Segmentation violations in RooStats / RooFit, related to variable range?

Roofit and RooStats
1

Dear experts,

I am experiencing several segmentation violations, in a context that I will try to describe. I am working with a version of ROOT that I compiled last week after pulling the code from git (6.07/07).

The context is that I’m trying to compute limits, using the RooStats tutorial macro StandardHypoTestInvDemo.C and using the asymptotic calculator. The input workspace contains a combined dataset with two categories (indexed with a RooCategory).

When I run this macro, I get a segmentation violation with the following trace:

Interestingly, I notice another (probably related) crash, when I try to fit the data myself. The fits converges without error only if I provide at least one of the following options to RooAbsPdf::fitTo(): Range(“MassWindow”), or NumCPU(2). If I don’t, most of the time (but not always?) I get a segmentation violation with a trace like the following one:

The second trace in particular makes me think that it is something related to the ranges of my observable (invariant mass, named “invMass”). Indeed, we start from a wide mass window and then reduce the dataset to a narrower one (with a range that we call “MassWindow”). I wonder if it could be that the invMass “version” has different ranges / binnings (different number of them or different names) in the dataset, the pdf, or the different components of them. In any case, I have tried a few things following this assumption, but without success.

Thanks a lot in advance for your help.

Cheers,
Émilien Chapon

2

Hi,

Can you please post your workspace and a minimal macro fitting your model, so I can reproduce the crash ?Otherwise it is difficult to understand the cause of your problem

Best Regards

Lorenzo

3

Hi Lorenzo,

Unfortunately my workspace contains CMS data which is not public and that I would rather not release freely on a forum. I have sent you a private message pointing you to an archive with some macros and workspaces to reproduce the issue. If anyone else would want to have a look, I can point them privately to the same archive. I will try to reproduce the issue with dummy data, so that I can make it public.

Cheers,
Émilien

4

I have generated pseudo-data, so that the crash can be reproduced without using CMS data. The archive is accessible at the address cernbox.cern.ch/index.php/s/eCbv7JjFiP9aOkj.

If I summarize our workflow: we consider fits of the invariant mass of two dimuons, separately in pp and PbPb collisions, and then combine these two workspaces into one (using the combinedWorkspace.C macro – the connection between pp and PbPb is that we express the yield of psi prime meson in PbPb as a function as the one in pp). At last we want to derive Feldman-Cousins intervals on our observable (the double ratio [N(psi’)/N(J/psi)]_PbPb / [N(psi’)/N(J/psi)]_pp, named RFrac2Svs1S_PbPbvsPP in the workspace) using runLimit_RaaNS_Workspace.C.

You can reproduce the first crash with the following:

You may also reproduce the second crash with the following:

5

Despite further investigation on my side, I haven’t found yet why I get this crash. I thought I’d share one of my findings though: it looks like the crash happens when the program encounters a pointer of the form 0x3ff0000000000000 (which is of course far beyond the range for the memory I have), either for _htableName in RooLinkedList::~RooLinkedList(), or for elem in void RooLinkedList::Delete(Option_t *). But I have no idea how this address gets corrupted…

6

Always try to precompile your macro using ACLiC, in order to find source code problems, e.g. try something like (inspect all reported warnings and errors):
root [0] .L YourRootMacro.cxx++

You can “debug” many problems using “valgrind”. See, for example: Segmentation Fault depending on linux distribution