What is Adaptyst?

Introduction#

In the current development stage, Adaptyst is an open-source and language-agnostic code profiler for Linux, built on top of “perf” with custom patches:

• It samples both on-CPU and off-CPU activity of every spawned thread and process.
• It minimises risk of broken profiled stacks for programs compiled with frame pointers (i.e. in case of GCC, with -fno-omit-frame-pointer and -mno-omit-leaf-frame-pointer if available) by detecting inappropriate kernel and CPU configurations automatically.
• It performs cache-aware roofline profiling using the CARM Tool from INESC-ID (only x86-64 at the moment, the AMD CPU support is very experimental).
• Thanks to Adaptyst Analyser, it produces data for rendering interactive flame graphs and charts in a web browser along with cache-aware roofline plots (with an option to plot specific code segments as points on the roofline graphs).
• Its main functionality is designed with hardware portability in mind. Adaptyst has been successfully tested on the x86-64, arm64, and RISC-V instruction set architectures.
• It supports custom sampling-based “perf” events for profiling low-level software-hardware interactions.
• It allows TCP streaming of profiling data to a separate machine for real-time processing.

On-CPU profiling uses perf with the task-clock event. Off-CPU profiling is based on eBPF-implemented sampling explained with the diagram below (using the example of a single process with interleaving on-CPU and off-CPU activity; the diagram uses the old project name “AdaptivePerf”). The sampling period is calculated from a user-provided off-CPU sampling frequency.

Adaptyst flow and components#

As shown in the diagram above, the Adaptyst program flow starts with running the adaptyst command-line tool, which starts the adaptyst-server component in the background by default (it is also possible to run the server on another machine, see External instance of adaptyst-server) and configures and runs “perf”. If roofline profiling is enabled and no cache-aware roofline benchmarks have been performed, adaptyst runs the CARM Tool to perform these benchmarks before starting “perf”.

During profiling, all events are pre-processed by “perf” in real time using the Adaptyst Python scripts and streamed to adaptyst-server (via file descriptors by default or TCP if an external instance is used), which takes care of proper processing such as generating flame graph data and saving the results in a non-encrypted format to persistent storage.

Once your profiled program finishes executing, the results can be inspected in a variety of ways, including by our dedicated tool Adaptyst Analyser producing an interactive website.