=/dev/zero of=/gpfs/projects/bsc15/bsc15557/kk bs=2M count=10'
Benchmark 1: dd if=/dev/zero of=/gpfs/projects/bsc15/bsc15557/kk bs=2M count=10
  Time (mean ± σ):      71.7 ms ± 130.4 ms    [User: 0.8 ms, System: 10.2 ms]
  Range (min … max):    14.7 ms … 1113.2 ms    162 runs
 
  Warning: Statistical outliers were detected. Consider re-running this
  benchmark on a quiet PC without any interferences from other programs. It
  might help to use the '--warmup' or '--prepare' options.

$ hyperfine 'dd if=/dev/zero of=/tmp/kk bs=2M count=10'
Benchmark 1: dd if=/dev/zero of=/tmp/kk bs=2M count=10
  Time (mean ± σ):      56.2 ms ±   5.7 ms    [User: 0.6 ms, System: 14.8 ms]
  Range (min … max):    45.8 ms …  77.8 ms    63 runs
 
$ hyperfine 'dd if=/dev/zero of=/dev/shm/kk bs=2M count=10'
Benchmark 1: dd if=/dev/zero of=/dev/shm/kk bs=2M count=10
  Time (mean ± σ):      11.4 ms ±   0.4 ms    [User: 0.5 ms, System: 11.1 ms]
  Range (min … max):     9.7 ms …  12.5 ms    269 runs