harvey/sys/doc/profiling.txt

Profiling (from the Akaros document)
===========================
2015-07-15 Barret Rhoden (brho)

Contents
---------------------------
"Oprofile"

"Oprofile"
---------------------------
Akaros has a very basic sampling profiler, similar to oprofile.  The kernel
generates traces, which you copy off the machine and process on Linux.

To get started, make sure #K is mounted.  

/ $ bind -a \#K /dev/

You control the profiler with the kpctl file.  The general style is to start
the events that trigger a sample, such as a timer tick, then you start and stop
the profiling.  The distinction between the two steps is that one actually
fires the events (e.g. the timer IRQ), and the other enables *collection* of profiling info when those events occur.

TODO from Akaros:
The optimer command takes the core id (or "all"), followed by "on" or "off".
As with all good devices, if you echo garbage, in, you should get the usage as
an errstr.  That'll be kept up to date more than documentation.

/ $ echo garbage > /dev/kpctl
echo failed: Unspecified, startclr|start|stop|clear|opstart|opstop|optimer

/ $ echo optimer garbage > /dev/kpctl
echo failed: Unspecified, optimer [<0|1|..|n|all> <on|off>] [period USEC]

Let's set up the timer on core 0:

/ $ echo optimer 0 on > /dev/kpctl

And then start oprofile system-wide.

/ $ echo opstart > /dev/kpctl
Enable tracing on 0
Enable tracing on 1
Enable tracing on 2
Enable tracing on 3
Enable tracing on 4
Enable tracing on 5
Enable tracing on 6
Enable tracing on 7

Run whatever command you want, then stop the profiler.

/ $ foo
/ $ echo opstop > /dev/kpctl

TODO
Might as well turn off the timers:
/ $ echo optimer all off > /dev/kpctl

Now we need to extract the trace.  The easiest way is via 9p.
/ $ cat /dev/kpoprofile > /mnt/trace

Once the trace has been read from kpoprofile, it cannot be read again.  The
read drains the kernel's trace buffer.

The trace that the kernel generates is in an Akaros-specific format.  There is
a go program at tools/profile/op2.go that translates from the Akaros format to
pprof format.  You could run this on Harvey, since we support Go programs, but
since we don't have a port of pprof, it's easier to do it all in Linux.

So now we're in linux, and say our 9p ufs server is rooted at mnt/netroot/.  Run op2:

(linux) $ op2 < mnt/netroot/trace > trace-pp

To get a sense for what the trace holds, you might want to start with looking at the raw addresses to distinguish between the kernel and the user.

(linux) $ pprof --addresses trace-pp
PPROF> top
	(shows some addresses)

Say the majority of the addresses are user addresses:

(linux) $ pprof obj/tests/foo trace-pp
PPROF> top
	(shows some functions)

Or you can visualize things:
(linux) $ pprof --evince obj/tests/foo trace-pp

The visualization is not of much user for user programs, since the kernel does
not record backtraces for userspace by default.  It's a little dangerous at the
moment.  In the future, we may have an op option to control whether or not the
kernel attempts a backtrace.

For more info on pprof, check out:
http://gperftools.googlecode.com/svn/trunk/doc/cpuprofile.html