P0089R0: On Quantifying Memory-Allocator Strategies Page 1 of 55 Doc No: P0089R0 Date: 2015-09-28

Performance requirements drive many of our most difficult design choices. In memory management, such choices can have surprising and far-reaching effects. Although performance of global memory allocators has improved markedly in recent years, use of local memory allocators can provide significant (sometimes even dramatic) benefits in commonly encountered circumstances we have tried to identify here. To make reasoned choices on the use of local memory allocators, we need to understand where and how their use may affect runtime performance. We have identified several measures of how systems can stress a global allocator, and may benefit by applying a well-chosen local allocator in its place. If we are to choose wisely where and how to apply a local allocator, we need objective measurements. We have identified several usage patterns which we have encoded into benchmarks to identify precisely where local allocators do (and where they do not) provide substantial benefits. This paper presents our results with limited analysis to help support informed discussion. Possibly the most significant result is that, where use of a local allocator does yield dramatic improvements, the number of operations are about the same: The slower benchmark run times for the global allocator are dominated by stalls waiting on cache interactions with main memory (due to a severe lack of physical and temporal locality); the ability to use a local allocator empowers us to act to avoid such stalls. Implementations of standard allocators (and others) are freely available today – accompanied by copious usage examples – in Bloomberg’s open-source distribution of the BDE library a