| Re: [linuxtools-dev] TMF indexing performance |
Polzer, Franz wrote:
How is it planned to efficiently creating the index of trace files that are holding multiple millions of events?It seems to be quite slow to parse every single event in a trace log file just for creating the index. I see that synchronization of multiple traces might be a problem but it can handled more efficient when we have a getNextEvent function that is able to skip a given amount of events.It might work this way:The framework calls getNextEvent( traces, contexts, numberOfEventsToSkip );
For LTTng traces, it won't work. The format is designed for extreme compactness and the events have variable sizes. You cannot jump to an arbitrary event _number_. The trace is however organized in blocks whose begin and end timestamp are known. Therefore you can jump to a specific time using bisection search.
Therefore, the current widgets in TMF should not need a time index. The event list scrollbar should eventually be implemented using a time scale rather than using an event number index. Using a scrollbar to jump to a location in a trace is *so* coarse that it simply makes no sense.
And why would you need to jump to a specific time index? Can anyone come up with a use case where the trace event number is really needed?