Re: I made a flame graph renderer for git's trace2 output
- Date: Mon, 20 May 2019 14:22:25 -0400
- From: Jeff Hostetler <git@xxxxxxxxxxxxxxxxx>
- Subject: Re: I made a flame graph renderer for git's trace2 output
On 5/10/2019 5:57 PM, Ævar Arnfjörð Bjarmason wrote:
On Fri, May 10 2019, Jeff King wrote:
On Fri, May 10, 2019 at 05:09:58PM +0200, Ævar Arnfjörð Bjarmason wrote:
As noted in TODOs in the script there's various stuff I'd like to do
better, and this also shows how we need a lot more trace regions to get
Hmm. My gut reaction was: doesn't "perf record -g make test" already
give us that granular data? I know "perf" isn't available everywhere,
but the idea of the FlameGraph repo is that it takes input from a lot of
sources (though I don't know if it supports any Windows-specific formats
yet, which is presumably a point of interesting to trace-2 authors).
But having generated such a flamegraph, it's not all that helpful. It
mainly tells us that we spend a lot of time on fork/exec. Which is no
surprise, since the test suite is geared not towards heavy workloads,
but lots of tiny functionality tests.
TBH, I'm not sure that flame-graphing the test suite is going to be all
that useful in the long run. It's going to be heavily weighted by the
types of things the test suite does. Flamegraphs are good for
understanding where your time is going for a particular workload, but
the workload of the test suite is not that interesting.
And once you do have a particular workload of interest that you can
replay, then I think the granular "perf" results really can be helpful.
I think the trace2 flamegraph would be most useful if you were
collecting across a broad spectrum of workloads done by a user. You
_can_ do that with perf or similar tools, but it can be a bit awkward.
I do wonder how painful it would be to alias "git" to "perf record git"
for a day or something.
Yeah I should have mentioned that I'm mainly linking to the test suite
rendering as a demo.
My actual use-case for this is to see what production nodes are spending
their time on, similar to what Microsoft is doing with their use of this
The test suite serves as a really good test-case for the output, and to
stress-test my aggregation script, since we're pretty much guaranteed to
run all our commands, and cover a lot of unusual cases.
It also shows that we've got a long way to go in improving the trace2
facility, i.e. adding region enter/leave for some of the things we spend
the most time on.
Yes, there is more work to do to add more regions to get more
granular data for interesting/problematic things. My primary
goal in this phase has been to get the basic machinery in place
and be vetted with some universally interesting regions, such as
reading/writing the index and the phases of status.
Going forward, we can trivially (permanently) add new regions as we
want. I tend to use temporary "experimental" regions during my perf
investigations so that I don't clutter up the mainline source with
WRT the TODO's in your script:
 I don't think data events will be useful for your usage. The data
values are orthogonal to the time values.
 I would add the child_start/_exit events to the stack. This will
give you the names of non-builtin/shell commands and hooks. The
various "child_class" and "use_shell" and "hook_name" fields will help
you avoid duplicate stack frames (which you'll get for builtin