Skip to content

debugging_and_profiling.md

Latest commit

 

History

History
65 lines (57 loc) · 3.44 KB

debugging_and_profiling.md

File metadata and controls

65 lines (57 loc) · 3.44 KB

Debugging and Profiling

Logging

  • Use logging instead of print statements
    • Allows output to go to files, sockets, remote servers, etc. instead of just stdout
    • Supports different severity levels (INFO, DEBUG, WARN, ERROR), allowing for filtering
    • Color code them for improved readability
  • systemmd in Linux systems places logs in /var/log/journal, accessed via journalctl command

Debuggers

  • Debugger benefits
    • Halt execution of program at a certain line or when conditions are met
    • Step through program incrementally
    • Inspect variable values after crashing
  • Debuggers
  • System calls are used by programs to perform actions via the kernel; strace allows you to track when they occur
  • Firefox/Chrome are good debugging tools for web dev - inspect site elements, live modification, etc.

Static Analysis

  • Static analysis programs take source code as input and analyze it with coding rules, allowing you to find issues without running code - especially helps for time-intensive programs

Profiling

  • Types of time
    • Real: time from start to finish of program, including time taken by other processes
    • User: time spent in CPU running user-level code
    • Sys: time spent in CPU running kernel-level code
  • Profiler types
    • Tracing: execute with code, take note of every function call. At the end, able to tell time spent in different functions
      • Problem: lots of overhead, affect performance of original program
    • Sampling: on a fixed time interval, stops program, looks at stack trace to see where in program you are, and enough samples will be able to tell where time is spent
  • Profiler examples
    • cProfile: Python tracing profiler that measures time per function call
      • python -m cProfile -s tottime SCRIPT.py 1000 '^(import|\s*def)[^,]*$' *.py
      • Can be hard to parse through, since you might want to measure a function that calls other third-party library functions
    • line_profiler: Python profiler that uses @profile decorator above functions to identify which one to measure
      • kernprof -lv SCRIPT.py
  • Memory profiling: can help identify memory leaks and inefficiencies
    • C++: valgrind
    • Python: memory_profiler
  • Event profiling: ignores specifics, treats program as a black box
    • perf: hardware-agnostic evaluation, reporting system events that occurred from a certain command
      • perf stat COMMAND ARG_1 ARG_2: gets counts of different events related to a command
  • Profiling visualization tools
    • Flame graph: perf COMMAND report flamegraph
    • Call graphs: pycallgraph graphviz -- ./SCRIPT.py
  • Resource monitoring: determine what resource is the main bottleneck/constraint
    • htop: general
    • iotop: I/O usage
    • du, ncdu: disk usage
    • free: memory
    • lsof: open files
    • ss, ip: network connections
    • nethogs, iftop: network usage
  • hyperfine: black box speed comparison between command line programs
    • hyperfine --warmup 3 'COMMAND_1' 'COMMAND_2'
  • eBPF: perform kernel tracing of user programs