TLAB (thread-local allocation buffer)

With many threads allocating from one heap, a shared allocation cursor would need a lock or an atomic operation on every allocation. A TLAB removes that cost: each thread reserves a private chunk of the heap in one shared-state operation, then bump-allocates inside it with no synchronization at all. Only when the chunk runs out does the thread go back to shared state for a refill.

The trade-off is waste: the unused tail of each thread's TLAB is memory no other thread can use until the TLAB is returned.

In OrionGC the TLAB is the AllocContext stored on each scheduler thread (a single static context serves instead before the scheduler runs, or when it is compiled out), the default TLAB size is 8 KiB, and every collection starts by returning all TLABs so the sweep sees a consistent heap. See AllocContext, TLAB refill and Returning TLABs.