PaperlayerView exact revision on arXiv

Source-linked AI summary

DT-RAID: A Software-Defined Tiered RAID Architecture for Heterogeneous SSDs

Kun-Chi Chiang, Radu Stoica, Animesh Trivedi, Chun-Lien Su, Liang-Chi Chen, Roman Pletka, Wei-Kuan Shih, Chien-Chung Ho

arXiv:2609.16003v1cs.ARcs.DC

TL;DR

Emerging SSDs expose heterogeneous tiers, but integrating them without application changes or losing system-wide placement visibility remains challenging. DT-RAID addresses this at the RAID layer with stripe-level heat tracking and online dual-tier placement. In SNIA enterprise-trace simulations, it improves modeled I/O performance by up to 6.8× and normalized lifespan by up to 20.9× under sufficiently asymmetric tiers.

  • Problem

    Integrating intra-drive heterogeneous SSD tiers into the storage stack without application modifications remains challenging.

  • Method

    DT-RAID uses RAID-layer stripe-level heat tracking and dual-tier placement to steer hot data and parity across host-visible performance and capacity tiers.

  • Results

    DT-RAID improves modeled I/O performance by up to 6.8× and extends normalized lifespan by up to 20.9× compared with uniform RAID deployments under sufficiently asymmetric tiers.

  • Takeaways & Limitations

    The RAID layer is a practical integration point for heterogeneous SSDs, enabling persistent stripe-level placement with bounded capacity reduction.

Abstract

from arXiv · show

The rapid proliferation of cloud and AI-driven workloads has led to increasingly complex requirements for modern storage subsystems. To meet these demands, SSD controller architectures have evolved into a fragmented landscape, offering tiers of drive types optimized for endurance, performance, or capacity. More recently, SSDs have begun to differentiate regions within the same device, enabling intra-drive heterogeneity. However, integrating such heterogeneity into the existing storage stack with minimal disruption remains challenging. In this paper, we argue that storage middleware, such as RAID, is an effective control layer to address these integration challenges. We present DT-RAID, an intra-drive heterogeneity-aware RAID architecture designed for emerging SSDs. DT-RAID monitors stripe-level I/O access patterns and makes online placement decisions without requiring application modifications. It employs a lightweight heat-tracking mechanism to dynamically place frequently accessed (hot) stripes onto the higher-performance, higher-endurance tier. Using simulations based on SNIA MSR enterprise I/O traces, we demonstrate that DT-RAID improves modeled I/O performance by up to $6.8\times$ under greater tier asymmetry and extends normalized lifespan by up to $20.9\times$ compared to uniform RAID deployments.

I. INTRODUCTION

Modern workloads demand diverse performance, endurance, and capacity characteristics, while emerging SSDs expose heterogeneous tiers within a device. DT-RAID uses RAID middleware to manage this heterogeneity transparently through stripe-aware placement and heat tracking.

  • The approach targets increasingly diverse cloud and AI-driven workloads whose storage requirements are difficult to predict.Examples include data lakehouses, transactional services, and AI training and inference workloads.
  • Application-level adaptation requires per-application changes, whereas SSD-internal tiering lacks system-wide visibility and may produce unstable performance.These limitations motivate an intermediate control layer for placement decisions.
  • DT-RAID addresses intra-drive SSD heterogeneity at the RAID layer, preserving the standard block interface without application or filesystem changes.The RAID layer also provides a global cross-device view of access patterns and introduces minimal SSD dependencies.
  • 6.8× modeled I/O performance and 20.9× normalized lifespan are the reported maximum gains over uniform-flash RAID under sufficiently asymmetric tiers.The evaluation uses a trace-driven RAID simulator and enterprise I/O traces from the SNIA IOTTA repository, with bounded capacity reduction.
  • DT-RAID combines lightweight heat tracking with dual-tier placement policies for arrays containing host-visible performance- and capacity-oriented flash tiers.The mechanism monitors stripe-level update intensity and differentiates data and parity placement.

A. NAND Flash Operating Modes

NAND operating modes trade capacity against latency and endurance, while parity-based RAID adds workload-dependent update overhead. Enterprise traces exhibit strong access skew, concentrating RAID activity on a subset of stripes.

  • A. NAND Flash Operating Modes: SLC-like flash provides lower latency and higher endurance, whereas TLC- and QLC-like flash provide greater density with higher latency and lower endurance.These differences arise from the number of bits stored per cell and the associated sensing, programming, and verification requirements.
  • B. RAID Basics and Parity Update Overheads: RAID-5 uses one parity chunk and RAID-6 uses two, so RAID-6 incurs higher parity-update overhead than RAID-5.Both schemes operate on stripes composed of one chunk from each member drive.
  • B. RAID Basics and Parity Update Overheads: Full-stripe writes minimize parity-update overhead, while partial-stripe writes require read–modify–write or read–reconstruct–write processing.The update strategy depends on how many data chunks a write changes relative to the stripe size.
  • B. RAID Basics and Parity Update Overheads: Higher RMW and RCW fractions correspond to a larger share of parity-induced operations in write-intensive traces such as hm_0 and prxy_0.Figure 1 breaks down update strategies and overall data/parity operations for four traces under 8-disk RAID-5 and RAID-6 with 4 KB chunks.
  • C. I/O Workload Skew: A small fraction of the allocated logical space accounts for most I/O activity, while most provisioned space remains cold across nearly all traces.Figure 2 compares cumulative access frequency over active LBAs and the full allocated LBA space.

D. Implications of Skewed Workloads for Parity-Based RAID

Parity-based RAID converts skewed logical updates into concentrated stripe-level activity, making tier placement consequential when flash tiers differ in performance and endurance. Prior RAID work has mainly used heterogeneous drives or faster tiers as staging areas, whereas DT-RAID studies persistent stripe-level placement.

  • D. Implications of Skewed Workloads for Parity-Based RAID: Repeated updates to a small set of logical addresses concentrate parity and data activity on a subset of RAID stripes.Parity consistency is maintained at stripe granularity, so partial-stripe updates propagate logical write skew into RAID-side activity skew.
  • D. Implications of Skewed Workloads for Parity-Based RAID: Stripe-level skew can create uneven flash activity and long-term wear stress when heterogeneous tiers have different performance and endurance characteristics.The cost of each parity-based update depends on the tier composition of its stripe.
  • III. RELATED WORK: Prior RAID studies generally assume homogeneous SSDs, exploit heterogeneity across drives, or use faster flash tiers as temporary staging space.Examples include homogeneous software RAID, HybRAID, and intra-drive designs using SLC partitions for parity-update staging.
  • III. RELATED WORK: DT-RAID instead treats host-visible heterogeneous flash as a substrate for persistent stripe-level placement within the RAID layer.This framing jointly considers parity semantics, workload skew, and tier asymmetry when studying performance and long-term wear.

IV. DT-RAID DESIGN PRINCIPLES

DT-RAID models host-visible heterogeneous SSDs as dual performance- and capacity-oriented tiers and makes placement decisions at stripe granularity. Its analysis isolates RAID-layer effects while assuming configurable host-visible performance and endurance budgets.

  • IV. DT-RAID DESIGN PRINCIPLES: The dual-tier abstraction preserves the essential performance-capacity trade-off while keeping the placement problem tractable.The tiers can represent SLC-like performance flash and TLC- or QLC-like capacity flash; extensions beyond two tiers are outside scope.
  • IV. DT-RAID DESIGN PRINCIPLES: DT-RAID considers flash cost asymmetry, parity-induced structural overhead, and workload-driven skew when determining stripe placement.Access recency and frequency identify repeatedly touched stripes, while read and write costs differ across NAND tiers.
  • C. Stripe Granularity as the Placement Unit: Stripe placement matches parity maintenance and other RAID operations while keeping metadata manageable.In the baseline 8-disk RAID-5 configuration, two 8-bit counters per stripe require about 4 GiB of DRAM, plus about 0.25 GiB for a hot-stripe bitmap.
  • D. Design Scope and Abstraction: The study isolates the causal impact of RAID-layer stripe management by abstracting away garbage collection, block allocation, wear leveling, and background maintenance.This modeling choice avoids additional nondeterministic effects that could obscure controlled comparisons.
  • D. Design Scope and Abstraction: The analysis assumes host-visible flash tiers with configurable performance and endurance budgets, which are not yet universally available in commodity SSDs.The scope is the implication of tier asymmetry for RAID-layer design rather than any specific commercial implementation.

V. DUAL-TIER RAID (DT-RAID) DESIGN

DT-RAID addresses structural parity-write cost and workload-driven stripe skew with two tier-placement policies that preserve RAID semantics. Static RAID places parity on the higher-endurance tier, while Dynamic RAID migrates hot stripes as whole units.

  • Design motivation: Parity-based RAID creates a workload-independent write cost because stripe updates require parity writes, while skewed workloads concentrate activity and parity updates on a small subset of stripes.These structural and workload-driven costs motivate complementary static and dynamic placement strategies.
  • Dynamic RAID: Dynamic RAID migrates whole hot stripes between tiers, preserving RAID geometry, parity rotation, and redundancy semantics while adapting placement to workload skew.Whole-stripe migration keeps data and parity aligned, unlike chunk-granular remapping.
  • Static RAID: Static RAID places parity chunks on the higher-endurance performance tier and data chunks on the capacity tier to reduce parity-induced wear.Parity placement follows the standard rotating RAID-5 layout without changing stripe geometry.
  • Policy comparison: Static RAID serves as a controlled baseline because it mitigates structural parity-write cost but does not react to workload variation.Dynamic RAID complements persistent parity placement with workload-aware migration.

D. Stripe-Level Cost-Aware Heat Metric

DT-RAID estimates stripe heat from host activity, parity-induced amplification, and NAND-level cost, then uses bounded counters and hysteresis for lightweight online placement control.

  • Heat construction: DT-RAID combines read and write activity with parity amplification and tier-dependent NAND costs to identify stripes with higher effective placement cost.The metric models host accesses, RAID-level amplification, and NAND-level cost rather than predicting absolute cost.
  • Placement signal: The resulting heat value serves as the control signal for identifying hot stripes and making stripe-level tier-placement decisions.Higher heat reflects more frequent updates, greater parity-related amplification, and higher write-side cost.
  • Activity tracking: Separate bounded read and write counters provide the activity measurements used to compute each stripe’s cost-aware heat.Counters are implemented as 8-bit values ranging from 0 to 255 and increment on each stripe access.
  • Online control: Heat regulation keeps average stripe heat below 127 by decrementing counters round-robin when the bound is exceeded, preserving relative hotness without saturation.Placement uses hysteresis: promotion occurs at heat 140, while demotion occurs below 127.

VI. SIMULATOR ARCHITECTURE

The evaluation uses a modular trace-driven simulator that preserves RAID semantics while modeling stripe heat, tier placement, and migration overhead. The study tests performance, lifespan, geometry, parity, and tier-asymmetry hypotheses under controlled conditions.

  • Simulator architecture: The framework preserves parity-based RAID geometry and redundancy semantics while explicitly modeling parity-consistent updates, heat tracking, tier placement, and migration overhead.Figure 4 presents the simulator architecture and request-processing flow.
  • Request-processing flow: The simulator replays block-level requests, maps each LBA to stripe metadata, updates accounting state, and dispatches requests through common read/write processing.Writes use a cache and update heat before writeback; reads check the cache before accessing the SSD array.
  • Evaluation hypotheses: The evaluation tests whether DT-RAID improves normalized I/O acceleration, device lifespan, robustness across RAID geometries, and gains under higher parity or tier asymmetry.The hypotheses compare RAID-5 with RAID-6 and SLC/TLC with SLC/QLC conditions.

A. Experimental Methodology

The methodology compares conventional, static, and dynamic tier placement under identical RAID organization and enterprise traces. It measures tier steering, flash cost, endurance, and sensitivity to RAID and tier parameters.

  • Compared configurations: The study compares ST RAID, DT-S RAID, and DT-D RAID while preserving identical stripe geometry and parity behavior, isolating tier-placement policy effects.ST places all data in the capacity tier; DT-S places parity-related chunks in the performance tier; DT-D migrates stripes using heat.
  • System model: The default model uses an 8-disk RAID array with 4 KB chunks and two host-visible tiers representing performance-optimized and capacity-optimized flash.The SSD abstraction tracks per-tier read and program activity, flash operation time, and endurance budgets.
  • Workloads: Enterprise SNIA IOTTA traces are replayed in original block-level order with identical trace segments across policies for fair comparison.The workloads are summarized by read/write ratios, total I/O counts, and working-set sizes.
  • Metrics: Metrics cover per-tier flash activity and capacity reduction, accumulated flash operation time, normalized I/O acceleration, per-tier program activity, lifespan, and parameter sensitivity.The evaluation targets tier steering, performance-related flash cost, wear, lifetime, RAID geometry, parity overhead, and tier asymmetry.
  • Tier-activity analysis: Figure 5 breaks down normalized flash operations by read/write type and tier across workloads and placement policies.Perf. R/W denotes performance-tier operations, while Cap. R/W denotes capacity-tier operations.

H1: DT-RAID Tiering Effectiveness

DT-RAID redirects RAID flash activity toward a performance tier, with dynamic placement steering substantially more operations than parity-only placement. The resulting modeled acceleration is evaluated with a simplified NAND-operation-time model, while dynamic placement achieves lower capacity reduction than static placement.

  • Placement policies: DT-S redirects parity activity without changing RAID geometry or redundancy semantics, while DT-D relocates hot data and parity using dynamic placement.The capacity-reduction expression uses the raw-capacity ratio ρ and the chunks resident in performance and capacity tiers.
  • Per-tier I/O breakdown: 74%–95.1% of total NAND flash operations reach the performance tier with DT-D, compared with 19.6%–41.1% under DT-S.DT-D redirects 20.3%–49.6% of writes and 37.5%–68.1% of reads; DT-S redirects 10.4%–21.1% of writes and 8.7%–20% of reads.
  • Capacity reduction: DT-D incurs 0.6%–5.1% capacity reduction, whereas Static RAID incurs a fixed 20% reduction in the default 8-disk RAID-5 SLC/TLC configuration.Capacity reduction measures usable capacity lost when chunks are mapped to the performance tier; Figure 6 presents the comparison against an all-TLC baseline.

H3: Normalized Lifespan Improvement

DT-RAID’s lifespan benefit depends on the endurance gap between performance and capacity tiers. Dynamic placement provides larger gains at higher endurance ratios, but both policies can reduce lifespan when the ratio is below 3:1.

  • Lifespan results: At a 100:1 endurance ratio, DT-D extends normalized lifespan by 3.1×–20.9× versus ST-RAID, depending on workload.At 4:1 the gain is 1.2×–1.3×, and at 8:1 it is 1.8×–2.6×.
  • Lifespan results: DT-S extends lifespan by 1.1×–1.6× versus ST-RAID when the endurance ratio exceeds 4:1.DT-D shows larger but more workload-dependent gains across the reported endurance ratios.
  • Condition for benefit: Below a 3:1 endurance ratio, both DT-S and DT-D can reduce lifespan rather than improve it.The wear model accounts for endurance-normalized program/erase cycles and the greater block usage of SLC for the same data volume.

H4: Sensitivity to RAID Geometry

DT-D remains effective as RAID-5 stripe width increases, whereas parity-only DT-S redirects less activity because wider stripes contain a smaller parity fraction. Dynamic placement also maintains minimal capacity reduction across the tested geometries.

  • Per-tier I/O distribution: DT-S performance-tier write share drops from 24.2% to 19.4% as RAID-5 width increases, while its performance-tier read share remains nearly unchanged.The capacity-tier read share correspondingly rises from 21.4% to 28%.
  • Per-tier I/O distribution: DT-D keeps capacity-tier writes at 3.2%–0.6% and reads at 3.3%–2.4% as RAID-5 width grows, while performance-tier reads rise from 48.1% to 81.9%.These results indicate greater robustness than DT-S as stripe width scales.
  • Capacity reduction: DT-D capacity reduction stays below 1% across tested stripe widths, changing only from 0.5% to 0.8%.DT-S reduction falls from 33.3% to 11.1% as wider stripes reduce the parity fraction.

H6: Sensitivity to Tier Performance/Capacity Asymmetry

DT-RAID gains more from tier asymmetry when dynamic placement steers hot-stripe I/O toward the faster tier. The strongest acceleration appears with the larger SLC/QLC performance gap.

  • DT-S: 1.2×–1.6× normalized acceleration from DT-S remains moderate and consistent across SLC/TLC and SLC/QLC.Parity-only placement benefits from greater asymmetry, but gains remain limited because only parity-related activity moves to SLC.
  • Asymmetry effect: Wider tier performance gaps increase acceleration because shifting I/O to the faster tier reduces total flash operation time.The SLC/QLC capacity tier has lower baseline performance than the SLC/TLC configuration.
  • Design implication: DT-D benefits from steering hot-stripe user I/O to the performance tier while preserving existing RAID geometry and redundancy semantics.Both placement schemes avoid application changes, and stripe granularity is identified as a practical placement unit.
Loading 2609.16003v1…