[lttng-dev] Babeltrace performance issue in live-reading mode

[Jonathan Rajotte Julien]

Hi,

On 2017-09-20 05:12 AM, liguang li wrote:
> 
> 
> On Tue, Sep 19, 2017 at 10:57 PM, Jonathan Rajotte-Julien <jonathan.rajotte-julien at efficios.com <mailto:jonathan.rajotte-julien at efficios.com>> wrote:
> 
>     On Tue, Sep 19, 2017 at 03:53:27PM +0800, liguang li wrote:
>     >    Hi,
>     >    On Mon, Sep 18, 2017 at 11:18 PM, Jonathan Rajotte-Julien
>     >    <[1]jonathan.rajotte-julien at efficios.com <mailto:jonathan.rajotte-julien at efficios.com>> wrote:
>     >
>     >      Hi,
>     >      On Mon, Sep 18, 2017 at 11:32:07AM +0800, liguang li wrote:
>     >      >    Hi,
>     >      >
>     >      >    Create a session in live-reading mode, run a application which
>     >      having very
>     >      >    high event throughput, then prints
>     >      >    the events with babeltrace. We found the live trace viewer are
>     >      viewing
>     >      >    events a few seconds ago, and as time
>     >
>     >      Could you provide us the version used for babeltrace, lttng-tools and
>     >      lttng-ust?
>     >
>     >     Babeltrace: 1.5.1
> 
>     Update to babeltrace 1.5.3.
> 
>     >     Lttng-tools: 2.8.6
> 
>     Update to lttng-tools 2.8.8
> 
>     >     Lttng-ust: 2.8.2
>     >
>     >      >    goes on, the delay will be bigger and bigger.
>     >
>     >      A similar issues was observed a couple months ago, which implicated
>     >      multiple delayed ack
>     >      problems during communication between lttng-relayd and babeltrace.
>     >
>     >      The following fixes were merged:
>     >
>     >      [1]
>     >      [2]https://github.com/lttng/lttng-tools/commit/b6025e9476332b75eb8184345c3eb3e924780088 <https://github.com/lttng/lttng-tools/commit/b6025e9476332b75eb8184345c3eb3e924780088>
>     >      [2]
>     >      [3]https://github.com/efficios/babeltrace/commit/de417d04317ca3bc30f59685a9d19de670e4b11d <https://github.com/efficios/babeltrace/commit/de417d04317ca3bc30f59685a9d19de670e4b11d>
>     >      [3]
>     >      [4]https://github.com/efficios/babeltrace/commit/4594dbd8f7c2af2446a3e310bee74ba4a2e9d648 <https://github.com/efficios/babeltrace/commit/4594dbd8f7c2af2446a3e310bee74ba4a2e9d648>
>     >
>     >      In the event that you are already using an updated version of babeltrace
>     >      and
>     >      lttng-tools, it would be pertinent to provide us with a simple
>     >      reproducer so we
>     >      can assess the issue.
> 
>     The version you are using does not include the mentioned fixes. Please update
>     and redo your experiment.
> 
> 
> Test this issue in the version you have listed, the issue still exists.

Given that previous versions had a major timing problem there I would expect to have some improvement.

In that case, we will need a lot more information on your benchmarking strategy.
We will need a simple reproducer, your benchmark code (r,gnuplot etc.), your overall methodology
to be able to reproduce the issue locally. Otherwise, it will be very hard 
to come to any conclusion.

>  
> 
>     Cheers
> 
>     >
>     >     
>     >     Steps:
>     >     lttng create session --live -U net://*
>     >     lttng enable-channel -s session -u ch1
>     >     lttng enable-event -s session -c ch1 -u -a
>     >     lttng start
>     >     
>     >     Run a high event throughput application, which is multithreaded
>     >    application> 
> 
> In the multithreaded application, each tracepoint will have the wall
> time of the system,then we can easily reproduce this issue through
> comparing the time of recorded event and the system wall time.
>  
> 
>     >     babeltrace -i lttng-live net://*
>     >     
>     >     After a while, we found the timestamp of the event in the babeltrace is
>     >    different with the time in host
>     >     which run the application. And the delay will be bigger and bigger with
>     >    time goes.
>     >     
>     >
>     >      Cheers
>     >      >    I checked the source code, found Babeltrace in live-reading mode
>     >      will read
>     >      >    the recorded events in the CTF
>     >      >    files, and then parse and print it in a single thread. The process
>     >      is a
>     >      >    little slow, do you have any ideas to
>     >      >    improve the process.
> 
> 
> From my understanding of the source code, the process of parse and
> print event will consume a lot of time. For example, the multithreaded
> application will consume 3 CPUs, in a specified time,3 subbuffers will
> be filled and sent to lttng-relayd daemon, recorded into the CTF files.
> If in the specified time, babeltrace only handled 2 subbuffers' event,
> thenthe issue will happens.

Did you perform a bisection for where the delay come from? Reception of packet? formatting of event?
What is the throughput of the application?
How many tracepoint definition?
Does babeltrace catch up if a quiescent period is given?
Could you provide us with statistics, timing data, etc.?
What type of delay are we talking about?


-- 
Jonathan R. Julien
Efficios