Awesome profiling

General utilites

• https://ebpf.io/: Revolutionary sandboxed kernel profiling technology that makes it easier to build various profiling utilities. Tons of options here in Python https://github.com/iovisor/bcc
• Dtrace: Available on Solaris (includes Mac but not Ubuntu) with notable highlights prstat and mpstat. prstat is not available on Ubuntu but can be replicated with htop and ps
• collectl: Full system level profiling including CPU, disk, memory and network
• perf: CPU level performance counters
• gprof: sampling and instrumentation aware profiling
• google perf tools
• Heaptrack: a heap memory profiler for linux
• jemalloc: another heap memory profiler
• ETW: Event tracing for windows
• Mac OS instruments: Mac OS instruments for profiling based on top of Dtrace
• Renderdoc: Multi platform graphics debugger for OpenGL and Vulkan
• Windows Perf Analyzer: If htop could plot lines, windows only but recently added support for android
• htop: Visualize utilization as bar charts or line charts, issue commands to processes
• Magic Trace: High resolution programmable traces
• pprof: pprof is a tool for visualization and analysis of profiling data

Continuous Profiling

• parca: Continuous profiling for analysis of CPU and memory usage, down to the line number and throughout time. Saving infrastructure cost, improving performance, and increasing reliability
• parca-agent: eBPF-based always-on profiler auto-discovering targets in Kubernetes and systemd, zero code changes or restarts needed! Supports multiple languages: C/C++, Rust, Go, Python, Ruby, Java, etc.

Python specific

• viztracer
• psutil: Like htop but from within your python code
• pyinstrument:python call stack visualizer
• pycallgraph: Visualize call stack as a graph (Maintenance mode)
• py-spy: Sampling profiler for Python
• line profiler: Line by line profiling
• palanteer: Fanciest UI, looks like something out of the matrix
• yappi: multi threaded profiling
• Pycharm profiler: Built in profiler in Pycharm
• TAU
• gprof2dot: Graphical call stack visualizer (Maintenance mode)
• snakeviz: Visualize python cprofile data
• scalene: CPU and GPU based profiling with a web GUI
• pprofile: Very low overhead line profile
• austin-python: Line-level very low overhead time & memory profiler with web & terminal UI
• py-perf A low-overhead, sampling CPU profiler for Python implemented using eBPF.

Ruby specific

• rbspy: Sampling CPU profiler for Ruby
• rbperf: Low-overhead sampling profiler and tracer for Ruby implemented in BPF
• vernier: Next generation CRuby profiler

Java specific

• JProfiler: Java profiler for cpu, multithreading, graphical call stack visualizer
• Java visual VM: Bundled with JDK

Mobile specific

• iOS: Instruments app in Xcode
• Android: Android Profiler in Android Studio

C# specific

• Unity profiler: profiling tools specific for game development

C++ specific

• Tracy: Windows only but very comprehensive and helpful for game development
• Callgrind: Valgrind extension

Web specific

• Chrome profiler: Support for throttling, emulating weak hardware,

PyTorch specific

• Pytorch profiler: Visual profiles of computations and data loading for PyTorch models, requires changes to code
• PyTorch memory profiler: Can help debug OOMs and memory spikes

CPU specific

• ARM profiling: ARM specific profiling tools, heavyweight UI
• Intel Vtune
• Intel GPA: Intel Graphics performance analyzer

GPU specific

• pynvml: Like nvidia-smi for your code with deeper level instrumentation
• NVIDIA visual profiler
• NVIDIA tools
• GPU View: Windows specific GPU profiling
• ROC profiler: AMD ROCm profiler
• Omniperf: AMD profiler for MI100 and MI200 accelerators
• NVIDIA NCU: Infinitely more useful than NVIDIA's nsys, does a godbolt style view on PTX and gives actionable performance hints

Books

• System Performance Enterprise and the Cloud
• BPF Performance tools

Blogs

• Flame Graphs: flame graphs vs flame charts, off cpu profiling, icicle charts and more
• How to read icicle and flame graphs: Flame graphs and icicle graphs are a great way to visualize performance profiles. In this post, we will learn how to read and interpret them.
• Sampling vs Tracing: sampling based profilers are easier to use since they don't require any code change while instrumentation based profilers require code changes but are generally more informative
• C++ performance tools: reddit post with tons of links

Talks

• scalene talk
• Why performance matters
• Common performance anti-patterns

Understanding code structure

• pyreverse: Get python classes and then visualize with graphviz
• pdb: Use step in functionality or line by line to understand how your code works
• IntelliJ UML Class diagrams