[lttng-dev] On lost events

[Mathieu Desnoyers]

* Matthew Khouzam (matthew.khouzam at ericsson.com) wrote:
> Some streams have overlapping packets. I was shocked.

Ah! yes. The time stamp begin is sampled at event space reservation, and
time stamp end is sampled at commit.

So you can have, with packets A and B in a stream:

- reserve space at end of A
- interrupted
- reserve space at beginning of B (get timestamp begin)
- commit at end of A (get timestamp end)
- commit at beginning of B

So yes, the timestamps can very slightly overlap indeed.

Thanks,

Mathieu

> 
> 
> On 13-07-05 06:13 PM, Mathieu Desnoyers wrote:
> > * Matthew Khouzam (matthew.khouzam at ericsson.com) wrote:
> >> Quick question. Some traces have overlapping packets, so in that case,
> >> the lost events are between the last event of a previous packet and the
> >> first event of the next packet, which just happens to be the previous
> >> packet's end time?
> > "Some traces have overlapping packets"
> >
> > do you mean
> >
> > "Some traces have overlapping packets"
> >
> > or
> >
> > "Some streams have overlapping packets"
> >
> > ?
> >
> >> On 13-07-04 06:29 PM, Mathieu Desnoyers wrote:
> >>> * Matthew Khouzam (matthew.khouzam at ericsson.com) wrote:
> >>>> Hi tracing werewolves,
> >>>>
> >>>> I want to clear up some stuff about lost events.
> >>>>
> >>>> Lost events are caused, in lttng 2.x when the tracer does not have
> >>>> enough memory available to write an event. This can be due to:
> >>>> * events coming in too fast (faster than we can write them)
> >>>> * a given event being too large
> >>>> * we have an event that does a nested wrap around
> >>>>
> >>>> LTTng will not write the whole packet that is lost.
> >>> LTTng does not write an _event_ discarded at all. However, a packet
> >>> "lost" is a different thing: it only happens in overwrite mode, when we
> >>> overwrite a packet.
> >>>
> >>>> It uses a ring
> >>>> buffer and the packet will just be overwritten in the buffer, but a
> >>>> counter will be incremented.
> >>> For an _event_ discarded, we're not even writing it into the buffer.
> >>>
> >>>> The counter is written in the packet header in the CTF trace.
> >>> Indeed, the event lost counters are written in the packet in the ctf
> >>> trace.
> >>>
> >>>> When reading a trace, if you just read the packet headers, you will see
> >>>> something as follows:
> >>>>
> >>>> Packet 1:
> >>>> ts_begin: 100
> >>>> ts_end: 200
> >>>> discarded: 0
> >>>>
> >>>> Packet 2:
> >>>> ts_begin: 300
> >>>> ts_end: 400
> >>>> discarded: 100
> >>>>
> >>>> Packet 3:
> >>>> ts_begin: 500
> >>>> ts_end: 600
> >>>> discarded: 100
> >>>>
> >>>> Packet 4:
> >>>> ts_begin: 650
> >>>> ts_end: 750
> >>>> discarded: 200
> >>>>
> >>>> With this example, the lost events are located between packets 1, 2 and
> >>>> packets 3,4.
> >>>>
> >>>> If the events are too large, they are discarded in the current packet,
> >>>> lttng does not split events into packets (yet?), and this is lost during
> >>>> a current packet.
> >>>>
> >>>> That means if the events are lost between 1 and 2, they are potentially
> >>>> in the range of [200-400] but probably in [200-300]
> >>>> The next lost event are between 3 and 4 (events discarded is a total
> >>>> count, not local to that packet), so 100 events are also lost between
> >>>> [600-750] but likely in between [600-650]. If you can confirm that all
> >>>> events WITH CONTEXTS are smaller than the packet sizes, I think it's a
> >>>> shoe in that it's between the packets and not containing the next packet.
> >>> "and not contained in the next packet".
> >>>
> >>> Thanks,
> >>>
> >>> Mathieu
> >>>
> >>>> Hope this clarifies some issues, it did for me!
> >>>>
> >>>> Matthew
> >>>>
> >>>> _______________________________________________
> >>>> lttng-dev mailing list
> >>>> lttng-dev at lists.lttng.org
> >>>> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com