Re: Support sample context in perf report
From: Jiri Olsa
Date: Wed Feb 27 2019 - 06:18:07 EST
On Mon, Feb 25, 2019 at 07:04:01PM -0800, Andi Kleen wrote:
> [Changes:
> Removed already merged patches.
> Address review feedback, see individual patches.
> Now compiles with gcc 8.
> Some minor bug fixes and improvements.]
>
> We currently have two ways to look at sample data in perf:
> either use perf report to aggregate everything, or use
> perf script to look at all individual samples.
>
> Both ways are useful. Of course aggregation is useful
> to quickly find the most expensive part of the code.
>
> But sometimes a single sample is not good enough to
> determine the problem and we need to look at context, either
> through branch contexts, or other previous samples (e.g. for
> correlating different micro architecture events or computing
> metrics)
>
> This can be done through perf script today, but it can
> be rather cumbersome to find the right samples to look
> at.
>
> Another problem with perf report is that it aggregates
> the whole measurement period. But many real workloads
> have phases where they behave quite differently, and it is
> often not useful to combine them into a single histogram.
>
> While this can be worked around with the --time option
> to report, it can be quite cumbersome.
>
> This patch kit attempts to address some of these
> problems in perf report by making it time aware.
>
> - It adds a new time sort key that allows perf report
> to separate samples from different regions. The time
> regions can be defined with a new --time-quantum option.
>
> - Then it extends the perf script support in the
> tui record browser to allow browsing samples for
> different time regions from within a perf report
> session.
>
> - Extends the report browser script display
> to automatically select sensible defaults
> based on what was recorded. For example it will
> automatically show branch contexts with -b.
>
> - Support browsing the context of individual samples.
> perf report can save a limited number of random samples
> per histogram entry with the new --samples option.
> Then the browser allows directly jumping to any
> of the saved samples and browsing the context on the current
> thread or CPU.
>
> There could be probably be done more to make
> perf report even better for such use cases (e.g. a real
> time line display), but this basic support is good
> enough for many useful usages.
>
> Also available in
> git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc.git perf/streams-2
your post is missing this patch, it's only in the branch:
perf tools: Add utility function to fetch executable
jirka