oprofile#

oprofile - How it Works#

How does it work?

  • Hardware based: CPUs have performance counters/events

  • NMI, trapped by Linux kernel

  • Samples sent to userspace

Compared to other tools …

  • Hardware $to$ low overhead

    • 1-3% they say

  • Support for wide variety of performance events

    • Cache miss

    • Branch prediction failure

    • Lots of others I don’t understand (⟶ ophelp)

  • Just works!

oprofile - Basic Usage (1)#

Profiling a single executable

$ operf ./program

  • Creates output directory oprofile_data

  • Used by reporting tools

    • opreport

    • opannotate

    • oparchive

    • opgprof

oprofile - Basic Usage (2)#

Report Everything#
$ opreport
... long list of processes, library and kernel symbols ...
Report Symbols and Their Counters#
$ opreport --symbols
Take Debug Information Into Account#
$ opreport --debug-info

oprofile - Too Much Information#

  • oprofile takes samples no matter what

  • ⟶ too much information

    • Used shared libraries

    • Kernel

Exclude Shared Libraries and Kernel ⟶ gprof Flat Profile#
$ opreport --symbols --exclude-dependent
...
samples  %        symbol name
2828     88.5410  contains
358      11.2085  find_duplicates
8         0.2505  main

oprofile - Call Graph#

  • oprofile collects samples in non-maskable interrupt

  • ⟶ time critical

  • does not (by default) record caller information with every sample

Have operf Record Caller On Every Sample#
$ operf --callgraph ./program
Report Callers and Callees#
$ opreport --callgraph
... no easy reading here ...

oprofile - Kernel Symbols (1)#

Symbols come from files ⟶ file mappings

  • Executables

  • Shared libraries

  • Kernel modules

Kernel itself is not a file!

  • Bootloader loads kernel image into memory

  • Kernel not necessarily contained in the file system

  • ⟶ Flash memory, network (PXE boot), …

$ operf ./program
$ opreport --symbols
...
8         0.4630  no-vmlinux               /no-vmlinux
...

oprofile - Kernel Symbols (2)#

operf needs some help …

Helping operf With Kernel Samples#
$ operf --vmlinux=/root/linux-3.16.5-gentoo/vmlinux ./program

  • vmlinux is not usually an artifact of the kernel build process

  • ⟶ make vmlinux

  • Redhat: kernel-debuginfo

oprofile - Kernel Modules#

Samples from kernel modules …

  • Attributed to module name in operf output

  • … not a file

  • … but can be found in a file

  • ⟶ opreport (and friends) need to look it up

Directing opreport To Module Tree#
$ opreport \
    --image-path=/lib64/modules/3.16.5-gentoo/kernel \
    --symbols
...

oprofile - System-Wide Profiling#

System-Wide Profiling: Why?

  • Application (“appliance’?) consists of multiple processes

  • One or more kernel drivers play a role

  • ⟶ one wants to know more about the big picture

System Wide Profiling (have to be root for that)#
# operf --system-wide
# operf --system-wide --vmlinux=/path/to/vmlinux

⟶ Samples from everywhere …

  • Userland processes

  • Kernel code used by system calls

  • Kernel code from interrupt service routines

  • Kernel code from kernel threads

oprofile - Offline Profiling#

Absolutely cool: oparchive

  • Samples gathered on production system

  • Analyzed on development system

  • ⟶ Transfer of every file involved

Archiving on Production Machine#
$ operf ./program   # or whatever ...
$ oparchive --output-directory=$HOME/operf-output
$ tar -C $HOME -jcf operf-output.tar.bz2 operf-output
Later, at Home#
$ tar jxf operf-output.tar.bz2
$ opreport archive:./operf-output

oprofile - Graphical Beauty#

Real men don’t need no graphics - but anyway, here it is …

Converting oprofile_data/ into callgrind Stuff#
$ opreport -gdf|op2calltree
$ ls oprof.out.*
oprof.out.firefox                oprof.out.NetworkManager
oprof.out.gnome-settings-daemon  oprof.out.nm-applet
oprof.out.gnome-shell            oprof.out.operf
oprof.out.gnome-terminal-server

  • One callgrind file for each process oprofile_data/ has samples from

  • Nicely viewable with kcachegrind