Re: [PATCH] Improve performance of IO locks

[Florian Weimer <fweimer@redhat.com>]

* Mark Wielaard:

> Hi Florian,
>
> On Tue, Aug 16, 2022 at 12:08:03PM +0200, Florian Weimer wrote:
>> * Mark Wielaard: 
>> > Yeah, I am actually surprised just reverting the patch fixed things.
>> > But looking at the bunsen data for glibc-debian-ppc64 does seem to
>> > show that the only thing changed is this particular patch. Runs before
>> > it pass, runs after show lots of fails.
>> >
>> > https://builder.sourceware.org/testruns/?has_keyvalue_like_k=testrun.git_describe&has_keyvalue_like_v=%25glibc-debian-ppc64%25
>> 
>> How do I read this table?
>
> Glad you ask! Although I think the mystery of the debian-ppc64 build
> failure has now been solved, the intent of the bunsenweb tool is to
> help figure out test result differences between builders. I have added
> Frank to the CC who can explain better than me. We'll also have a BoF
> at the GNU Tools Cauldron for suggestions on how the improve it and
> the integration with other sourceware services [*].
>
> So the table above is all the builds that bunsen knows about for
> glibc-debian-ppc64. Note that currently we only have one ppc64[be]
> worker that does glibc builds.
>
> You can select a couple of builds and compare them. Pick the failing
> one, which was 498f51, and the last succeeding one. And compare the
> results, configs, etc.

> There are a couple of ways to explore the diffs. One is by looking at
> the buildbot "upload to bunsen" step and select that URL (this really
> should be improved to make this easier to find). This then gives you
> different "clusters" of similar (but possible different) other builds
> to compare to. e.g. look at
> https://builder.sourceware.org/testrun/498f51f327afdd7030516455b709a31a0038b432
> 
> Another is to look at all results for a particular builder (as above)
> or to look at all the glibc builds using the (sql) search form. e.g.
> https://builder.sourceware.org/testruns/?has_keyvalue_like_k=testrun.git_describe&has_keyvalue_like_v=%25glibc%25

Thank you for these links, they surely look useful.  I assume the zebra
pattern in the table is not actually meaningful (it's too regular for
that), it's just to improve readability, right?

What confuses me still is that the run lists do not seem to have numbers
with failure counts.

Thanks,
Florian