Source-linked AI summary
Mole: Tier-Specific Hotness Profiling Driven Memory Tiering for Multi-Tiered Memory Systems
Mingyang Liu, Congming Gao, Xufeng Yang, Fang Wu, Youmin Chen, Renhui Chen, Jiwu Shu
TL;DR
Multi-tiered memory faces migration failures when greedy promotion and stepwise demotion compete for limited tier capacity. Mole separates their destinations and profiles tiers at different frequencies; evaluation on a four-tier system reports that it outperforms state-of-the-art tiering systems.
Problem
Method
Mole combines staging demotion to the lowest tier, hotness-matched promotion, and tier-specific profiling to detect reactivated pages and identify cold pages.
Source: §1 · p. 2Results
On a four-tier memory system, Mole outperforms state-of-the-art tiering systems.
Source: §1 · p. 2, §8 · p. 13Takeaways & Limitations
Separating migration destinations preserves upper and intermediate tiers for promotions while the lowest tier stages demoted pages that may later become hot.
Source: §1 · p. 2, §8 · p. 13Tier-specific hint-page-fault profiling requires periodic virtual-address scans to check page locations and whether previously unmapped pages remain unaccessed.
Source: §4.2 · p. 5
Abstract
from arXiv · showhide
Multi-tiered memory systems combine fast, small upper tiers with slow, large lower tiers to improve performance and cost efficiency. Existing designs, such as AutoTiering and MTM, rely on greedy promotion and stepwise demotion, which can intensify contention for limited capacity in faster tiers. We observe that promotions directly improve performance, whereas demotions primarily reclaim space. Based on this asymmetry, we propose Mole, a memory tiering system that separates promotion and demotion destinations. Mole employs staging demotion to move cold pages directly to the lowest tier, bypassing intermediate tiers, and targeted promotion to place hot pages in tiers that match their current hotness. The lowest tier serves as a staging area from which pages can be promoted when they become hot again. This separation preserves intermediate-tier capacity for promotions, reducing tier contention and migration failures. However, staging demotion requires timely identification of pages that become hot after demotion. To meet this requirement with low profiling overhead, Mole uses tier-specific profiling: it profiles the lowest tier at high frequency to detect reactivated hot pages, while profiling upper tiers at lower frequency to identify cold pages. Experimental results show that Mole reduces migration failures and improves performance under dynamic access patterns.
1 Introduction
Multi-tiered systems must place pages according to hotness, but greedy promotion and stepwise demotion can intensify competition for limited capacity. Mole separates promotion and demotion destinations and varies profiling frequency by tier to reduce contention while detecting reactivated pages in time.
1 Introduction
- Multi-tiered systems place frequently accessed pages in fast, small upper tiers and cold pages in slow, large lower tiers, with page migration generating tier-transit traffic.Emerging technologies such as persistent memory and CXL-attached memory are contributing to hierarchies with more tiers.Source: §1 · p. 1
- Existing systems such as AutoTiering and MTM use greedy promotion toward the top tier and stepwise demotion to the first tier with available capacity.These policies can increase demand for limited faster-tier capacity.Source: §1 · pp. 1–2
- Greedy promotion and stepwise demotion can intensify tier contention, causing migration failures even when many hot pages are correctly identified.Under saturated conditions, promotion failure rates exceed 97% for Silo and Liblinear, while demotion failures reach 37.9% and 48.2%, respectively.Source: §1 · pp. 1–2
- Mole sends cold pages directly to the lowest tier and targets hot pages to tiers matching their current hotness, preserving upper-tier capacity for promotions.The lowest tier acts as a staging area for pages that may later warrant promotion.Source: §1 · p. 2
- Mole profiles the lowest tier frequently to detect reactivated hot pages, while lower-frequency profiling in upper tiers identifies cold pages with less overhead.Reactivated pages can become hot again after demotion, while upper-tier cold pages have long access intervals and tolerate moderate detection delays.Source: §1 · p. 2
- Existing profiling systems commonly use the same frequency across tiers, coupling lowest-tier detection latency with profiling overhead in upper tiers.Some prior systems profile only lower or remote tiers but do not support scalable, per-tier frequency control.Source: §1 · p. 2
2 Background and Related Works
Multi-tiered memory combines heterogeneous tiers, and page placement depends on profiling hotness and managing migration paths. Prior systems use several profiling mechanisms and extend two-tier migration policies, but their multi-tier policies can leave intermediate tiers as contention points.
2.1 Multi-Tiered Memory System
- The paper studies a four-tier hierarchy on a dual-socket machine, exposing each socket’s memory as separate NUMA nodes and binding application threads to one socket.The setup isolates page placement from thread scheduling and cross-socket effects.Source: §2.1 · pp. 2–3
2.2 Tiered Memory Management
- Tiered memory management combines page-hotness profiling, which detects access patterns, with migration-path management, which moves pages among tiers.The section identifies three popular profiling mechanisms: hint page faults, PTE access bits, and hardware performance counters.Source: §2.2 · p. 3, §2.2.1 · p. 3
- Hint-page-fault profiling makes selected virtual addresses inaccessible and captures subsequent accesses through minor page faults; scans also measure how long pages remain unaccessed.In this paper, “unmap” refers to changing PTE permissions, not page reclamation.Source: §2.2.1 · p. 3
- PTE access-bit profiling periodically clears and scans access bits to estimate hotness, while hardware-counter sampling can be limited by sampling coverage and platform support.Access-bit sampling gives coarse-grained information because it records only whether a page was accessed since the prior scan.Source: §2.2.1 · p. 3
- AutoTiering and MTM extend two-tier greedy promotion and stepwise demotion to multi-tier systems, but intermediate tiers can remain contention points under capacity pressure.Proactive demotion in two-tier systems does not directly solve multi-tier placement because multiple demotion destinations exist.Source: §2.2.2 · p. 3
3 Limitations of Existing Solutions
Existing multi-tier systems face two distinct limits: capacity pressure can cause migration failures, while uniform profiling rates fail to balance detection needs and overhead.
3.1 Page Migration Failures in Multi-tiered Systems
- Across both capacity configurations, promotion failures reach 97.4% for Silo and 99.4% for Liblinear; demotion failures reach 25.4%–37.9% and 46.1%–48.2%, respectively.Most failed promotions target Tier-0; intermediate tiers can receive both promoted hot pages and demoted cold pages.Source: §3.1 · p. 4, Figure 2 · p. 4
- Limited available tier capacity can hinder promotions and demotions, leaving pages in suboptimal tiers.AutoTiering was evaluated with Liblinear and Silo under 1:1:2:2 and 1:1:8:8 tier-capacity configurations.Source: §3.1 · pp. 3–4
3.2 Profiling Overhead Varies Across Tiers
- Hint-page-fault profiling imposes greater performance impact on hot pages because reaccess can trigger TLB misses, page-table walks, and hint faults on the critical path.Cold pages may remain unaccessed for a long time and impose little performance impact.Source: §3.2 · p. 5
- At 75 ms, profiling Tier-0 reduces GUPS by 57.5%, while profiling Tier-3 reduces GUPS by 5.7%; at 1000 ms, losses are 28.0% and 2.1%, respectively.The evaluation used a GUPS-like workload across four 8 GB tiers, with skewed access shares of 64%, 25%, 10%, and 1%.Source: Figure 3 · p. 4, §3.2 · p. 5
- Most existing systems use identical profiling rates across tiers, while selective profiling of only certain tiers does not adapt the period separately for each tier.The uniform rate couples lowest-tier detection latency with profiling overhead in upper tiers.Source: §3.2 · p. 5
4 Design of Mole
Mole combines tier-specific hotness profiling with migration management to address profiling overhead and tier contention. Its four modules collect hotness, control profiling frequency, derive tier thresholds, and guide migrations.
4.1 Overview
- PAP and DPF provide tier-specific profiling, while staging demotion and targeted promotion reduce tier contention by sending cold pages to the lowest tier and matching hot pages to suitable tiers.PAP selects pages through physical-page scans; DPF applies different profiling frequencies across tiers.Source: §4.1 · p. 5
4.2 Physical Address based Hotness Profiling
- PAP scans tier-specific Linux LRU lists and stores profiling state in page metadata, avoiding repeated page-table checks during later scans.Hint-page faults record hot-page reaccess intervals, while scans update unaccessed durations for pages that remain unmapped.Source: §4.2 · pp. 5–6
4.3 Differentiated Profiling Frequency Control
- DPF keeps low-cost metadata scans frequent across tiers but varies costly unmapping by tier, unmapping lower tiers more often and upper tiers less often.An unmap budget limits pages processed per cycle and is distributed in proportion to tier capacity.Source: §4.3 · pp. 6–7
4.4 Hotness Histogram Construction
- HHC uses reaccess intervals and unaccessed durations as a unified hotness signal, with shorter intervals indicating hotter pages and capacity-based thresholds partitioning tiers.Its segmented encoding uses a base-2 logarithmic scale below T and linear steps of T above it; the evaluation uses T=1024 ms.Source: §4.4 · p. 7
4.5 Page Migration Management
- PMM uses HHC thresholds to guide migrations, prioritizing promotions because placing hot pages in faster tiers directly improves performance.Staging demotion sends cold pages to the lowest tier, while targeted promotion places hot pages in tiers matching their hotness.Source: §4.5 · p. 8
4.5.1 Staging Demotion.
- Staging demotion bypasses intermediate tiers to free them for promotions, while high-frequency lowest-tier profiling can detect and promote warm pages demoted too far.Mole also queues cold pages for asynchronous proactive demotion, with kswapd assisting under urgent memory demand and sending pages to the lowest tier.Source: §4.5.1 · p. 8
4.5.2 Targeted Promotion.
- Targeted promotion uses HHC thresholds to select a hot page’s destination, falling back to a lower tier when the target lacks free space and retrying failed promotions asynchronously.Mole promotes pages in fault handling and batches scan-detected promotions asynchronously to reduce overhead.Source: §4.5.2 · p. 8
5 Implementation
Mole implements tier-specific profiling and compact hotness tracking alongside asynchronous migration, with fault-triggered promotion and support for large pages.
Profiling
- Mole profiles four tiers with a 1-second-per-1280-MB scan period, assigning the lowest tier factor 1 and the other tiers factors 2, 3, and 4.The scan period and factors were selected based on evaluation results.Source: §5 · p. 8
Unmapping
- RMAP walks locate PTEs for unmapping; PageUnmap filters redundant attempts, while batched TLB flushes and deferred PROT_NONE invalidation reduce overhead.The tracking thread runs asynchronously and is reported to add negligible performance overhead.Source: §5 · pp. 8–9
Metadata and Hotness Histogram
- Mole reuses the 21-bit last_cpupid field and a one-bit PageUnmap flag; its per-tier 32-counter hotness histograms keep memory use at 0.2%.Per-CPU counters are periodically aggregated, and dynamic thresholds use RCU locks.Source: §5 · p. 9
Page Migration
- Demotion and promotion can run asynchronously in the background, while page-fault-triggered synchronous promotion moves hot pages to faster tiers.The synchronous path adds critical-path latency but ensures hot pages are migrated to a faster tier.Source: §5 · p. 9
Large Page
- Mole defaults to 4 KB pages but is page-size agnostic; its profiling and identification work at page granularity, and Linux supports large-page migration without further modifications.The design supports large pages without requiring additional Linux modifications.Source: §5 · p. 9
6 Evaluation
Across the evaluated four-tier configurations, Mole outperforms the comparison systems and reduces promotion failures. The evaluation also examines workload sensitivity, placement and profiling behavior, and runtime overhead.
6.1 Experiment Setup
- Evaluation uses six workloads on a dual-socket four-tier server, tests 1:1:2:2 and 1:1:8:8 capacities, and compares Mole with AutoNUMA, Tiering-0.8, and AutoTiering.The setup disables Hyper-Threading, Turbo Boost, and prefetching to reduce measurement variability.Source: §6.1 · pp. 9–10
6.2 Performance Comparison
- 70.2% and 20.6%: Mole outperforms the second-best system in the 1:1:2:2 and 1:1:8:8 configurations, respectively; average promotion failures fall to 39.5% and 31.2%, versus over 90% for baselines.The lower failure rates reflect separating demotion destinations from promotion targets.Source: Figure 7 · p. 10, Figure 8 · p. 10, §6.2 · p. 10
- Mole beats the second-best baseline by 20% on Liblinear (1:1:8:8), 7.4% on GUPS, and 12% and 4% on PageRank across the two configurations.On Liblinear, AutoTiering has a 48% demotion failure rate; on GUPS, Mole's promotion success rate is 56.5%.Source: §6.2 · pp. 10–11
- 1.38× over the second-best system on Graph500 (1:1:2:2); on Silo, Mole achieves 24% and 15% higher performance than the second-best baseline across the two configurations.Mole's promotion failure rates are 34.9% on Graph500 (1:1:2:2) and 17.8% and 6.8% on Silo, while Silo baselines remain above 91%.Source: §6.2 · p. 10, Figure 9 · p. 11
6.3 Understanding Mole Performance
- 2.16× and 1.74× average speedups over Tiering-0.8 and AutoTiering as Graph500 RSS rises; Mole-TD instead increases execution time by 2.06× on average versus Mole.With low memory pressure, methods perform similarly; staging demotion is more effective when the lowest tier has available space and holds little data.Source: §6.3.1 · p. 11, §6.3.2 · p. 11
- Mole concentrates Liblinear's hottest pages in Tier-0 and most cold pages in Tier-3, while identifying more GUPS hot pages than each comparison method.It identifies 1.17×, 1.07×, and 1.25× more hot pages than AutoNUMA, Tiering-0.8, and AutoTiering, respectively.Source: §6.3.3 · p. 11, §6.3.4 · p. 11
- On GUPS, the lowest tier devotes all cycles to unmapping, versus about 25% in the top tier; over 90% of lowest-tier pages can be skipped using PageUnmap.For scanning-based profiling, asynchronous scans at identical frequencies keep scan cycles similar across tiers.Source: §6.3.5 · p. 11
6.4 Overhead Analysis
- Mole cuts profiling CPU overhead by 11% versus AutoNUMA and average per-page operation time from 271.0 ns to 246.7 ns; PTE updates and TLB flushes account for 96.8% of unmap overhead.PTWs and reverse-mapping walks account for 2.3% and 0.9% of unmap overhead, respectively.Source: §6.4 · p. 11, Table 4 · p. 11, Table 5 · p. 11
7 Related Work
The related work reviews tracking approaches, hotness identification methods, and migration strategies, contrasting their limitations in multi-tier systems with Mole's design.
7.1 Memory Access Tracking Optimization
- Prior tracking approaches include software-based, hardware-enhanced, and library-based methods; TPP restricts tracking to one tier in two-tiered memory.The passage contrasts Mole's software profiling, which operates across tiers and platforms, with limitations of some tracking approaches.Source: §7.1 · pp. 11–12
7.2 Page Hotness Identification
- MEMTIS relies on PEBS sampling and access frequency, while Tiering-0.8 and Chrono apply interval tracking to hot pages but use LRU for demotion; Mole tracks unaccessed duration before hint page faults.The passage identifies sampling randomness, the need for multiple samples and periodic cooling, and limited multi-tier demotion effectiveness as constraints of prior methods.Source: §7.2 · p. 12
7.3 Page Migration
- Prior migration work optimizes demotion rate or timing, but the passage says these approaches do not address tier contention in multi-tier systems.AutoTiering's greedy promotion is also described as producing suboptimal tier utilization and degraded performance under memory pressure.Source: §7.3 · pp. 12–13
8 Conclusion
Mole combines tier-specific profiling with staging demotion and targeted promotion to address tier contention in multi-tier memory. Its four-tier evaluation reports better performance than state-of-the-art tiering systems.
Conclusion
- Mole sends cold pages directly to the lowest tier and places hot pages in tiers matching their hotness, alleviating tier contention.These mechanisms are enabled by tier-specific hotness profiling.Source: §8 · p. 13
- Tier-specific profiling detects reactivated hot pages promptly while limiting profiling overhead on upper tiers.The conclusion describes the profiling as low overhead.Source: §8 · p. 13
- Evaluation on a four-tier memory system shows that Mole outperforms state-of-the-art tiering systems.Source: §8 · p. 13