EDM Memory usage

[paulgessinger]

The peak memory consumption of the track state EDM strongly increases when pileup and even complexity is increased. For ttbar events at $\mu=200$ from Pythia8, it can go up to a few gigabytes per event, which is suboptimal.

This issue tracks progress to improve and optimize the memory consumption.

Associated PRs:

• 1507

  This helps centralize the MultiTrajectory usage: a single instance is used by the examples algorithms to accumulate tracks from all seeds

• 1512

  Switches from overallocating vectors and matrices for the measurements to using a jagged vector structure that looks like this:

  | x x x | x | x x |  ...
   ^       ^     ^
  M1,D=3 M2,D=1 M3,D=2

• 1511

  Implements statistics on the memory usage of Acts::VectorMultiTrajectory to give us an idea of what's going on.

Status quo

The tooling from #1511 gives the following measurements from MultiTrajectory:

ODD full chain example with Pythia8 ttbar @ mu200

other:
               count: 1877425.000000
               index:   128.91M
             parPred:    85.94M
             covPred:   515.65M
             parFilt:     0.00M
             covFilt:     0.00M
             parSmth:    85.94M
             covSmth:   515.65M
                meas:    25.08M
             measCov:   150.46M
                 jac:   515.65M
         sourceLinks:    18.50M
          projectors:     4.18M
               total:  2045.97M
meas:
               count: 653022.00
               index:    44.84M
             parPred:    27.47M
             covPred:   164.82M
             parFilt:    29.89M
             covFilt:   179.36M
             parSmth:    29.89M
             covSmth:   179.36M
                meas:    29.89M
             measCov:   179.36M
                 jac:   164.82M
         sourceLinks:     9.96M
          projectors:     4.98M
               total:  1044.66M

ODD truth tracking with Pythia8 ttbar @ mu200

               count: 40146.000000
               index:     2.76M
             parPred:     1.84M
             covPred:    11.03M
             parFilt:     0.00M
             covFilt:     0.00M
             parSmth:     1.84M
             covSmth:    11.03M
                meas:     0.00M
             measCov:     0.00M
                 jac:    11.03M
         sourceLinks:     0.31M
          projectors:     0.00M
               total:    39.82M
meas:
               count: 37159.00
               index:     2.55M
             parPred:     1.70M
             covPred:    10.21M
             parFilt:     1.70M
             covFilt:    10.21M
             parSmth:     1.70M
             covSmth:    10.21M
                meas:     1.70M
             measCov:    10.21M
                 jac:    10.21M
         sourceLinks:     0.57M
          projectors:     0.28M
               total:    61.24M

Other ideas

• [ ] Switch to float for storage of parameters and covariances to float
• [ ] Only store lower left triangle of covariance matrices. @tboldagh points out there is actually a triangular view directly in Eigen

[stale[bot]]

This issue/PR has been automatically marked as stale because it has not had recent activity. The stale label will be removed if any interaction occurs.

[github-actions[bot]]

This issue/PR has been automatically marked as stale because it has not had recent activity. The stale label will be removed if any interaction occurs.