[offline] Reduce memory footprint for I3Position and I3Direction

[David Schultz]

One thing we should do at some point is to check low-level code optimizations:

• struct packing
• cpu cache usage vs memory layout
  • cache lines
  • data structure layout
  • loop ordering
• profile-guided optimization?
  • expensive, but might be worth it for production

Probably should also run profiling tools over the whole code base and find hot spots to look into more deeply.

[David Schultz]

I investigated struct packing today. Here are some findings:

• Consider making OMKey.stringNumber_ or OMKey.omNumber_ a short.

• I3Position and I3Direction have variables for cartesian and spherical. Eliminate one of them, and save 24 bytes per object (greatly shrinking I3Particle).
  • Above elimination would get rid of isCalculated_, saving 8 bytes because of bool + padding.

• Use c++11 enum in I3Particle to define number of bytes for each (defaults to int, but we could use char for many). Careful of the serialization here.
  • Same to I3DOMStatus, and likely other classes.

• I3Orientation has lots of duplicate orientations cached.

• I3EventHeader has 4 bytes of padding. fixing enum type and changing subRunID_ to short would save 8 bytes.

• Stop using dataclasses/external/tree.hh (lots of padding issues that I don't want to fix).

• I3FlasherInfo seems to waste space, using larger variables than necessary (which causes padding problems).

[David Schultz]

Commits r157654/IceCube and r2958/IceTray fix some alignment/padding issues that were simple and not mentioned above.

[David Schultz]

Attempted a PGO build on a slimmed down offline-software, using the tests as a benchmark. No noticeable different in test time. Likely this is because the unit tests are diverse, checking every possibility. Probably need to run only the production tests in order to make a difference.

Notes on how to PGO build:

CXXFLAGS="-fprofile-dir=$PWD/pgo -fprofile-generate=$PWD/pgo" cmake -DCMAKE_BUILD_TYPE=Release ../src
make all test-bins
./env-shell.sh make test
make clean
rm -rf CMakeCache.txt CMakeFiles
CXXFLAGS="-fprofile-dir=$PWD/pgo -fprofile-use=$PWD/pgo -fprofile-correction" cmake -DCMAKE_BUILD_TYPE=Release ../src
make

Note that -j cannot be used, a serial make is required.

[Alex Olivas]

Want memory/resource tracking in cicada and benchmark off the Winter 2019 release.

[Alex Olivas]

Merged mkarl's I3Position branch to the trunk, which should reduce the memory footprint of I3Position (and by extension I3Particle). Next up are the other classes: I3Direction, I3Orientation, etc...

[Alex Olivas]

Milestone Summer Solstice 2020 deleted

[Alex Olivas]

One issue was addressed so far, but the rest were broken into several tickets and scheduled for future milestones.