Combining Machine Learning and Lifetime-Based Resource Management for Memory Allocation and Beyond

Memory management is a decades-old research area24 that is fundamental to the performance of all applications. On modern architectures, memory managers determine a workload's ability to use 2MB (and 1GB) huge pages instead of traditional 4KB pages. The use of huge pages is crucial for performance on modern servers since they substantially reduce the cost of address translation by producing a wider reach in Translation Lookaside Buffers (TLB), reducing misses on the CPU's critical path.5 Current huge page-aware memory managers13 trade-off huge page usage with memory utilization, breaking up huge pages when they become inefficient. Figure 1 visualizes the source of this trade-off: When a C program allocates memory, it calls into a memory allocator library (e.g., TCMalloc13), which places the object at a particular address in memory until the program deletes it. The object may not move.