Local ('Arena') Memory Allocators (part 2 of 2)
Are allocators worth the trouble? What situations merit their use? How are they applied effectually? What’s the performance impact? This practical talk by large scale C++ expert Dr. John Lakos demonstrates that having allocators in your tool box may lead to orders of magnitude speed improvements. The runtime implications of the physical location of allocated memory is often overlooked, even in the most performance critical code. In this talk, we will examine how the performance of systems can degrade when using `new`/`delete` and `std::allocator` . We will contrast these global allocators, which allocate memory globally for a system, with local allocators that each allocate memory for a proper subset of objects in the system. We will also demonstrate how local allocators can reduce or entirely prevent the degradation seen in systems that rely on the global allocator. Six dimensions – fragmentability, allocation density, variation, locality, utilization, and contention – will be introduced to depict the potential for performance penalties and aid the listener in determining which local allocator will offer the best performance in their subsystems. Evidence will be presented that identifying these dimensions, and selecting a local allocator based upon them, can lead to *order-of-magnitude* reductions in run time compared to systems using a global allocator.