Re: Simplistic profiling with the runtime

[fche at redhat dot com (Frank Ch. Eigler)]

charles.spirakis wrote:

> Enclosed is a simplistic example of using the runtime to generate time
> based profiling data. [...]

Nice.

> Has any discussion happen on what we want the systemtap translator
> syntax to look like in this regard? [...]

There are two aspects to it: hooking into the timer of your choice,
and having the pc->string decoding.

The plain jiffies-based timer has just received its own support in the
translator recently.  Adding a new variant for this profiling timer
would not require much code.  See the translator's src/tapsets.cxx,
look for "timer", near the bottom after all the strange dwarf stuff.

Second, the pc->string decoding function (get_image_path) would
probably need to be recast, probably in the form of an embedded-C
function and placed into tapset/SOMETHING.stp.  It would take a PC
argument and return a string.  A few more similar functions would be
needed to return the PC (REGS_IP(CONTEXT->regs->ip)), current->comm
(if any), rdtsc, etc.

Then, the final user script that ties it together might look like this:

global times, samples
probe timer.profiling(100) { # timer.jiffies(10) already works
  s=rdtsc()
  samples[comm(), pc(), path(pc())] ++
  e=rdtsc()
  times <<< e-s
}
probe end {
  print "format1",samples
  print "format2",times
}

... except that this code uses aggregates and the "print" statement,
neither of which are supported by the translator yet.  It will be a bunch of weeks before we get to those.

Until then, one might not bother with the times aggregate (or use an
ordinary array, indexed with a manually quantized e-s value), and use
an explicit foreach() loop in "probe end" to print the array(s).

- FChE