An adaptive memory allocatorfor better reallocation performance

Currently, the solution for dynamic memory reallocation is often left to a simple malloc-copy-free sequence. However, with the growing size of the memory object, this copying operation will have considerable CPU and memory costs before the content finishes its copy.This report aims to achieve a nearly zero-copy reallocation scheme for reallocations, especially with large memory objects. We want to replace the malloc-copy-free sequence with a better solution to save the performance losses during the copy. Inside the virtual memory system in Linux, there is a possible solution for solving this problem. Mremap is the Linux kernel function to operate on the virtual page table directly. It modifies the page table to achieve the same goal. When reallocating new data sizes, we provide a larger virtual memory chunk to the process and redo the virtual-physical memory linking. Thus, the copy overhead of physical memory is resolved. With comparison experiments we made, we found out the mremap has the potential for better performance compared with the state-of-art memory allocators. We developed this adaptive memory allocator which could achieve 5% performance when dealing with large memory reallocations. With the experiments with Redis, we also obtained performance enhancement in basic, list-based, and large object Redis commands.Electrical Engineering | Embedded System