Profiling
profiling
perf tool
- Install
apt-get install linux-perf
- Produce profile data
Compile with '-fno-omit-frame-pointer' options as CFLAGS. An alternative exists, using the '--call-graph dwarf' option, however it apparently generates a lot of more disk data. To know more about such option, take a look at the link http://www.brendangregg.com/perf.html#StackTraces and use 'perf help record' to show all the options.
For instance, use
-Og -g3 -fno-omit-frame-pointer
or the more gdb specific
-Og -ggdb3 -fno-omit-frame-pointer
There are a lot of event that can be monitored, and they can be specified using the -e option. The default event is just 'cycles' (or 'cycles:pp' apparently), but the option accept multiple event types separated by a comma, like '-e cpu-clock,faults' for instance.
perf record -q -e cycles -g --call-graph fp -- ./COMMAND ARG1 ARG2 ... ARGN
- Visualization
perf report --stdio -g none --sort comm,dso # load per module perf report --stdio -g none # load per function perf report --stdio -g graph # call chains perf report -g graph -s period,comm,dso,symbol # TUI interface
Example of output of STSM program:
- 92.18% 0.00% stsm stsm [.] main
- main
- 91.77% STSM::run
+ 56.86% STSM::generateCandidates
- 25.22% STSM::detectBlocksOfAllSolidSequences
+ 23.42% STSM::detectSolidSequenceBlocksFromSolidSequence
0.81% Segment::unify
+ 5.25% STSM::updateKernelsOfAllCandidates
1.80% RangedSequence::range
+ 1.45% STSM::updateMatchingPositions
0.99% Segment::intersects
+ 92.18% 0.00% stsm libc-2.24.so [.] __libc_start_main
+ 92.18% 0.00% stsm [unknown] [k] 0x4d96258d4c544155
+ 91.77% 0.00% stsm stsm [.] STSM::run
+ 56.86% 6.74% stsm stsm [.] STSM::generateCandidates
+ 49.99% 49.99% stsm stsm [.] Segment::intersects
+ 25.22% 0.00% stsm stsm [.] STSM::detectBlocksOfAllSolidSequences
- GUI
There is a script to generate graphs.
Requisites:
apt-get install python3 graphviz
Test it after using perf record command:
perf script | c++filt | gprof2dot.py -f perf | dot -Tpng -o output.png
References
- perf Tutorial
- perf command manuals: record and report
- Linux perf Examples
- tools considerations
perf example (offline), webarchive link- simplest perf example
valgrind tool
valgrind --tool=callgrind --callgrind-out-file=I100_Ooriginal_S26_FS10_FB05_MS0.grind ./stsm 100_sax-26_sample-46x951.csv I100_Ooriginal_S26_FS10_FB05_MS0.js\ onn I100_Ooriginal_S26_FS10_FB05_MS0.log 10 05 qcachegrind I100_Ooriginal_S26_FS10_FB05_MS0.grind
gprof tool
It is included in binutils package. To use:
1) compile and link sources using -pg flag
2) run the compiled source, it will generate a gmon.out file
3) use gprof to evaluate
- Example
cc -Wall -pg main.c -o a.out ./a.out gprof ./a.out gmon.out > report.txt