[lttng-dev] difference of timestamp between babeltrace and python binding

[Philippe Proulx]

On Wed, Jul 22, 2020 at 9:17 AM Kim, Seongab via lttng-dev
<lttng-dev at lists.lttng.org> wrote:
>
> Hi,

Hello.

See my comments below.

>
> I found that the timestamps from babeltrace and python binding are different in nanoseconds level.
>
> skim at d54030999178:~/ssd_work/traces$ python3 test.py ./trace/ | head -n 10
> 2019-12-12 13:45:40.098212096 (+0.000000 s): syscall_exit_ppoll
> 2019-12-12 13:45:40.098213608 (+0.000001 s): rcu_utilization
> 2019-12-12 13:45:40.098214632 (+0.000001 s): syscall_entry_ioctl
> 2019-12-12 13:45:40.098214888 (+0.000000 s): kmem_kfree
> 2019-12-12 13:45:40.098214144 (+0.000000 s): rcu_utilization
> 2019-12-12 13:45:40.098215168 (+0.000001 s): sched_stat_runtime
> 2019-12-12 13:45:40.098220776 (+0.000004 s): sched_switch
> 2019-12-12 13:45:40.098221312 (+0.000002 s): timer_hrtimer_cancel
> 2019-12-12 13:45:40.098221312 (+0.000000 s): kmem_kmalloc
> 2019-12-12 13:45:40.098223360 (+0.000002 s): kmem_kfree
> Traceback (most recent call last):
>   File "test.py", line 33, in <module>
>     print(fmt.format(dt, ns_from_origin % 1e3, diff_s, msg.event.name))
> BrokenPipeError: [Errno 32] Broken pipe
> skim at d54030999178:~/ssd_work/traces$ babeltrace2 ./trace/ | head -n 10
> [13:45:40.098211969] (+?.?????????) test-sys syscall_exit_ppoll: { cpu_id = 3 }, { ret = 0, nfds = 1, fds_length = 0, overflow = 0, fds =[ ] }
> [13:45:40.098212481] (+0.000000512) test-sys rcu_utilization: { cpu_id = 5 }, { s = "Start context switch" }
> [13:45:40.098213729] (+0.000001248) test-sys syscall_entry_ioctl: { cpu_id = 2 }, { fd = 24, cmd = 1075331408, arg = 547894601712 }
> [13:45:40.098213953] (+0.000000224) test-sys kmem_kfree: { cpu_id = 3 }, { call_site = 0xFFFFFF800153B274, ptr = 0x0 }
> [13:45:40.098214017] (+0.000000064) test-sys rcu_utilization: { cpu_id = 5 }, { s = "End context switch" }
> [13:45:40.098215201] (+0.000001184) test-sys sched_stat_runtime: { cpu_id = 5 }, { comm = "Positioning", tid = 1056, runtime = 38048, vruntime = 2125241692 }
> [13:45:40.098219681] (+0.000004480) test-sys sched_switch: { cpu_id = 5 }, { prev_comm = "Positioning", prev_tid = 1056, prev_prio = 15, prev_state = 1, next_comm = "Sensor", next_tid = 1061, next_prio = 15 }
> [13:45:40.098221313] (+0.000001632) test-sys timer_hrtimer_cancel: { cpu_id = 5 }, { hrtimer = 0xFFFFFFC257807CD8 }
> [13:45:40.098221409] (+0.000000096) test-sys kmem_kmalloc: { cpu_id = 0 }, { call_site = 0xFFFFFF80014A27EC, ptr = 0xFFFFFFC04EDFA400, bytes_req = 739, bytes_alloc = 1024, gfp_flags = 37748928 }
> [13:45:40.098223361] (+0.000001952) test-sys kmem_kfree: { cpu_id = 5 }, { call_site = 0xFFFFFF8008164C58, ptr = 0x0 }
>
> The difference is not consistent and Trace Compass shows the same timestamp with the babeltrace output, so I think something is wrong in my python code.
> I just modified the example script from https://babeltrace.org/docs/v2.0/python/bt2/examples.html#get-an-event-s-time to show nanoseconds as below.
>
> diff:
>
> 32,33c32,33
> <         fmt = '{} (+{:.6f} s): {}'
> <         print(fmt.format(dt, diff_s, msg.event.name))
> ---
> >         fmt = '{}{:03.0f} (+{:.6f} s): {}'
> >         print(fmt.format(dt, ns_from_origin % 1e3, diff_s, msg.event.name))
>
> Full code:
>
> import bt2
> import sys
> import datetime
>
> # Create a trace collection message iterator from the first command-line
> # argument.
> msg_it = bt2.TraceCollectionMessageIterator(sys.argv[1])
>
> # Last event's time (ns from origin).
> last_event_ns_from_origin = None
>
> # Iterate the trace messages.
> for msg in msg_it:
>     # `bt2._EventMessageConst` is the Python type of an event message.
>     if type(msg) is bt2._EventMessageConst:
>         # Get event message's default clock snapshot's ns from origin
>         # value.
>         ns_from_origin = msg.default_clock_snapshot.ns_from_origin
>
>         # Compute the time difference since the last event message.
>         diff_s = 0
>
>         if last_event_ns_from_origin is not None:
>             diff_s = (ns_from_origin - last_event_ns_from_origin) / 1e9
>
>         # Create a `datetime.datetime` object from `ns_from_origin` for
>         # presentation. Note that such an object is less accurate than
>         # `ns_from_origin` as it holds microseconds, not nanoseconds.
>         dt = datetime.datetime.fromtimestamp(ns_from_origin / 1e9)

This rounds the µs part.

With 1509556783964792907, the µs part becomes 964793.

That's one issue.

>
>         # Print line.
>         fmt = '{}{:03.0f} (+{:.6f} s): {}'
>         print(fmt.format(dt, ns_from_origin % 1e3, diff_s, msg.event.name))

`type(1e3)` is `float`.

You get an inaccurate floating point division operation here instead of
an integral one.

Do:

    ns_from_origin % 1000

or

    ns_from_origin % int(1e3)

instead.

Hope this helps,

Phil

>
>         # Update last event's time.
>         last_event_ns_from_origin = ns_from_origin
>
> Could someone help what was wrong in the python code? Or is this a bug in ns_from_origin implementation?