[erlang-questions] profile question

[Kaiduan Xie]

Thanks a lot, Raimo for the help.

More questions,

1. From my understanding, Erlang trace is not targeted to find performance
bottleneck, it is used to help you understand how function is executed. I
can not figure out how to use trace to find which function calls the
io_lib_format:pcount in my case most. (In my case, io_lib:format is
dispersed in the code almost everywhere.)

2. For fprof, how to profile a function without specifying the arguments?
For example, I want to fprof io_lib:format, but I want to profile all calls
on io_lib:format regardless of the real arguments.

3. In C, we can use gprof to achieve what I want in 2). Although gprof
increases the system load substantially. What is the counter-part of gprof
in Erlang.

4. What is the tool you guys in Ericsson use to profile a high-load system
without degrading the performance dramatically, at the same time, we can
still get gprof like output?

Thanks,

kaiduan

On Wed, Jan 28, 2009 at 3:17 AM, Raimo Niskanen <
raimo+erlang-questions@REDACTED<raimo%2Berlang-questions@REDACTED>
> wrote:

> On Tue, Jan 27, 2009 at 11:00:10PM -0500, Kaiduan Xie wrote:
> > Hi, all,
> >
> > I have a question on how to find the hot spot of large base Erlang
> system.
> > For example, after running cprof.start(), cprof.pause(), cprof.analyze(),
> I
> > got the following results (only the first item is listed below),
> >
> > {1847012003,
> >  [{io_lib_format,722457356,
> >                  [{{io_lib_format,pcount,2},109483111},
> >                   {{io_lib_format,collect,2},109483111},
> >                   {{io_lib_format,build,3},109483111},
> >                   {{io_lib_format,iolist_to_chars,1},83312743},
> >                   {{io_lib_format,indentation,2},60014886},
> >                   {{io_lib_format,field_value,2},26618934},
> >                   {{io_lib_format,field_value,3},17824836},
> >                   {{io_lib_format,precision,2},17706522},
> >                   {{io_lib_format,pad_char,2},17706522},
> >                   {{io_lib_format,field_width,3},17706522},
> >                   {{io_lib_format,field_width,2},17706522},
> >                   {{io_lib_format,decr_pc,2},17706522},
> >                   {{io_lib_format,control,7},17706522},
> >                   {{io_lib_format,collect_cseq,2},17706522},
> >                   {{io_lib_format,collect_cc,2},17706522},
> >                   {{io_lib_format,print,7},14402937},
> >                   {{io_lib_format,chars,2},8912422},
> >                   {{io_lib_format,term,5},8912417},
> >                   {{io_lib_format,min|...},8912416},
> >                   {{io_lib_format|...},8912416},
> >                   {{...}|...},
> >                   {...}|...]},
> >
> > So here comes the question, how can we find which part calls
> > (io_lib_format,pcount,2) most? I tried to use fprof(), but it seems that
> you
> > need to know the arguments of the function first.
> >
> > What is the best practice to address this problem? Or did I do something
> > wrong?
>
> No, you are on track. This is the curse of cprof - it only gives you
> the call count but with minimal strain on your system. The curse of
> fprof is that it gives you lots of information for the price of
> a very high system load.
>
> Now you can read some code. The module io_lib_format is an internal
> module for io_lib and io. So the user code calls either io:format/1,2,3,
> io:fwrite/1,2,3, io_lib:format/2 or io_lib:fwrite/2. You can search
> for those in your source and see which are suspicious.
>
> You can also hand craft a trace by tracing only the io and io_lib functions
> in the previous paragraph using the match spec function {caller}.
> This you might do with the dbg module, maybe some other trace tool,
> or write one yourself. That would give you an idea of who is to blame.
>
> >
> > Thanks,
> >
> > kaiduan
>
> > _______________________________________________
> > erlang-questions mailing list
> > erlang-questions@REDACTED
> > http://www.erlang.org/mailman/listinfo/erlang-questions
>
> --
>
> / Raimo Niskanen, Erlang/OTP, Ericsson AB
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20090128/bb17e883/attachment.htm>