Re: [tracecompass-dev] How to use Trace Compass with perf record/trace

[Matthew Khouzam <matthew.khouzam@xxxxxxxxxxxx>]

Hi,

I have to warn, I am far from an expert on Perf, to know who to get it
working, you're better off asking on the LKML.

Is your perf record command working?
(https://lists.gt.net/linux/kernel/1972041)

The CTF output will give raw events, the aggregate data is generated by
trace compass itself.

I hope this helps.

Matthew.


On 17-03-02 11:50 AM, Stefano Doni wrote:
> Hi,
>
>
> I'm to to Trace Compass and I'm trying to use it to analyze traces
> collected on standard Linux kernel using perf. I'm particularly
> interested in visualizing the time each application thread spends in
> userspace, syscalls and when scheduled out of CPUs.
>
> The first attempt I made was using perf record and then converting to
> CTF, however it seems Ubuntu did not compile CTF data conversion into
> perf:
>
> $ sudo perf data convert
> No conversion support compiled in.
>
> Do you know if there are other distros that have included CTF
> conversion capabilities in perf packages?
>
>
> The second question is related to importing text files into Trace
> Compass. I have a number of perf trace text outputs that I collected
> from live production systems that I would like to analyze. I've seen
> that Trace Compass has the capability to import arbitrary text files,
> however it is not clear from the guide what are the steps needed to
> actually import the file and which kind of analysis are supported.
>
> The perf trace output has the following format:
>
> 5295.425 ( 2.302 ms): recvfrom(fd: 459<socket:[3282172773]>, ubuf:
> 0x7feedeff4e90, size: 8  ) = 8
>   5295.586 ( 0.034 ms): recvfrom(fd: 459<socket:[3282172773]>, ubuf:
> 0x7feedeff4e90, size: 4  ) = 4
>   5297.259 ( 0.069 ms): write(fd: 468<socket:[3282201926]>, buf:
> 0x7fedd62c1f68, count: 1954  ) = 1954
>   5329.828 (32.545 ms): read(fd: 468<socket:[3282201926]>, buf:
> 0x7feedeff0900, count: 16384  ) = 16384
>   5329.905 ( 0.018 ms): read(fd: 468<socket:[3282201926]>, buf:
> 0x7feedeff0900, count: 16384  ) = 16384
>   5329.961 ( 0.022 ms): read(fd: 468<socket:[3282201926]>, buf:
> 0x7feedeff0900, count: 16384  ) = 16384
>   5330.007 ( 0.018 ms): read(fd: 468<socket:[3282201926]>, buf:
> 0x7feedeff0900, count: 16384  ) = 16384
>   5330.055 ( 0.017 ms): read(fd: 468<socket:[3282201926]>, buf:
> 0x7feedeff0900, count: 16384  ) = 13726
>
> Interesting bits: first field is timestamp, syscall duration between
> brackets and then we have the syscall name. Here I have traced a
> single thread, but if a pid is traced we also have the thread id (tid)
> in the output.
>
>
> Is it possible to import a perf trace output and analyze it to see
> different threads over time with syscall durations, similar to the
> Control Flow analysis?
>
>
> Thanks a lot in advance!
>
>
> _______________________________________________
> tracecompass-dev mailing list
> tracecompass-dev@xxxxxxxxxxx
> To change your delivery options, retrieve your password, or unsubscribe from this list, visit
> https://dev.eclipse.org/mailman/listinfo/tracecompass-dev