Northstar, 30 September: fast growth, fragile arrivals, a search for CPU headroom

Between the first and last daily readings, Northstar grew from about 1.11 million to about 1.35 million simulated residents. The day had several phases: quick expansion, slower growth, an evening slump, a recovery and a renewed decline. The cause of the slump remains open. The CPU target set for the evening, a sustained low load at more than two million residents, was not reached; a short trial at a lower population did get below 30 percent.

The series has documented gaps, so the endpoints span nearly a full day but are not an exact midnight balance. The population peaked in the evening and had fallen again by the last point. Labour-market values exist for only some of the population points. Within that window the city gained jobs while vacancies fell and unemployment rose. Residents, labour force and advertised jobs do not grow at the same rate, and the table alone identifies no single bottleneck.

New districts need infrastructure before move-in

The building agent continued the growth pipeline with further sections in the northern west field and at the southern edge. Later came several eastern areas, a lakeside section and more land for housing and services. Roads, water, power connections, schools and burial capacity were added before or alongside zoning. Zone cells named in the build logs are designated building land, not finished homes or residents.

Shortly after midnight the southern feeders and the east gate came under pressure. Measures came in stages: new external connections, ramps, road widening and changes to selected traffic lights. The load on one ramp was later reported as lower than before, while the southern path stayed heavily loaded. Around midday another northern connection took traffic, and lower loads were reported on parts of the old main route. These local before-and-after observations are indications, not controlled proof of a citywide improvement; traffic volume, time of day and the growing housing stock changed as well.

For power, additional high-voltage links and three gas power plants were logged. The last structured reading shows generation well above consumption, both in the unchanged internal unit "Spielwert", which is not installed capacity in megawatts. Water and sewage were extended with pumps and treatment plants, and school reserves received building extensions. Forecasts in the plans for a larger population are distinct from current consumption.

More housing did not automatically fix the evening slump

During the day the building agent reported empty office space with no office demand. Part of it was rezoned to mixed residential use. That creates places to live but does not remove the need for suitable jobs. Commercial zoning and a pilot with cheaper rental housing followed. In one rezoning an existing office building fell inside the affected area, and the building agent logged the reversal. This failed attempt belongs in the account; it is not a successful job expansion.

In the evening, move-ins recorded in the operating statistics dropped sharply while households willing to move in accumulated and some new districts lost occupancy. An expert analysis read this as a disturbance in the last step of moving in, pointing to vacant homes and jumps in growth after loading a save. It also discussed traffic routes, the labour market and higher housing costs as competing or contributing explanations. The exact cause remains open. The timing next to a load event proves neither a specific game defect nor the exclusion of other influences.

The population series shows a low of about 1.331 million at 19:23 UTC and then a recovery to above 1.363 million, after which the decline resumed. A restart or an extra building section should therefore not count as a lasting fix because of a short-term surge. Toward the end, according to the operating plan, garbage truck fleets at recycling centres were enlarged and further housing and labour-market measures were prepared. No completed check of their effect existed at the end of the day.

Saving CPU while keeping the simulation

Around 17:00 UTC a human target sharpened the work plan: growth beyond two million residents together with a demonstrated CPU load below 30 percent, at normal speed and without deliberately causing population loss. The verification plan calls for a longer window with sufficient measurement coverage, covering both the game process and the whole system. This target was not met at the end of the day. Short trials below the target population do not replace the planned sustained proof.

The first affinity series shows the conflict of goals. Restricting the game to 12 cores cut CPU time per 1,000 frames to a little more than half, but simulation throughput fell clearly compared with the configuration using all 32 logical processors. A lower CPU display alone therefore does not satisfy the optimisation requirement.

Later trials combined affinity with higher priorities for selected threads. One window with twelve released logical processors kept throughput close to the 32-processor level, with both process and system-wide load below 30 percent. The first 60 seconds of the trial were left out of the CPU values, following the existing evaluation; frame throughput was computed over the full window.

The caveat matters: during that trial the population fell slightly. The other variants were also measured one after another under changing load and population. The numbers hold for their short windows but do not yet show a lasting improvement independent of population. By the end the controller was back at 32 threads, according to the operating plan, because throughput dropped under workload with narrower affinity. Start arguments for the worker count, according to the building agent, also did not give the hoped-for read-back result.

Restarts: a protective rule under examination

Frequent precautionary restarts had repeatedly interrupted growth. At a human suggestion, expert reviews examined the existing memory limit. They distinguished short memory peaks from the lasting baseline and weighted the risk differently. The small number of historical crashes, differing run times and the earlier operation at speed 4 allow no hard safe limit to be derived.

The building agent then began an extended test at normal speed. Later a restart took place at a logged baseline of about 16.1 GiB. That is a tested operating step, not proof that any load below is safe or above is unsafe. Recovery after loading was also limited in time.

Status and open points

For the contractual performance series, qualified guard readings are still missing for 30 September; the trial windows above do not replace it.