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

[Alexandre Montplaisir <alexmonthy@voxpopuli.im>]

On 2015-02-09 11:55 PM, Wang Nan wrote:
> On 2015/2/10 4:13, Alexandre Montplaisir wrote:
>> [...]
>>
>> So in this example we could see right away that the average duration is around 10-15 ms, but we have some outliers that took more than 20 ms, so it would be interesting to zoom in there to get more details. And eventually clicking on one particular occurrence could bring you to the exact point in the trace where it happened (or create bookmarks for the outlier entries, etc.)
>>
> Looks good. It should be better if users are able to define their own matching rules.

Absolutely! We can have pre-defined rules for the known domain (system 
calls, IRQs, and so on), but then it'd be easy to let the user enter an 
event name for the "start" and an event name for the "end".

I was also thinking, we could even use general filters to define what is 
a start and an end event. It would require some refactoring due to the 
way filters are organized at the moment, so it wouldn't be trivial, but 
eventually we may want to get there.

I'm currently working on the side on a prototype of the "event matching" 
table view, I will let you know once I have an initial version working, 
so we can see if/how useful it is.

>
> In addition, do you think about chaining more than 2 events together? Some procedures, such as IO and network, can be traced at different layers, and we are able to chain events caused by same sources together with some cules. Please refer to:
>
> http://comments.gmane.org/gmane.linux.kernel/1883569
>
> Which we are working on.
>

Geneviève (added in CC) is currently working on virtual machine 
analyses, where we can match events from a kernel trace on a host to a 
kernel trace in a guest, and so on. The "layering" and general matching 
of events should be similar. Maybe she has some ideas if the event 
matching part could be reused somehow?


Cheers,
Alexandre