Attendees: Erica Snider, Herb Greenlee, Tracy Usher, Thomas Junk, Katherine Lato
1) Discussion with Jim Amundson
- On the Event Display project. We have a proposal for how to proceed with adopting a web-based event display and recruiting effort to support it. Erica is charged with following up and reporting back on progress. Tracy raised two points.
- Running on laptops that are not connected to the internet
- The problem of passing data about display objects over the network
- The former problem is solved if the server can run locally (which is a requirement)
- The latter is solved if there is a protocol for matching display data to the screen resolution. There are many successful examples of doing this in the industry.
- On Spack. Plan is to bring more resources to bear on this project, but not clear how or when this will happen.
2) Multi-threading, vectorization and other optimizations
- Tracy commented about talking with Mike and Giuseppe about optimized versions of deconvolution and noise suppression code for ICARUS, building on the multi-threading work Mike did for CalWire_DUNE.
- ICARUS is willing to look at Dave Adams data-prep code as a possible common framework for performing this and similar calculations. Interfaces can be common if inputs are raw digits or wires, and outputs are wires. Provides flexibility to customize algorithms as experiments need.
- They also discussed 2D drift model that has been worked on a little. They demonstrated that it’s working and running 6 to 8 times slower than 1D, and are working to make it run faster. That could be useful in production if it could be made faster. That part of the code should lend itself to vectorization.
3) LArSoft Workshop planning
- Workshop planning and agenda – https://indico.fnal.gov/event/20453/
- The basic plan is a tutorial that starts on the afternoon of the 24th. Second day is multi-threading and vectorization in the morning followed by long-term vision of LArSoft discussion in the afternoon.
- Could ICARUS contribute to long-term vision? Yes.
- And the deep learning people could contribute.
- ICARUS & DUNE should be talking to each other to build the event in a way that allows the possibility of sequentially processing blocks of low-level event data, such as raw digits, on the scale of a single TPC, while limiting the memory footprint to that block.
ICARUS – Tracy
- Have a lot of stuff to get done and not a lot of time left. Basic structure is in place. Need more people.
- Deep learning people at SLAC are using ProtoDUNE and ICARUS data to develop DL algorithms. In the next six months, there should be feedback from that effort that could be useful in long-term strategy.
- We are running simulations, Pandora, and light reconstruction, generally the full simulation and reconstruction workflow. The pathway works. Need to get serious about being ready for data taking. Have hooks in place for noise model. Doesn’t leave time for thinking about larger-scale problems. Optical reconstruction could be interesting. The one imported from MicroBooNE is basic, but the potential is there for something more sophisticated.
DUNE – Tom
- Personally working on GArSoft more than LArSoft last few months.
- ProtoDUNE analysis are coming along well. We don’t simulate electron diverters. The space charge parameterization is smooth, has an analytic smoothing function, doesn’t have sharp corners, as occurs near the diverters, so the edges are all wrong. But it’s probably on us to provide a map.
- The photon libraries were gigantic, Alex Himmel has a solution. A nearly fully parameterized solution. (This was presented at a LCM.)
- Data chunking (processing low-level data one TPC at a time in order to minimize the memory footprint) is important for the future. Data preparation has to work on one piece of data at a time. Framework is a waterfall workflow, not designed to loop over pieces in place. Can code it outside of art, but would like to have the framework do the work.
- Another complication is that due to various DAQ changes, it would be convenient to loop data products with different labels. But doing this in art causes all of the data to be read in, which make data chunking impossible.
- A DUNE problem, not LArSoft, is that the Geometry isn’t the best. Field cages are bigger than APAs, box bounded by field cage is bigger than the box bounded by the wires. We’re not simulating that stuff.
MicroBooNE – Herb
- Nothing to report.