|Version 10 (modified by nfrisby, 3 years ago) (diff)|
Linux perf tool
Since Linux 2.6.31, linux has had a new performance counter subsystem (initially called "perf counters" and later renamed to "perf events"). The facilities provided by perf events, and the associated tool "perf", are generally a superset of what you can do with oprofile and qprof, so it's a good idea to use perf if you can. It does support fewer processors than the other systems, although more are being added over time.
The perf events subsystem is compiled in by default in the kernel shipped with most distros (e.g. Ubuntu) which means there's no fiddling around compiling your own kernel or modules.
On Debian and Ubuntu you can install "perf" by installing the linux-tools-<your kernel version> package:
$ sudo apt-get install linux-tools-<your kernel version>
If your distro doesn't include "perf" as a package you can build it manually like so:
- apt-get install binutils-dev libdwarf-dev libelf-dev (or equivalent on your distro)
- download a kernel source tree, matching your kernel version, from http://kernel.org
- unpack it
- cd tools/perf
- if you want, make install, or just copy the binary somewhere appropriate
Check that it works:
$ perf stat true Performance counter stats for 'true': 3.684489 task-clock-msecs # 0.410 CPUs 1 context-switches # 0.000 M/sec 0 CPU-migrations # 0.000 M/sec 158 page-faults # 0.043 M/sec 2920646 cycles # 792.687 M/sec 2962512 instructions # 1.014 IPC 687407 branches # 186.568 M/sec 24356 branch-misses # 3.543 % <not counted> cache-references <not counted> cache-misses 0.008976351 seconds time elapsed
if you see some zeroes here it probably means your processor isn't fully supported by the kernel's perf events subsystem.
Now to profile a GHC-compiled executable:
$ perf record ./queens $ perf report
The output looks something like this:
# Samples: 9161149923 # # Overhead Command Shared Object Symbol # ........ ....... ................. ...... # 30.65% queens queens [.] s1ql_info 18.67% queens queens [.] s1qj_info 12.17% queens queens [.] s1qi_info 9.94% queens queens [.] s1o9_info 5.85% queens queens [.] r1nI_info 5.33% queens queens [.] s1sF_info 5.18% queens queens [.] s1sG_info 3.69% queens queens [.] s1oP_info 1.68% queens queens [.] stg_upd_frame_info 0.88% queens queens [.] stg_ap_2_upd_info 0.62% queens queens [.] s1sE_info 0.56% queens [kernel] [k] read_hpet 0.39% queens queens [.] stg_ap_p_info 0.35% :2030 f76beb [.] 0x00000000f76beb 0.31% queens queens [.] s1oD_info 0.28% swapper [kernel] [k] mwait_idle_with_hints 0.25% queens queens [.] __stg_gc_enter_1 0.23% queens queens [.] evacuate 0.18% swapper [kernel] [k] read_hpet 0.12% queens queens [.] scavenge_block
which is great for pointing to the hotspots. You can also annotate the source code (of the RTS) or the assembly, using perf annotate.
There is a slight subtlety to using perf on an individual nofib test.
To more accurately measure changes in binary sizes, nofib strips the symbols out of the final executable. Therefore, the useful names like "stg_ap_p_info" and "s1ql_info" will not show up in the perf report output. IE The following commands will yield information that is difficult to use.
$ cd nofib/your/favorite/test $ make clean; make boot; make NoFibRuns=0 >& log $ perf record ./test <test inputs> $ perf report
One workaround is to find the ghc options used in the log file and invoke ghc <options from log> --make Main.hs -o test directly. The useful symbol names will then show up in the perf report output. NB The symbols are in the .o files that nofib's make generates; only the linking must be done without make.