Ptlbmalloc2: Reducing TLB Shootdowns with High Memory Efficiency

Abstract

The cost of TLB consistency is steadily increasing as we evolve towards ever more parallel and consolidated systems. In many cases the application memory allocator is responsible for much of this cost. Existing allocators to our knowledge universally address this issue by sacrificing memory efficiency. This paper shows that such trade-offs are not necessary by presenting a novel memory allocator that exhibits both excellent memory efficiency and (TLB) scalability: ptlbmalloc2. First, we show that TLB consistency is becoming a major scalability bottleneck on modern systems. Next, we describe why existing memory allocators are unsatisfactory regarding this issue. Finally, we present and evaluate ptlbmalloc2, which has been implemeted as a library on top of glibc. Ptlbmalloc2 outperforms glibc by up to 70% in terms of cycles and execution time with a negligible impact on memory efficiency for real-world workloads. These results provide a strong incentive to rethink memory allocator scalability in the current era of many-core NUMA systems and cloud computing.