Reference Baselines
Tree sizes that a given engine produces for given inputs. Use them to check a build, and to tell which engine state a published number belongs to.
A tree size is determined by the input deck and the Input/ tree, the truncation time, the probability cutoff in Input/Scheduler/ZCtrlPar.txt, the branch floor, and the engine. It does not depend on serial or MPI execution, the rank count, the restart mode, compression, or the pivotal-event layout.
Two engine states
| Engine | How to get it | Behaviour at a branch point |
|---|---|---|
| Current | This branch | A child sequence restarts from the branch point and receives the content of its own alternative (the control value, the action time, the component failure, the procedure step) |
legacy-baseline-2026-08-18 | git checkout legacy-baseline-2026-08-18 (commit 5afb952, 2026-08-18) | A child sequence restarts from the branch point without that content. The alternative is recorded in the pivotal-event node; the child runs as the mainline does |
Every tree size, timing and scaling figure published before 2026-08-28 belongs to the legacy baseline. On the current engine, ADS_RESTART_INJECT=0 selects the legacy behaviour at a branch point; the BV4 rows below that carry this setting were measured that way.
Current engine
Measured 2026-09-18 on commit 85a93de plus the working tree, macOS 26.7 on Apple M4 Pro, Apple Clang with gfortran 16.2.0, Open MPI 5.0.10, Release build.
| Deck | Truncation time (s) | Cutoff | Settings | Sequences (ZESNode) | PE nodes | KPS sets | ES_summary lines |
|---|---|---|---|---|---|---|---|
tests/BV4 | 15.5 | 5E-5 | serial | 36 | 97 | 36 | 37 |
tests/BV4 | 15.5 | 5E-5 | mpirun -n 3 | 36 | 97 | 36 | 37 |
tests/BV4 | 15.5 | 5E-5 | serial, ADS_RESTART_INJECT=0 | 20 | 60 | 20 | 21 |
tests/BV4 | 100 | 5E-5 | serial, --memory-limit 4096 | 514 | 2,628 | 514 | 515 |
tests/BV4 | 100 | 5E-5 | serial, ADS_RESTART_INJECT=0 | 57 | 269 | 57 | 58 |
The first row is what ctest, the packaged run-test.sh and Expected Counts check.
Cluster runs of the designed-tree deck studies/sgtr-tree (cutoff 1E-5, ADS_BRANCH_FLOOR=0.1, ADS_BRANCH_FLOOR_SINGLE=0.5, --aggregate-pe, --compress-queue, ADS_RESTART_TRIM=1; 2026-08-28 to 2026-09-04):
| Truncation time (s) | Sequences | Note |
|---|---|---|
| 1000 | about 5,090 | 5,047 to 5,090 over the recorded runs |
| 3600 | 34,455 to 35,613 | Size varies by about 3% between runs of identical inputs, because recovery forks in this deck are sensitive to timing |
legacy-baseline-2026-08-18
| Deck | Truncation time (s) | Cutoff | Sequences | PE nodes | Recorded |
|---|---|---|---|---|---|
tests/BV4 | 15.5 | 5E-5 | 20 | 60 | test suite, 2026-03 to 2026-08 |
tests/BV4 | 50 | 5E-5 | 46 | benchmark matrix, 2026-03 | |
tests/BV4 | 100 | 5E-5 | 57 | 269 | benchmark matrix, 2026-03 |
tests/BV4 | 150 | 5E-5 | 66 | benchmark matrix, 2026-03 | |
tests/BV4 | 200 | 5E-5 | 76 | benchmark matrix, 2026-03 | |
tests/BV4 | 300 | 5E-5 | 83 | benchmark matrix, 2026-03 | |
tests/BV4 | 1000 | 5E-5 | 883 | 7,260 | long checkpoint tests |
studies/sgtr-10k | 1000 | 1E-5 | 4,020 | strong-scaling ladder, 2026-08 | |
studies/sgtr-10k | 10000 | 1E-5 | 12,306 | 168,702 | 768 ranks, 2026-08; about 376 GB of output |
The six ENABLE_LONG_CHECKPOINT_TESTS tests expect the BV4 T=1000 row and therefore apply to the tag, not to the current engine.
Reproducing a row
cd build/bin
./adsidac --project-dir=../tests/BV4 --input-deck=../tests/BV4/BV4.i --truncation-time=<T>
find ../tests/BV4/Output/ES -name 'ZESNode*' | wc -l
python3 ../../tests/scripts/pe_canonical.py ../tests/BV4/Output/PE | wc -lRuns longer than the BV4 default grow quickly on the current engine. Use --memory-limit on a workstation (see Choosing performance options).