Stefan Hajnoczi wrote: > On Fri, May 21, 2010 at 5:52 PM, Jan Kiszka <jan.kiszka@xxxxxxxxxxx> wrote: >> I would just like to avoid that too much efforts are spent on >> re-inventing smart trace buffers, trace daemons, or trace visualization >> tools. Then better pick up some semi-perfect approach (e.g. [1], it >> unfortunately still seems to lack kernel integration) and drive it >> according to our needs. > > I agree we have to consider existing solutions. The killer is the > usability: what dependencies are required to build with tracing? Is a > patched kernel or module required? How easy is it to add static trace > events during debugging? > > If there are too many dependencies, especially to unpackaged software, > many people will stop right there and not bother. A patched kernel or > module isn't acceptable since the hassle of reconfiguring a system for > tracing becomes too great (or in some cases changing the kernel is not > possible/allowed). > > Adding new static trace events should be easy, too. Ideally it > doesn't require adding information about the trace event in multiple > places (header files, C files, etc). It also shouldn't require > learning about the tracing system, adding a trace event should be > self-explanatory so anyone can easily add one for debugging. > > A lot of opinions there, but what I'm saying is that friction must be > low. If the tracing system is a pain to use, then no-one will use it. No question. I mentioned LTTng as it is most promising /wrt performance (both when enabled and disabled). But LTTng was so far not best in class when it came to usability. > > http://lttng.org/files/ust/manual/ust.html > > LTTng Userspace Tracer looks interesting - no kernel support required > AFAICT. Toggling trace events in a running process supported. > Similar to kernel tracepoint.h and existing report/visualization tool. > > x86 (32- and 64-bit) only. Sure? I thought there might be an arch dependency due to urcu but it has generic support as well now. > Like you say, no correlation with kernel trace data. It would be good if we could still hook into trancepoints and stream them out differently. That would allow for add-hoc tracing when performance does not matter that much (trace to file, trace to kernel). But we would still benefit from enabling tracepoints during runtime and keeping them built in. Jan
Attachment:
signature.asc
Description: OpenPGP digital signature