Compaction and pause budgets: pay for movement deliberately

Consider sixteen separated free regions of two MiB each. Their total is 32 MiB, but they do not provide one contiguous eight-MiB region if the allocator cannot combine them or obtain more address space. The actual outcome depends on the runtime's allocation regions, large-object strategy, available mappings, and collection mechanism. This example is a shape argument, not a prediction that every allocator with those free bytes must fail. Internal fragmentation, such as unused bytes inside size-class slots, is a separate accounting category.

A moving collector can pack selected survivors together and make larger free regions available. It must preserve references to the moved objects, including roots and fields that can still be used. Pinning prevents an object from moving for a defined interval, often to support a native operation holding a stable address. Pinned objects can constrain compaction, but their impact depends on placement and the collector's design. Keeping an object alive and keeping its address stable are different obligations; rooting alone does not imply pinning in a moving runtime.

For a simplified stop-the-world phase, separate coordination time, live bytes moved, and reference-update work. Divide each work quantity by an appropriate measured service rate, then add only components assumed to execute serially in the model. A bytes-per-second rate must specify whether it counts logical copied payload or physical read-plus-write traffic. Parallel workers, bandwidth contention, cache effects, and overlapping phases can invalidate a naive sum. Use measurements from comparable heap shapes and load, not a device's advertised maximum bandwidth.

A collector's pause target is not automatically a hard application deadline. Coordination can wait for safe execution points, a workload can change the survivor set, and a concurrent phase can consume CPU while requests continue running. Report global pauses separately from concurrent CPU time and throughput. Do not add unrelated component p99 values and call their sum the end-to-end p99: quantiles do not generally compose that way. Trace actual requests through the relevant intervals and inspect how bursts, queueing, and scheduling amplify or hide a pause.

First ask whether retained objects are useful. Reducing unnecessary survivors can reduce tracing and movement while improving memory headroom. A larger heap may allow less frequent collection, yet retain more data or produce larger work batches; the tradeoff depends on the collector and application. Reducing pin duration or copying a small native input can sometimes make relocation easier, but an additional copy also has a cost. Compare alternatives under the same request distribution, memory limit, and latency objective, including rare large batches.

The queue lab shows service-rate imbalance rather than actual collector pauses. Press Reset, set arrivals to four and service to two, predict two extra waiting jobs per second, then Play or Step. For a separate experiment, press Reset before setting arrivals to four and service to five; predict the restored four-job initial backlog drains in four seconds, then Play or Step. This can illustrate how reduced effective service capacity accumulates work. The controls do not create periodic stop-the-world intervals, pinned regions, or relocation. In a real service, measure the backlog immediately before and after pauses and the subsequent recovery rate. That evidence connects a collection event to user delay more directly than a heap-size setting alone.