Monday 2018-10-22
—————————
Attendees: Guilherme, Oksana, Enric, Stefan x 2, Javier, Vassil, Philippe, Yuka.
Shift/Oliver:
All the forum questions were basically covered or followed.
Last week, apart the forum, I spend quite a lot of time trying to
understand why the reference guide was not building.
I finally found that two dataframe examples are blocking (df10*.C). They
execute but then ROOT does not quit. That shows on the machine:
ec-ubuntu-14-04-x86-64-2
For now I disabled these two examples. The doc build worked
when I stated by hand. I am now redoing it via jenkins
PPP: 2nd part of RForest discussion including Macro Klemencic. There is a PR not-for merging that is open for comments. Guilherme already commented, others are also welcome to comment.
PPP this week cancelled.
TMVA: no attendees.
Godparents meeting:
- Some comments on how to improve publicizing next workshop, offers for helping with that.
- Suggestion to offer recordings of presentations rather than vidconf, also to keep motivation for budget holders to *send* people
- Query for affiliation (experiment) next time, ideally in registration
- Suggestion to have topical "lab" sessions where audience can contact devs and discuss with them
- Suggestion to present feedback summary also to the audience, e.g. on a daily basis (e.g. next morning)
v6.16:
————
Merging of Stefan’s PR.
Vassil: a bunch of things for the module that we might be able to work-around/solve, we still have 7 tests blacklisted but those are bug fixes not major feature.
Vassil: on Lorenzo’s plate, automatic differentiation. Vassil thinks that it is too risky for the release but it is Lorenzo’s call.
Guilherme: blocker right now when building CLAD is on, CLAD will try to install in /usr/local and fail the build. CLAD was crashing with gcc 8 but this is fixed.
Oksana: I am in the process of review my status on compression. I am hoping to have ZLIB fixed probably not LZ4 though. I will give an update tomorrow.
Guilherme: Should clear out the warning for gcc 8 and latest version of clang.
v7 bump related to module?
—————————————
Vassil mention that llvm upgrade is not urgent because patch are minimal so far. So maybe upgrading in Spring. OrcJIT v2: enables parallel JIT compilation, still experimental, v1 is soon going away. working on (maybe) resolving symbol from static libraries. (still likely a multi-year project).
forum.root.cern:
———————
general statement: not worth the effort.
lack of good argument in favor.
root.cern/forum is a good alternative.
Round Table:
——————
Javier: RDataFrame, partition leads to duplicate data reads even when using cluster boundaries.
Philippe: File seems very odd …. [Later found that the problem can not be reliably reproduced and might be a tooling issue]
Yuka: LLVM developers meeting, totally sick now, I still have to fix a few test for modules. We should check out LLVM MCA that can be used to analyze performance of llvm byte code, including vectorization and other architecture dependent optimization. -fstrict-with-table-pointer (spelling?) that turns virtual function call, when possible, they are turned into direct calls, result 0.8% improvement in google infrastructure; I am planning on trying to use it in cling.
Guilherme: see also the intel-archicture-analyzer.
Yuka: Also discussing to start requiring C++17, i.e. skip C++14.
many bugs in libC that google is ‘relying’ on. Nintendo is using LLVM for the SWITCH (mario cart, etc.).
Many talk about vectorization and parallelization.
Memory tagging: amazing talk to enable memory corruption debugging; need support for hardware vendor.
Oksana: New PR build that is module enable but non-obvious name. Working on compressions PRs.
Stefan: In roofit, there are many problem and JIRA tickets and I am running down that (long) list. No large development started yet.
Vassil: Spent a week in California, including Berkley, Computation Chemistry interested in CLAD (They use 10% of DOE resources to do that type of work). Had a talk on automatic differentiation, discovered that SWIFT has automatic differentiation as a first class feature (related to machine learning). They are finding efficient way to multiple matrices. Invited, Oct 30, tensor-flow/SWIFT guy, will also travel to Google/Zurich. They are planning to rebuild TensorFlow with proper SWIFT support, in particular to control what is executed where. [TensorFlow on ‘steroid’ steer by a language they can influence [they can’t do that with C++ or Python]]. Announcement likely next year. Vassil very optimistic that we might be benefit from this as much as we did from LLVM.
Vassil: Investigating how to reduce memory usage in the module.
Guilherme: Focussing on writing CHEP paper. Had 2 babies recently, so I was distracted :). On the technical side, I will continue to reduce the use of globing in our cmake files.
Stefan: CERN computing school in Tel Aviv. I have worked on the RDataFrame example, in particular the physics one. I updated to switch from read from web server to reading from xrootd/eos. [Still need to disable them when xrootd support is disabled]
Enric: Working on PR with inheritance pythonization. Also talking to Wim on cppy behavior, for derived class it is ‘automatic’ if (and only if) the derived class does not over-ride the function that is pythonization.
Enric: We had tutorials for the ATLAS PhD students. 1.5 hours tutorials focusing somewhat on RDataFrame, student choose to stay an hour longer to finish exercises.
Enric: Working with Danilo on the ATTRACT proposal (100,000 euros = one technical student) on data analysis workflow (using RDataFrame).