[gvegayon]

Pins 1b10904d8fb11e4d775a9eca154305f0b6cd9e22 (UofUEpiBio/epiworld#341).

The stage covers two epiworld changes since 3d-1:

• 3d-2 (Phase 3d-2: measles models on a mixing layer epiworld#340, #345) ports the measles models, which these scenarios don't use.
• 3d-1b (3d-1b: build network layers without maps (#339) epiworld#341) builds networks without std::map. It's the fix for the 3d-1 finding, 1.0 network layers: building from an edge list is 5-9x slower and uses up to 2.6x the memory of 0.x (AdjList) epiworld#339.

The runners are unchanged: the AdjList constructor they use kept its signature. Run in the same image as 3d-1 (9a51e13558f5), both C++ engines, replicates interleaved. make container-check and make container-smoke pass; all 1,280 runs ok.

Rows

Each cell is 0.x, then 1.0. Simulate and build in seconds, peak in MiB.

Scenario	n	Simulate	Ratio	Build	Ratio	Peak	Ratio	Attack rate
00	10,000	0.00538, 0.00547	1.02	0.00149, 0.00131	0.88	8, 8	0.98	0.385, 0.385
00	100,000	0.00925, 0.00985	1.06	0.0171, 0.0122	0.71	38, 33	0.85	0.060, 0.060
01	10,000	0.00248, 0.00249	1.00	0.00133, 0.00112	0.85	8, 8	0.98	0.119, 0.119
01	100,000	0.00937, 0.00934	1.00	0.0133, 0.011	0.82	43, 41	0.94	0.014, 0.014
02	10,000	0.00236, 0.00245	1.04	0.00129, 0.00111	0.86	9, 8	0.99	0.119, 0.119
02	100,000	0.009, 0.00921	1.02	0.0136, 0.0107	0.79	44, 41	0.94	0.014, 0.014
03	1,000,000	0.0326, 0.0306	0.94	0.172, 0.137	0.80	352, 276	0.78	0.006, 0.006
04	165,865	0.185, 0.175	0.95	0.0166, 0.015	0.90	90, 88	0.97	0.622, 0.622

Against 3d-1 (the 10% rule)

• Build time and peak memory fell in every row. The build ratio went from 5.2–9.0 to 0.71–0.90, and the peak ratio from 1.06–2.56 to 0.78–0.99.
• Attack rates are the same as 0.x's and as at 3d-1.
• 1,000,000 agents is no longer noisy: the simulation ratio is 0.94, against 1.58 at 3d-1.
• The simulation ratio grew by more than 10% in every scenario up to 165,865 agents: to 1.00–1.06 from 0.69–0.97, and to 0.95 from 0.81 on GeoPops. I profiled it before anything else. The time per step didn't change; page faults moved from the build into the run.
  • The cause: the 3d-1 build freed tens of MiB of map nodes that glibc kept, and the run's first allocations reused them.
  • The measurement: I built scenario_01's runner at 100,000 agents against three commits (40 interleaved replicates in this image). It simulates in 0.0069 s at 3d-1 (5811a28), 0.0070 s after 3d-2 (b699adb), and 0.0101 s now.
  • The fault counts: inside run() it takes 0 minor page faults before #341 and about 3,400 (13 MiB) after. At under a microsecond each, that accounts for the difference.
  • So: the 3d-1 ratios below 1 were the same artifact, and the engines now simulate at parity. Build plus simulation at 100,000 agents takes 0.020 s, against 0.104 s at 3d-1. Nothing to fix in epiworld.

The note is under "3d-1b" in docs/epiworld-v1.md.

🤖 Generated with Claude Code