Object Lifetime Models and Their Effect on Performance and Memory Fragmentation in Specialized C++ Allocators

Abstract

Modern C++ applications place heterogeneous demands on dynamic memory management. Allocator throughput and memory consumption depend on the workload request profile and on the lifetime model of allocated objects. This work presents an empirical study of four allocators: the system malloc, a monotonic arena, a fixed-size pool allocator, and a segregated free list. A matrix of 39 controlled scenarios is built by factorial variation of the lifetime model (long-lived, FIFO, LIFO, and random order), and the workload parameters: fixed small size, variable size, and churn. Each scenario records throughput and the ratio of reserved memory to peak live allocation volume. Additional experiments at 10× scale and at a reuse fraction of 0.9 expose the stability limits of each strategy. Specialized allocators outperform malloc in throughput by one to two orders of magnitude when the lifetime assumptions of their architecture hold. The pool allocator delivers 86.4 million operations per second on long-lived, fixed-size objects with an overhead coefficient of 1.000× and sustains 81.4 million operations under a 10× load increase. The segregated free list exposes a gap between throughput and memory efficiency: overhead reaches 236.5× under heavy churn, and the 10× scale test pushes the system into a fallback regime, with 2,002,194 capacity-exhaustion operations. A controlled ideal_fit_32 experiment confirms that making all requests fit a single size class does not reduce the overhead of an allocator that eagerly preallocates all size classes. The results justify treating the object lifetime model as a primary design axis for allocators.