SFrame Resources
I have been learning about the SFrame tracing effort and figured I should document the resources I have reviewed. Indu Bhagat has been actively involved in the development of SFrame. This is one of her talks giving an overview of the objectives of SFrame. The overall idea is that profiling tools (e.g. perf) usually need to generate stack traces. She lists some methods used to generate stack traces, e.g. using frame pointers, EH frame, last branch record (LBR), and other heuristics. Each of these have their own advantages and pitfalls. SFrame encodes the minimal info required for stack tracing.
I found additional videos by searching for sframe indu (there are lots of unrelated sframe results out there). This one by Steven and Indu covers potential issues that need to be addressed for JITted code.
There are various informative discussions about SFrame out there, e.g.
- SFrame based stack tracer for user space in the kernel [LWN.net]
- Fedora’s tempest in a stack frame [LWN.net]
Indu and Steven also talk about SFrame at the Linux Storage, Filesystem, MM & BPF Summit | LF Events (linuxfoundation.org). This video feels a bit more detailed than the previous one. In this video, it is explicitly outlined that frame pointers require setup in every function and increase register pressure. The cost of this cites the Fedora’s tempest in a stack frame [LWN.net] article. Steven also discusses how the Orc unwinder was created to solve the need for accurate stack unwinding. This is needed, for example, for live kernel patching (see What is Linux kernel live patching? (redhat.com) for more info). SFrame is based on orc unwinder but for user space.
These resources provide a high-level view of stack tracing/unwinding concerns. Tools like async-profiler/async-profiler: Sampling CPU and HEAP profiler for Java featuring AsyncGetCallTrace + perf_events (github.com) will probably look very different if they switch to using SFrame.
Leave a Reply