Generational collection: remember edges across the boundary

Many workloads allocate objects that soon become unreachable, motivating a young allocation region that can be collected frequently. Objects surviving selected collections may be promoted according to the runtime's policy. This is a workload hypothesis and implementation strategy, not a theorem that every newer object dies sooner than every older one. Long-lived input batches, cached request objects, and persistent graph construction can produce very different survivor patterns. Measure age and retention rather than treating promotion as an application-level measure of usefulness.

Suppose a root reaches an old object A, and A contains the only reference to young object B. A collector scanning only young roots and young objects would miss B if it ignored A's edge. Scanning the entire old heap would recover correctness but defeat much of the partial-collection benefit. A remembered set records relevant edges or conservative source regions. The young collection then treats their young targets as additional starting points and traces within its collection scope. A false-positive remembered entry can add work; losing a required live edge can corrupt the program.

A write barrier is runtime work associated with a reference mutation. In a card-table design, the heap is divided into fixed-size address regions called cards; a store can mark the source card dirty. Later scanning that card discovers relevant references. This stores less information than recording every target exactly, at the cost of inspecting additional bytes or objects. Other runtimes use different remembered-set representations and filtering. A barrier may be inlined, batched, or optimized when the compiler can prove the recording unnecessary.

Generational barriers and concurrent-marking barriers address different invariants, though one implementation may combine their work. During concurrent tracing, a mutator can change edges while marking progresses; a particular algorithm may need to record a newly installed reference, preserve an overwritten reference, or enforce another specified condition. Do not transfer the details of one algorithm to all collectors. At a native-code boundary, writes to managed reference fields must use the runtime's supported interfaces so its required barriers and root tracking are not silently bypassed.

A workload updating many old reference fields can create significant barrier and remembered-set work even with a small allocation rate. A workload retaining most young allocations can make young collections expensive because many objects must be traced or evacuated. Diagnose those mechanisms separately: record allocation rate, survivor or promotion volume, dirty-region work where available, and pause distribution. Reducing the frequency of one phase can shift work elsewhere. A larger young region may reduce collection frequency while increasing the amount considered in a single collection.

The probability lab isolates a hypothetical survivor fraction, not barrier execution. Set n=20 and p=0.10, predicting two survivors on average; increase p to 0.50 and predict ten. This compares equal-sized independent objects in one artificial cohort. The actual controls cannot change generation size, promotion age, card width, or mutation rate, and real object lifetimes are often dependent. Explain why a higher survivor fraction may increase copying or tracing work, then draw the old A→young B edge separately to show a correctness requirement that the probability bars cannot represent.