[Rivet] Extension of mcplots/Rivet for Heavy Ion Analyses

[Benedikt Volkel]

Hej Andy, hej Frank,

thanks for your replies! It is great that you are interested in a
discussion.

Basically, the proposed solution #2 would be a very easy solution
because it is both fast and simple to implement and there is no need to
extend Rivet. However, there are major drawbacks arising from further
desired capabilities which affects also proper resource management such
that #1 or #3 should be preferred over #2.

Especially in the case of R_AA analysis it is interesting to combine the
output of different AA generators with different pp generators. If only
a direct read-in from HepMC files is possible, there are two ways of
doing that. Firstly, one can save the entire HepMC files of single runs
in order to pass the desired combinations to Rivet afterwards. This
method requires a lot of disk space and deleting the files means a
complete new MC run in case it is needed again. The other way is the one
Frank suggested. In this approach the problem is the large amount of
time because for every combination, two complete MC runs are required.

To overcome the drawbacks it would be nice to be able to recycle
generated YODA files and to put them together afterwards in any desired
combination. This saves both computing power/time and disk space.

More generally, the fact that #2 means not a full integration in Rivet
leads to the question of how certain pre/post-processes are handled in a
standardized manner and how the right ones are matched to certain
analyses. It might be difficult to ensure that something like

$ rivet -a EXPERIMENT_YEAR_INSPIRE fifo.hepmc

still works. Consequently, there might be some actual paper-based
analyses which cannot be handled by non-Rivet-experts.


Finally, a solution according to #3 might be preferred over #2 and #1
because

-> everything, including general post-processing steps, can be handled
only by Rivet,

-> resources are saved (in a more general way, not only regarding R_AA
analyses),

-> there are other scenarios like those Andy mentioned which could be
also handled in this approach.

What do you think about that? Again, we are glad about getting your
feedback and additional ideas.

Cheers,

Benedikt

On 11.08.2016 22:10, Andy Buckley wrote:
> Hi Benedikt,
>
> Thanks for getting in touch -- sounds like a productive project.
>
> Actually, version 3 sounds a lot like what we have had in mind for 
> some time, as part of the development branch for handling events with 
> complex generator weightings: we planned to be able to initialise 
> Rivet analyses with pre-existing YODA files, populating the 
> "temporary" data objects and re-running the finalize() function.
>
> We intended this mainly so that YODA files from homogeneous, 
> statistically independent parallel runs could be merged into one 
> mega-run, and then the finalize() steps re-run to guarantee that 
> arbitrarily complicated end-of-run manipulations would be correctly 
> computed. But it sounds like it would be similarly useful for you...
>
> Thanks for the pointer to your code. Can I ask how long you will be 
> working on this project for? I look forward to the discussion and 
> hopefully some of us will be able to meet in person at CERN, too... 
> but if you would still be available at the end of September you would 
> be very welcome (and we can pay) for you to attend our 3-day developer 
> workshop.
>
> Cheers,
> Andy

On 11.08.2016 09:41, Frank Siegert wrote:
> Dear Benedikt,
>
> thanks for your mail and for describing your problem and solutions very clearly.
>
> I want to throw into the discussion a 4th alternative, which is
> somewhat similar to your #2 but doesn't need any modifications of
> Rivet itself. I have used this approach with a Bachelor student when
> we were trying to determine non-perturbative corrections with
> iterative unfolding, i.e. we needed both hadronised and non-hadronised
> events in the analysis at the same time to fill the response matrix.
> Thus, for us it was important to preserve the correlation between
> hadronised and non-hadronised event, which for you is not an issue, so
> maybe this method is not necessary or more complicated for you, but I
> thought I'd mention it nonetheless.
>
> We are running a standalone pre-processor script which combines the
> HepMC files from the two generator runs, and by using appropriate
> particle status codes embeds the non-hadronised event into the
> hadronised one. We then wrote an analysis plugin including a custom
> projection, which can extract separately (based on the particle
> status) the non-hadronised event and the hadronised event from the
> same HepMC file. This allowed us not just to fill the two sets of
> histograms in the same run (and divide them in finalize), as you would
> want to do it, but also fill a response matrix with correlated events,
> which you probably don't care about.
>
> So basically all you would need is a pre-processing script to combine
> the HepMC files, which could possibly be included in your HI generator
> interface and thus not disrupt the workflow. But maybe this is too
> complicated, and given that you don't need the correlations you might
> be better off with your approach #1.
>
> Cheers,
> Frank

>
>
> On 10/08/16 21:55, Benedikt Volkel wrote:
>> Dear Rivet developers,
>>
>> my name is Benedikt Volkel and I'm working on a summer student
>> project in ALICE with Jan Fiete Grosse-Oetringhaus and Jochen Klein.
>> The goal is to extend the mcplots project to cover specific needs
>> arising from heavy-ion analyses. In particular, we want to implement
>> a post-processing step, which is frequently required for heavy-ion
>> analyses. This step must take place after the production of certain
>> analysis output, e.g. to combine results from different generator
>> runs. As mcplots is based on the standard Rivet work flow, the
>> questions do not apply just to mcplots but more general to Rivet. To
>> sketch the problem in more detail and start a discussion on a
>> possible implementation of a standardized post-processing step we
>> use the example of an R_AA analysis as a starting point.
>>
>> The conceptual problem of an R_AA analysis is the combination, here
>> a division, of heavy ion (AA) data and pp data. The two types of data
>> are provided by different generator runs. We will always assume that
>> Rivet can figure out whether it gets events from an AA generator or
>> a pp generator. This differentiation could be done by evaluating the
>> heavy ion block 'H' in a given HepMC file and/or by reading the beam
>> types. We have investigated the following 3 approaches and would
>> like to ask for comments and feedback:
>>
>> 1) External script: In this approach we don't modify the Rivet
>>    framework at all. The analysis is run independently for the two
>>    generators (pp and AA), in each case only one type of histograms
>>    is filled while the other stays empty. In the end, we use an
>>    external program/script to process the YODA output of both runs
>>    and perform the division. This can be done by using the YODA
>>    library and hence easily in Python or C++.
>>
>>    Comments: So far there is no standard way to distribute or
>>    execute such a post-processing executable. A standard work flow
>>    to include a post-processing step would be desirable. A
>>    standardized hook to execute an arbitrary external script might
>>    provide more flexibility because those external scripts could be
>>    written in Python or C++ and could have an almost arbitrary level
>>    of complexity.
>>
>> 2) Specific executable, Rivet as library: In this case I wrote an
>>    executable which takes care of creating one instance of
>>    Rivet::AnalysisHandler and manages the read-in of two HepMC
>>    files. I based the code on the source code of the executable
>>    $MCPLOTS/scripts/mcprod/rivetvm/rivetvm.cc implemented in
>>    mcplots. My modifications are sketched in [1]. In this way, both
>>    data sets are available in the same analysis and the division can
>>    simply be done in the finalize step.
>>
>>    Comments: It is also already possible on the commandline to pass
>>    two or more HepMC files to Rivet for sequential processing.
>>
>> 3) The goal of my last approach was to enable Rivet to produce
>>    reasonable analysis output without external dependences.
>>    Furthermore, it should be possible to have asynchronous
>>    production of pp and heavy ion YODA files independent from each
>>    other, bringing those together using only Rivet. Therefore, Rivet
>>    was modified to allow reading back the YODA output. This
>>    allows us to implement also the post-processing in the analysis
>>    class.
>>
>>    Comments: You can find the code on
>> https://gitlab.cern.ch/bvolkel/rivet-for-heavy-ion/tree/working.
>>    The basic steps can be found in [2] and more comments can be
>>    found directly in the source code.
>>
>>    For the R_AA analysis, Rivet can be first run with a pp generator.
>>    In the resulting YODA file only the pp objects are filled. In a
>>    second run, with the AA generator, Rivet can be started passing
>>    the first YODA file as additional input. In the Rivet analysis
>>    itself the heavy ion objects are filled. However, after
>>    finalize(), a method Analysis::replaceByData is called. The
>>    objects normally produced by the analysis are replaced by those
>>    from the provided YODA file if they have a finite number of
>>    entries. Hence, after the replacement there are filled and finalized
>>    pp objects coming from the first run and AA objects from the second
>>    run. Those can now be used in an newly introduced post() method
>>    which manages e.g. the division of histograms in case of the R_AA
>>    analysis. It is also possible to provide a YODA file where both
>>    the pp and the AA objects are filled and the R_AA objects have to
>>    be calculated. No actual analysis is done (0 events from an MC), but
>>    init(), analyze(...), finalize() and the post() method are called. 
>> The
>>    histograms are booked, nothing happens in analyze(..) and
>>    finalize(), the corresponding histograms are replaced after 
>> finalize()
>>    and post() handles the division. Basically, this is a similar
>> approach as
>>    the one in scenario 1) but no external dependences are involved. 
>> Also,
>>    scenario 2) remains possible if the YODA input is avoided.
>>
>> All methods work and lead to the desired output. The first two do
>> not need an extension of the Rivet source code. While first one allows
>> for the largest amount of flexibility, the second one is the one
>> which can be implemented most quickly and where all steps can be
>> encapsulated in one analysis class in Rivet. However, always two MC
>> generators runs are required at the same time. Finally, there is one
>> thing all approaches have in common, though, namely the extension of
>> the rivet executable and related ones, in order to account for these
>> analyses types in the command line: 1) linking to the external
>> post-processing script in a consistent way, 2) parallel processing
>> of at least two HepMC files, 3) read-in of a YODA file.
>>
>> I hope that I could explain the issues in a reasonable manner. Jan,
>> Jochen and me are looking forward to a fruitful discussion of how to
>> implement analyses like the one mentioned above in a reasonable way.
>> Please give us critical feedback concerning the approaches and let us
>> know if there are more appropriate ways of solving our problems which I
>> haven't accounted for yet. A consistent and straightforward way of
>> implementing those analyses in Rivet would be extremely helpful.
>>
>> Best regards and many thanks,
>>
>> Benedikt
>>
>>
>>
>> ---------------------Appendix---------------------
>>
>>
>> [1] Sketch of the modification of
>> $MCPLOTS/scripts/mcprod/rivetvm/rivetvm.cc
>>
>> ---------------------------------
>> ...
>>
>> ifstream is1( file1 );
>> ifstream is2( file2 );
>>
>> ...
>>
>> AnalysisHandler rivet;
>>
>> HepMC::evt1;
>> HepMC::evt2;
>>
>> rivet.addAnalyses( RAA_analysis );
>>
>>    while( !is1 || !is2 ) {
>>
>> ...
>>    evt1.read(is1);
>>    evt2.read(is2);
>>
>>
>>    analyze(evt1);
>>    analyze(evt2);
>>    ...
>>
>> }
>>
>> ...
>> ---------------------------------
>>
>> [2] Basics of the small extension of Rivet
>>
>> Rivet::Analysis:
>> Introducing member: bool _haveReadData
>> Introducing: void post()
>> Introducing: void Analysis::replaceByData( std::map< std::string,
>> AnalysisObjectPtr > readObjects )
>>
>> Rivet::AnalysisHandler:
>> Introducing members: std::map< std::string, AnalysisObjectPtr >
>> _readObjects and bool _haveReadData
>> Introducing: void AnalysisHandler::readData(const std::string& filename)
>> _______________________________________________
>> Rivet mailing list
>> Rivet at projects.hepforge.org
>> https://www.hepforge.org/lists/listinfo/rivet
>
>