Tiered memory systems, which integrate heterogeneous memory components, offer a promising solution to balance capacity and efficiency. While Transparent Huge Pages (THP) are essential for reducing TLB misses and improving performance, they introduce significant hot bloat in tiered memory. This phenomenon causes cold sub-pages to reside on hot huge pages, occupying scarce DRAM resource and forcing performance-critical data into slower tiers, thereby degrading the Quality of Service (QoS). Existing mechanisms to this problem, such as MEMTIS, employ coarse-grained thresholds and lack the sensitivity required for the highly dynamic workloads, then it often yields suboptimal performance under varying memory configurations. To address this phenomenon, we introduce AHPS, an adaptive huge page splitting framework. AHPS continuously monitors key performance indicators to characterize the fast-tier hit ratio over both long-term and short-term intervals. By employing a variable-size sliding window, it accurately captures workload fluctuations and quantifies deviations from the stable state. Furthermore, AHPS dynamically regulates the number of huge pages to split based on memory access skewness, available memory capacity, and real-time hit ratios. Extensive evaluations on memory-intensive benchmarks demonstrate that AHPS consistently outperforms the state-of-the-art solution MEMTIS, delivering performance improvements of 10% to 27% across various workloads.
Wang et al. (Tue,) studied this question.