ROOT 6 planning
Present: John H., Marcin, Rolf, Peter, Bertrand, Benedikt, Marco, Andrea, Bill
IMPORTANT NOTE
From Bill: if something is broken in the ROOT repository and if something is fixing the originating problem, it is important that the root developers advertise this clearly to avoid losses of times.
ACTIONS
Andrea: Try w/o the atlas fix if pycool is broken
Marcin
o Data access methods now available after improvement in PyCool
o Two more problems appeared: First issue was about a type name change, fixed on the atlas side. Wim fixed the second issue about the pythonisation of the return value.
o Inclusion of header files at runtime: not fixed yet. Workaround not implmeneted yet
Peter
o Problem of creation of raw data with root. A class has custom streamer, written in splitmode 99, never changed in the past 4 years. Part of the data members are with root6 zero. This is not a priority as the issue is present for datasets dating 2007.
o By next week, the full chain will be complete, numbers will be reported.
Marco
o Complete the exercise of caching the cuts made with python
o All compiles and run: no test of the physics. Enough for a proof of concept. It is not yet complete.
o Ran 1 single event for the stripping (the heaviest workflow). Using that cache instead of python bindings the VSIZE is reduced by 500 MB and the RSS of 450MB.
o It is semi-automatic, 100 packages touched, 1 Month to have the full blown change in hands. The idea in the long term plan is to deprecate the bindings and have python objects which instead of representing c++ objects produce c++ code to be complied.
o With the new method, if a cut is not found in the cache the system fallsback on the present method. HLT is not a worry. The spotlight is on the user jobs: a warning will be prompted if the cut is not found in the cache.
o The exercise can be pushed a bit further eliminating dictionaries at the bootstrap, saving another 70 MB. Going to python-cmodules one can save 100 MB.
Benedikt
o It was proposed to reduce the max-page-size for libraries. An updated custom version of lcg70 brought an improvement of 50 MB in VSIZE, as expected, 0 MB in RSS.
o A proposal could be to enable this by default. The worry is that the alignment is correlated to the ability of the kernel to swap pages of 4MB. We agree that this is useful if you need these 50MB.
Bill
o Good news: CMSSW is functional on ROOT6
o Memory: if one runs with the pch which is built remotely, a +200MB are used.
o Problem is a missing enum even if on the selection XML last week
- Verified that this is fixed this week, and autoload on enums was added for both the rootmap and seal mechanisms
o Still some other test failure