[lttng-dev] [diamon-discuss] My experience on perf, CTF and TraceCompass, and some suggection.

[Alexandre Montplaisir]

Hi Wang,

First of all, thank you very much for posting this use case. This is 
exactly the type of user feedback that will help make the toolchain 
better and more useful for users!

Some comments and questions below,


On 01/23/2015 04:35 AM, Wang Nan wrote:
> [...]
>
> Then I need to convert perf.data to ctf. It tooks 140.57s to convert
> 2598513 samples, which are collected during only 1 second execution. My
> working server has 64 2.0GHz Intel Xeon cores, but perf conversion
> utilizes only 1 of them. I think this is another thing can be improved.

Out of curiosity, approximately how big (in bytes) is the generated CTF 
trace directory?

>
> The next step is visualization. Output ctf trace can be opened with
> TraceCompass without problem. The most important views for me should be
> resources view (I use them to check CPU usage) and control flow view (I
> use them to check thread activities).
>
> The first uncomfortable thing is TraceCompass' slow response time. For
> the trace I mentioned above, on resource view, after I click on CPU
> idle area, I have to wait more than 10 seconds for event list updating
> to get the previous event before the idle area.

Interesting. It is expected that opening a very large trace would take a 
long time to load the first time, as everything gets indexed. But once 
that step is done, seeking within the trace should be relatively quick 
((log n) wrt to the trace size). In theory ;)

The perf-to-CTF conversion brings a completely new type of CTF traces 
that was not seen before. It is possible that the CTF parser in Trace 
Compass has some inefficiencies that were not exposed by other trace 
types. Are you able to share that trace publicly? Or a trace taken in 
the same environment, with no sensible information in it? It could be 
very helpful in finding such problem.

> Then I found through resources view that perf itself tooks lots of CPU
> time. In my case 33.5% samples are generated by perf itself. One core is
> dedicated to perf and never idle or taken by others. I think this should
> be another thing needs to be improved: perf should give a way to
> blacklist itself when tracing all CPUs.

I don't want to start a tracer-war here :) but have you investigated 
using LTTng for recording syscall/sched events ? Compared to perf, LTTng 
is only about "getting trace events", and is a bit more involved to set 
up, but it is more focused on performance and minimizing the impact on 
the traced applications. And it outputs in CTF format too.

I remember when testing the perf-CTF patches, comparing a perf trace to 
an LTTng one, perf would be doing system calls continuously on one of 
the CPUs for the whole duration of the trace. Whereas in LTTng traces, 
the session daemon would be a bit active at the beginning and at then 
end, but otherwise completely invisible from the trace.

> TraceCompass doesn't recognize syscall:* tracepoints as CPU status
> changing point. I have to also catch raw_syscall:*, and which doubles
> the number of samples.

This is a gap in the definition of the analysis it seems. I don't 
remember implementing two types of "syscall" events in the perf 
analysis, so it should just be a matter of getting the exact event name 
and adding it to the list. I will take a look and keep you posted!

> Finally I found the syscall which cause idle. However I need to write a
> script to do statistics. TraceCompass itself is lack a mean to count
> different events in my way.

Could you elaborate on this please? I agree the "Statistics" view in TC 
is severely lacking, we could be gathering and displaying much more 
information. The only question is what information would actually be useful.

What exactly would you have liked to be able to see in the tool?

> [...]
>
>
>   5. Ad-Hoc visualization and statistics. Currently TraceCompass only
>      support dwaring pre-defined events and processes. When I try to
>      capture syscalls:*, I won't get benefit from TraceCompass because it
>      doesn't know them. I believe that during system tuning we will
>      finally get somewhere unable to be pre-defined by TraceCompass
>      designer. Therefore give users abilities to define their own events
>      and model should be much helpful.

As I mentioned earlier, the pre-defined "perf analysis" in Trace Compass 
should be fixed to handle the syscall events.


But it's interesting that you mention wanting to add your own events and 
model. I completely agree with you, we will never be able to predict 
every and all use cases the users will want to use the tool for, so 
there should be a way for the user to add their own.

Well good news, it *is* possible for the user to define their own 
analysis and views! This is still undergoing a lot of development, and 
there is no nice UI yet, which is why it is not really advertized. But 
starting from any supported trace type, a user today can define a time 
graph view (like the Resource View for example) or a XY chart, using a 
data-driven XML syntax.

If you are curious, you can take a look at a full example of doing such 
a thing on this page:
https://github.com/alexmonthy/ust-tc-example
(the example uses an LTTng UST trace as a source, but it could work with 
any supported trace type, even a custom text trace defined in the UI).

>
> Thank you.

Thanks again for taking the time to write about your experience!

Cheers,
Alexandre