HTLL: Latency-Aware Scalable Blocking Mutex

作     者：Yu, Ziqu Gu, Jinyu Wu, Zijian Liu, Nian Guo, Jian 

作者机构：Shanghai Jiao Tong Univ Inst Parallel & Distributed Syst IPADS Shanghai 200240 Peoples R China Guangdong OPPO Mobile Telecommun Co Ltd Dongguan 523860 Peoples R China 

出 版 物：《IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS》 (IEEE Trans Parallel Distrib Syst)

年 卷 期：2025年第36卷第3期

页      面：471-486页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China [62202292, 62432010] Fundamental Research Funds for the Central Universities 

主　　题：Throughput Tail Low latency communication Spinning Protocols Libraries Codes Scalability Degradation Accuracy Synchronization primitives lock over-subscription latency-aware 

摘      要：This article finds that existing mutex locks suffer from throughput collapses or latency collapses, or both, in the oversubscribed scenarios where applications create more threads than the CPU core number, e.g., database applications like mysql use per thread per connection. We make an in-depth performance analysis on existing locks and then identify three design rules for the lock primitive to achieve scalable performance in oversubscribed scenarios. First, to achieve ideal throughput, the lock design should keep adequate number of active competitors. Second, the active competitors should be arranged carefully to avoid the lock-holder preemption problem. Third, to meet latency requirements, the lock design should track the latency of each competitor and reorder the competitors according to the latency requirement. We propose a new lock library called HTLL that satisfies these rules and achieves both high throughput and low latency even when the cores are oversubscribed. HTLL only requires minimal human effort (e.g., add several lines of code) to annotate the latency requirement. Evaluation results show that HTLL achieves scalable performance in the oversubscribed scenarios. Specifically, for the real-world database, LMDB, HTLL can reduce the tail latency by up to 97% with only an average 5% degradation in throughput, compared with state-of-the-art alternatives such as Malthusian, CST, and Mutexee locks;In comparison to the widely used pthread_mutex_lock, it can increase the throughput by up to 22% and decrease the latency by up to 80%. Meanwhile, for the under-subscribed scenarios, it also shows comparable performance than state-of-the-art blocking locks.