Magic-trace: Diagnose tricky performance issues with Intel Processor Trace

It seems like this is basically unavoidable on existing hardware, though, right?

if we imagine there existed some visualization that could more accurately represent the complexity of a core, I don’t know how it would be possible to get the data, because AFAIK there are no methods to trace processor execution for modern processors at higher fidelity than this.

even sampling profilers have similar issues with being limited to the model of sequential instruction streams, since each sample gives a single program counter, not the full view of everything the core has in flight.

Doesn’t VTune support processor trace? Some VMs support PT. And AWS has support also.

This says post-skylake, but both my SKX workstation and i5-6200U laptop have 1 in /sys/bus/event_source/devices/intel_pt/caps/psb_cyc which seems to be the condition, though I haven't tried to use it.

Yes, I agree that higher-resolution data is not readily available. LLVM MCA has a timeline view that attempts to visualize the overlapping execution of instructions (https://llvm.org/docs/CommandGuide/llvm-mca.html#timeline-vi...), but this is based on models of how the CPU works (not runtime-collected data), and these models are not perfect.

I also agree that sampling profilers have the same issue: instruction-level views of sampling profiles should be taken with a grain of salt.

My concern is that flame graphs with 1-3ns of resolution are presented as a selling point of the tool, without any mention of the caveats around how this model really breaks down at this time scale. I would like to know more details of how the PT data actually relates to the out-of-order execution. Does a branch's timestamp correspond to when that branch was retired? Do we actually know what the timestamp corresponds to, or is it not well-specified? Are there cases where the timestamp is known to be misleading about the true bottleneck?

I don't know the answers to these questions, but I see a tool like this, I really want more information about the strengths and limitations of the data.

Hey - I wanted to say that I came across a comment of yours from more than a decade ago (https://news.ycombinator.com/item?id=2328627), and I was startled at how accurate it is as a prediction of how parsers and IDEs are combined today, about 11 years later. I'm glad you're still commenting on here (and what a criminal understatement it is for that page to characterise their tool's flaws as "timing information is less accurate" - that's bloody execution order that you're talking about!).

Anyway, I wanted to say how much I appreciate your comment of 10 years ago. I'm also a parser nerd, and a performance nerd, and I feel strongly that programmers have a professional responsibility to write code in a way that expresses our intent by a logical minimum of instructions/work. I strongly suspect that this will become important again in the future, not because the ratio of software-efficiency to hardware-power decreases again, but because climate concerns will drive us to measure our code in performance-per-watt rather than performance-per-dollar (depending on what action is taken on carbon pricing, it may be a distinction without a difference).

I look forward to the day when grossly inefficient software is rightly considered to be as unacceptable as grossly inefficient SUVs, and people in our profession are forced to take responsibility for the damage that their obscenely inefficient crap is doing. I hope Python 4 comes with a snorkel.