Boost logo

Boost Interest :

From: Beman Dawes (bdawes_at_[hidden])
Date: 2008-05-21 10:28:01


On Wed, May 21, 2008 at 8:49 AM, Doug Gregor <doug.gregor_at_[hidden]> wrote:

> On Wed, May 21, 2008 at 7:49 AM, Beman Dawes <bdawes_at_[hidden]> wrote:
> > How does the work on CMake that Troy and Doug are doing relate to the
> work
> > on Bitten that Dave is doing?
>
> Like BBv2, CMake has its own way to do regression testing as part of
> the build. With CMake, it's through a separate (but bundled) program
> called CTest that is configured by CMake. CTest is responsible for
> running tests locally and producing XML results, which it can upload
> to a server through XML-RPC.

FWIW, one of my problems with the current system is that the volume of XML
results chews up too much bandwidth (DSL, 3.0 mbs down/ .7 mbs up) for me to
contribute as much testing as I'd like. Moving the servers to OSL will help,
since I won't have to run the release server here, and Rene won't have to
run the trunk server on his machine.

> CTest is meant to work with Dart, Dart2,
> and CDash; part of what Dave and Troy are doing is to make CTest work
> with Bitten as well, because Bitten will integrate with Trac and
> handle our regression reporting. Dave will also be working to give
> Bitten a more Boost-friendly user interface, categorizing sets of
> tests by library and so on: this is the major feature that nearly
> every regression-reporting system is missing.

Sounds good.

Dave has also been talking about the need for overall measures of what he
calls the health of Boost. To me, that should include QA testing, such as
the Boost inspect program, as part of the regular testing and reporting
system. So I'd like to see a library marked as failing if it is getting
inspect failures. That would be a big incentive to expand the inspect
program to cover more areas, such as documentation probes.

> > Who will be responsible for the scripts that testers rely on? Rene?
>
> I'm guessing Troy or Dave. Troy wrote the original tester scripts for
> the CMake-based build system.

My concern about scripts are ease-of-use and robustness. Before Rene started
working on the scripts last fall, it took several (and sometimes a lot more)
emails back and forth to get a new tester up and running. Even then testing
often stalled for long periods, and more emails went back and forth, because
something broke. If there were changes to testing setup or configuration,
more emails and more delays resulted.

As Rene adding ease-of-use and robustness features like better command help,
self-bootstrapping, and choosing defaults with reliability in mind, the
number of emails and testing stalls dropped dramatically. The Linux tests I
run used to require my intervention perhaps every couple of days. Now they
run many months without intervention. Other testers have had similar
success.

Rene also been able to respond quickly, and keep an eye on the testing
process.

I want to be sure we don't regress in these areas.

> Is the plan to start small with a few libraries and testers, and then
> > gradually roll-out to all libraries and all testers?
>
> All libraries need to build and test so that we can verify that the
> build system is working properly. Plus, we would like to be able to
> build binary installers with CMake sooner, and not wait for all of the
> regression-testing work to be completed.
>

Yes, that would be helpful. Building daily release snapshots has been very
helpful, and the logical conclusion would be to both build and test at least
one binary installer as part of each snapshot.

Does cpack also build 7zip archives? They are becoming more popular, and
Boost has been providing them for several releases.

We'll start with a few testers, and widen the testing base when we
> feel that the testing system is relatively robust. We don't want to
> take resources away from the current testing system too early.

Makes sense.

I'd like to start testing with the new system (does it have a name?) as soon
as a script becomes available.

> > At what point do we
> > turn off the current build/test/report system?
>
> I don't know. I think it will have to coincide with a switch to CMake,
> unless someone goes back and makes bjam/BBv2 support Bitten.

OK. I'd like to be sure both a significant number of testers find the test
runner portion acceptable, and a significant number of developers find the
developer reporting side acceptable before we switch. It also wouldn't hurt
to make sure user reporting is meeting user needs.

> > Are any metrics being collected? I'm interested in time and bandwidth
> > utilization for both test clients and the central server, since these are
> > limiting bottlenecks with the current mechanism.
>
> CTest collects some statistics for compile-time (in aggregate) and
> run-time of individual tests. It's in the XML, and we could certainly
> report that back to Bitten.

What I'm interested in (as a release manager concerned with having enough
testers for good coverage) are some overall measures of what kinds of
demands we are placing on testers and servers. For example, how much time
(wall clock? CPU?) is the tester using per day? How many bytes received?
Transmitted? Similar for servers. No need to beat these to death; just three
or four overall indications of resources consumed. Useful surrogates if
those particular numbers aren't available.

> > Who will host the central server or test results repository?
>
> Since Bitten integrates with Trac, we'll need to host this at OSL
> where the Trac lives.

Excellent! OSL has been very reliable, and a key element in the success of
Boost!

>
> > Which is the best mailing list for keeping the entire set of
> > build/test/report needs coordinated?
>
> CMake-related build/test issues should go to this list. Once the
> testing system is being used by others, boost-testing will still be
> the coordination point for regression testers.

Thanks for the responses. Very helpful.

--Beman



Boost-cmake list run by bdawes at acm.org, david.abrahams at rcn.com, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk