[PATCH] Make LNT compatible with PostgreSQL

Tobias Grosser tobias at grosser.es
Wed Jul 30 12:26:33 PDT 2014


On 30/07/2014 20:45, Daniel Dunbar wrote:
> On Tue, Jul 29, 2014 at 9:56 PM, Tobias Grosser <tobias at grosser.es
> <mailto:tobias at grosser.es>> wrote:
>
>     On 30/07/2014 00:30, Daniel Dunbar wrote:
>
>         Here are my thoughts on what we should do:
>
>
>     Hi Daniel,
>
>     thanks for your feedback.
>
>
>         1. Before doing anything substantial, I want to get Chris'
>         patches to
>         rerun tests with significant changes (~= a form of adaptive
>         sampling)
>         landed. I have high hopes for that approach in helping making
>         results
>         more reliable and actionable.
>
>
>     Adaptive sampling is a very neat idea.
>
>     Out of interest. How does a database change require these changes?
>       From my naive perspective, I would rather have a set of already
>     running buildbots with some history to allow to understand the
>     effectiveness of Chris' changes. Hence, having a stable database in
>     place would be nice.
>
>
> It doesn't, but if the changes work well there is some value in having
> the data set be consistent, I thought. The other reason was that if the
> changes have bugs or need tweaks, it would be nicer to sort out the
> issues before bringing up a new database to keep things "clean" later.

We could still wipe the db after some experimental phase.

>         2. In the past, when bringing up new databases I have reimported
>         some
>         historical data using the JSON files that the server archives
>         (as Chris
>         noted). I could do that again here if useful.
>
>
>     Sure. We have the last 500 builds going back to July 16 for the
>     'clang -O3 builder'. That's 15 days history. Nothing huge, but just
>     enough to get us history starting from the 3.5 branch.
>
>
> The server actually has much more data than that, I have the files to
> import back to 2012.

Amazing.

>         3. I'm not sure exactly when I will have time to bring up a
>         PostgreSQL
>         instance on llvm.org <http://llvm.org> <http://llvm.org>. I
>         would really love to move to a
>
>         PaaS solution like Heroku to make managing this kind of thing easier
>         (and easier to collaborate on), but we might not yet have the
>         organizational clout for that.
>
>
>     You seem very busy and Yi Kong has done a great job in moving LNT
>     ahead the last months. Maybe he could help out with the installation
>     work?
>
>
> Undoubtedly, the problem with having the server on llvm.org
> <http://llvm.org> currently is we have to manage access carefully and
> also be careful. If it was hosted elsewhere, we wouldn't need to worry
> nearly as much about changes.

Right. Maybe someone has a virtual machine to do exactly this? Or the 
LLVM Foundation has funds to get such a virtual machine?

>         4. Last I looked, the llvm.org <http://llvm.org>
>         <http://llvm.org> instance mostly had
>
>         polly related bots. It would be nice to start off with a more
>         standard
>         set of bots trying to cover the diversity of platforms, in the
>         hopes of
>         making the results more interesting to the larger community.
>
>
>     Right, it is important that we get wider test coverage. On the other
>     side, one server is just badly named:
>
>     http://llvm.org/perf/db___default/v4/nts/recent_activity
>     <http://llvm.org/perf/db_default/v4/nts/recent_activity>
>
>     Those are five machines building 'clang -O3' on X86 without any
>     Polly involved. So for X86 at least one configuration is rather
>     strongly tested.
>
>
> Ah, ok. Can we rename that bot?

r214324

This breaks all history on the perf builder so it would be in fact a 
good time to introduce a new database. ;-) (Possibly with old data 
reimported with matching builder names).

Cheers,
Tobias




More information about the llvm-commits mailing list