95 lines
3.8 KiB
Markdown
95 lines
3.8 KiB
Markdown
# 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](../trace_spec) 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
|
||
```
|
||
|