Modern workloads are demanding increasingly larger memory capacity. Compute Express Link (CXL)-based memory tiering has emerged as a promising solution for addressing this trend by utilizing traditional DRAM alongside slow-tier CXL-memory devices in the same system. Unfortunately, most prior tiering systems are recency-based, which cannot accurately identify hot and cold pages, since a recently accessed page is not necessarily a hot page. On the other hand, more accurate frequency-based systems suffer from high memory and runtime overhead as a result of tracking large memories. In this paper, we propose FreqTier, a fast and accurate frequency-based tiering system for CXL memory. We observe that memory tiering systems can tolerate a small amount of tracking inaccuracy without compromising the overall application performance. Based on this observation, FreqTier probabilistically tracks the access frequency of each page, enabling accurate identification of hot and cold pages while maintaining minimal memory overhead. Finally, FreqTier intelligently adjusts the intensity of tiering operations based on the application's memory access behavior, thereby significantly reducing the amount of migration traffic and application interference. We evaluate FreqTier on two emulated CXL memory devices with different bandwidths. On the high bandwidth CXL device, FreqTier can outperform state-of-the-art tiering systems while using 4$\times$ less local DRAM memory for in-memory caching workloads. On GAP graph analytics and XGBoost workloads with 1:32 local DRAM to CXL-memory ratio, FreqTier outperforms prior works by 1.04$-$2.04$\times$ (1.39$\times$ on average). Even on the low bandwidth CXL device, FreqTier outperforms AutoNUMA by 1.14$\times$ on average.

翻译：现代工作负载对内存容量的需求日益增长。基于计算快速链路（CXL）的内存分级技术通过在同一系统内同时使用传统DRAM与慢速CXL内存设备，为应对这一趋势提供了颇具前景的解决方案。然而，现有分级系统大多基于最近访问时间策略，由于最近被访问的页面未必是热页面，因此难以准确识别冷热页面。相较之下，更精确的基于频次策略的系统因需追踪大规模内存而面临高昂的内存与运行时开销。本文提出FreqTier——一种面向CXL内存的快速精确频次分级系统。我们发现内存分级系统可容忍少量追踪误差而不影响应用整体性能。基于这一发现，FreqTier通过概率追踪各页面访问频次，在保持极低内存开销的同时实现对冷热页面的精准识别。此外，FreqTier能够根据应用的内存访问模式智能调节分级操作强度，从而显著减少页面迁移流量与应用干扰。我们在两个不同带宽的模拟CXL内存设备上评估了FreqTier。在高带宽CXL设备上，处理内存中缓存类工作负载时，FreqTier在性能超越现有最优分级系统的同时，本地DRAM内存使用量降低至其四分之一。在1:32本地DRAM与CXL内存配比条件下，于GAP图分析与XGBoost工作负载中，FreqTier相较先前方案实现1.04-2.04倍（平均1.39倍）性能提升。即使在低带宽CXL设备上，FreqTier平均性能仍优于AutoNUMA达1.14倍。