[Rivet] installation

[Andy Buckley]

On 27/01/14 13:43, Frank Siegert wrote:
> Hi Andy,
> 
> I think there is one simple issue explaining the difference in
> perception between you and Frank et al.:
> One actually needs to install *Rivet-dev* to get current analyses.

Or 1.9.0... depends how much you want YODA.

I think this highlights (again) that having a robust way to decouple the
analysis releases from the base system would help a lot of people. The
hard bit is that "robust", since analysis requirements are often the
things that inform how we should improve the interface, i.e. so far
system changes have been quite tightly coupled to analysis changes.
Decoupling would be great, but might also introduce more ways to screw up...

> And installing Rivet-dev is the annoying bit, in particular for Rivet2
> vs. Rivet1 due to:
> 
> 1. Cython
> It's simply not possible to install cython properly into the install
> prefix automatically in a bootstrap script, which is why we have
> worked around by setting the PATH and PYTHONPATH to find it during the
> bootstrap. But that "solution" fails in case you want to actually
> update your Rivet later (new analyses!) and compile again, since those
> variables aren't set anymore. Swig was much easier in that respect.
> 
> 2. hg
> While SVN was basically available on all systems that I have worked
> on, I often had to install hg myself.
> 
> 3. YODA from hg (release sometimes isn't new enough for Rivet-dev)
> + its own dependencies.
> 
> 4. autotools
> I have seen two independent clusters where the autotools from SL6 (or
> SL5, not 100% sure) lead to very hard to understand linker problems,
> which went away with recent autotools.
> 
> 5. yaml_cpp
> This needs cmake for installation, i.e. yet another dependency. So if
> that's bundled in the future, it will already make quite a difference.

The cmake dependency will still be there, unless someone can work out a
way to avoid it. In Rivet 1's bundled version I had replaced the native
yaml-cpp build with autotools, but that was a problem in itself. Given
the LCG determination to go down the cmake route, I think we're a bit
stuck with that as an unavoidable dependence. Presumably you are having
to install cmake even for tarball builds in your new bootstrap script,
otherwise you won't be able to build an external yaml-cpp?

> Bottom line: no, it's not just about the Boost version in YODA.

Ok, thanks. But I'm now even less sure about what we can do to improve
the situation! I wish it were neater, too.

Andy


> On 27 January 2014 13:03, Andy Buckley <andy.buckley at cern.ch> wrote:
>> On 27/01/14 11:17, Frank Krauss wrote:
>>> Hi Andy,
>>> well, that's great.  I found out about the dependencies the hard way,
>>> but I still did not manage to produce an installation.  And since you
>>> guys appear to claim zero time of developer time it has about zero
>>> usability at the moment for anyone who cannot rely on it being already
>>> installed on lxplus.  I guess however that Rivet was also meant for the
>>> wider MC community, including theorists not sitting at CERN.  For those
>>> it is pretty much useless, and, the way I see it, because the developers
>>> mainly aimed at proving how smart they are rather than at having
>>> something useful.  So, up to now, you just showed me that you guys do
>>> not give a toss about this part of the community.  Good to know.  I
>>> sincerely hope this attitude of producing monumental software for the
>>> nerds is not the new trend in HEP computing.
>>
>> I'll try to reply nicely... but not exactly the style of feedback that
>> makes me want to be Mr Helpful, Frank.
>>
>> First off, of course we care about MC/pheno theorists. But as far as I
>> was aware the part of the community that Rivet was started for (i.e. you
>> guys) were more-or-less happily using it without significant trouble. So
>> the recent work (shall I moan about manpower again?) has been aimed at
>> getting the experiment uptake to the same level as in the generator
>> community. Understandable, no? So it looks like time for the balance to
>> swing a bit back toward the non-LCG users again.
>>
>> I _really_ hope we have better ways to look (be!) clever than introduce
>> new pain-in-the-ass library dependencies. It's more the other way
>> around: when there's a better implementation somewhere else, which we
>> have no chance/time to reimplement internally, we had to be stupid and
>> acquire the dependency. Most of this happened > 5 years ago. Where there
>> have been changes they have been to *reduce* build system problems (cf.
>> the previous way of bundling yaml-cpp). I didn't think that much had
>> changed with the release of version 2, but clearly some changes have had
>> painful impact: it would be good to know exactly what these are. As I
>> mentioned in another email, I suspect that YODA's Boost version
>> requirement is too high -- is that right?
>>
>> The dependencies that we have were chosen (I thought fairly carefully)
>> for compatibility with the "canonical" HEP system, i.e. SLC6 and the LCG
>> libraries. Frank's new script should handle tarball builds on systems
>> where LCG isn't present, which is great. Installation on personal
>> laptops _should_ be far easier because so many dependencies can come
>> direct from system packages... but you're right that that's assuming a
>> degree of nerdity in that you're already administrating your own machine
>> so can presumably choose what to install from where. If not, Frank's new
>> "install everything" script should help... but be aware that if you
>> build one library with a system Boost/GSL/whatever and Rivet with this
>> all-in-one bootstrap then the fact that it fails to run isn't really a
>> Rivet problem! Finally, building an unreleased development version is of
>> course a bit harder: heck, *I* don't try to build the developer version
>> on lxplus direct from the repository: I make a tarball on my laptop,
>> copy it over, and run the standard bootstrap.
>>
>> Ok, enough. The feedback and the more general bootstrap script are
>> really good to have, and I wish we'd known the strength of feeling about
>> Rivet much earlier... but if no-one reports it then we don't know.  Let
>> us know (nicely :-P) if there are more problems -- particularly,
>> concrete help like Frank S' is hugely appreciated.
>>
>> Andy
>>
>> PS. Current manpower, since there was incredulity: I would guess < 1
>> day/month on average for me, plus something similar or less for David G,
>> and occasional (very welcome!) contributions from others like Frank S.
>> Much of that time goes on analysis integration and (slow) feature
>> development. Does it look like more from outside?
>>
>> PPS. When yaml-cpp is bundled back inside Rivet, the tarball
>> dependencies will be exactly the same as Rivet 1, except that the
>> histogramming now has to be built separately. So I think the 1 vs. 2
>> dichotomy is a false one, and the issue is more to with the change to
>> bootstrap scripts. You'll probably still need to have cmake installed,
>> unless we bundle that as well... and why not an entire operating system?
>> But while I dislike that cmake requirement, it's the same for a *lot* of
>> other HEP packages now since LCG standardised on it. There will be a lot
>> of pressure for HepMC3 to go that way, I think.
>>
>>
>>> On 26/01/14 22:59, Andy Buckley wrote:
>>>> Hi Franks,
>>>>
>>>> Well, it's usable in proportion to the amount of developer time that we
>>>> have had to spend on making it nice... the main users what we have had
>>>> to win over are the experiment ones who access it through experiment
>>>> frameworks or from the Genser-built copy. So certainly that's who I've
>>>> been aiming my limited build system dev time at, and to be honest I
>>>> still stand by that choice although I'd rather not have *anyone* annoyed
>>>> by our code.
>>>>
>>>> For a tarball build I think the dependencies are as follows:
>>>>
>>>> YODA requires:
>>>>   * Boost (headers only, but currently asks for quite a new version...
>>>> can maybe be relaxed: any testers?)
>>>>
>>>> Rivet requires:
>>>>   * HepMC, FastJet (unavoidable, I think)
>>>>   * YODA (yes, could have been built-in, but it has wider utility)
>>>>   * yaml-cpp (for 2.1.0 I'd like to use a built-in copy, cf. the latest
>>>> LHAPDF6)
>>>>   * GNU Scientific Library (volunteers to eliminate that dependency? But
>>>> is typically an easily-met requirement, I think)
>>>>   * Boost (headers only, fairly easy version requirement, I think)
>>>>
>>>> If we eliminate yaml-cpp, how much is there to object to? Anyway, yes
>>>> Frank I'd be very happy for you to make a new rivet2-bootstrap script
>>>> which builds everything from scratch... and/or attempts to work out what
>>>> is already on the system? At some level users should make their own
>>>> minds up about what to install from packaging systems and what by
>>>> hand... but by saying that I probably give away what side of Frank's
>>>> nerd fence I'm sitting on ;-)
>>>>
>>>> Anyway, in short I don't see so much room for cutting things down, but
>>>> always think that smoothing the entry paths (with scripts or otherwise)
>>>> is a good thing as long as it's kept manageable for the developers, too.
>>>> Since that's translated as "next to no manpower" for the last... two
>>>> years?, I'm afraid the situation isn't so surprising. But we'll do what
>>>> we can -- certainly I want to get rid of the yaml-cpp explicity
>>>> requirement, and hopefully we can get a couple of new part-time
>>>> developers (re)activated in the coming months.
>>>>
>>>> Andy
>>>>
>>>>
>>>> On 26/01/14 19:47, Frank Siegert wrote:
>>>>> I hate to admit that I agree with Frank's rant here, and I have had
>>>>> multiple users recently who made the same point, though admittedly
>>>>> more politely. The bootstrap script currently is close to useless to
>>>>> everybody who isn't installing on lxplus. I have it on my agenda to
>>>>> replace it with a working version, but haven't got around to it yet.
>>>>> And to be honest, all the changes and external dependencies that came
>>>>> with Rivet2 have left me as a second-class developer (if even),
>>>>> because there is a lot of stuff in the codebase which is beyond my
>>>>> capabilities to fix -- I'm normally quite annoyed myself with even
>>>>> installing Rivet on a new cluster, where
>>>>> cython/yaml/yoda/boost/cmake/hg/fastjet/hepmc... isn't available. Some
>>>>> of those are only necessary for the development version, but it's
>>>>> still rather annoying.
>>>>> This goes far beyond the switch to YODA, which I agree with as
>>>>> necessary, though tbh I originally thought would be Rivet-in-house as
>>>>> well.
>>>>>
>>>>> Anyway, as I'm clearly not the majority of Rivet developers, and
>>>>> mostly everybody else on the team seems to be happy about this usage
>>>>> of fancy new technologies, separation and dependencies, there is no
>>>>> point in discussing that in itself. But we definitely need a usable
>>>>> bootstrap script much more urgently than we needed it for Rivet1.
>>>>>
>>>>> Andy, what about renaming the current rivet2 bootstrap script to
>>>>> rivet-2-bootstrap-lcg, and (me) creating a separate one which actually
>>>>> does all the installations on its own (but without support for Rivet1
>>>>> of course, to avoid it becoming too bloated as the old one did)?
>>>>>
>>>>> Cheers,
>>>>> Frank
>>>>>
>>>>>
>>>>> On 26 January 2014 17:13, Frank Krauss <frank.krauss at durham.ac.uk>
>>>>> wrote:
>>>>>> Dear Rivet Authors,
>>>>>> I am just towards the end of my second hour running trying to
>>>>>> install Rivet 2.0.  Admittedly, I am not half as geeky and nerdy as
>>>>>> you are, so I accepted/anticipated that this would take me an hour.
>>>>>> But, please, can you explain the rationale why you use
>>>>>> cython, yaml, yoda, boost, fastjet & hepmc (I am sure i missed some!)
>>>>>> in a non-trivial way with non-trivial version specifications without
>>>>>> providing a meaningful installation guide?  You certainly do not
>>>>>> want the common people to use your code - is that it?  If you think
>>>>>> this helps the user base you are mightily mistaken.  If you think I
>>>>>> will continue to advertise your code, you're delusional - this is
>>>>>> next to uninstallable as it stands now, with a pretty superficial
>>>>>> installation help.
>>>>>> Best wishes
>>>>>>      Frank
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Rivet mailing list
>>>>>> Rivet at projects.hepforge.org
>>>>>> http://www.hepforge.org/lists/listinfo/rivet
>>>>> _______________________________________________
>>>>> Rivet mailing list
>>>>> Rivet at projects.hepforge.org
>>>>> http://www.hepforge.org/lists/listinfo/rivet
>>>>>
>>>>
>>>
>>
>>
>> --
>> Dr Andy Buckley, Royal Society University Research Fellow
>> Particle Physics Expt Group, University of Glasgow / PH Dept, CERN


-- 
Dr Andy Buckley, Royal Society University Research Fellow
Particle Physics Expt Group, University of Glasgow / PH Dept, CERN