ElasticLSM: Adaptive Storage and Elastic Lattice Models

• ElasticLSM is a term encompassing both an adaptive LSM-tree design for key-value stores and elastic lattice spring formulations in mechanics.
• In storage systems, ElasticLSM removes fixed structural constraints to enable continuous, workload-driven compaction and global write throttling.
• In computational mechanics, ElasticLSM models use image-driven and continuum-consistent approaches to simulate elastic and brittle-fracture behavior.

Searching arXiv for papers related to “ElasticLSM” to ground the article in the relevant literature. I’ll look up the exact term and closely related work on arXiv. ElasticLSM is a context-dependent term in recent arXiv literature rather than a single formalism. In storage systems, it denotes an LSM-tree variant that removes fixed structural constraints and enables workload-driven “AnyTime–AnyRuns” compactions and global run-count-based write throttling, as introduced as the structural foundation of ArceKV (Liu et al., 5 Aug 2025). In computational mechanics, the same label is used descriptively for lattice spring models centered on elastic response: the image-driven CuLSM/Img2Particle workflow for elastic and brittle-fracture characterization, and the more formal Isotropic Elastic Lattice Spring Model (IELSM), which augments classical axial-spring lattices with a volumetric constraint term to recover isotropic elasticity with arbitrary Poisson’s ratio in 2D (Chiang et al., 2022). This suggests that ElasticLSM is best understood as a field-specific notion of “elasticity” applied either to system structure or to constitutive mechanics, not as a cross-domain unified theory (Li et al., 3 May 2026).

1. Terminological scope and disambiguation

The ambiguity of the term arises because the acronym LSM has distinct meanings in different disciplines. In key-value storage, LSM means Log-Structured Merge Tree. In mechanics, LSM means Lattice Spring Model. The shared label “ElasticLSM” therefore refers to different technical objects.

Usage	Expansion of LSM	Defining characteristic
ElasticLSM in ArceKV	Log-Structured Merge Tree	Removes fixed size ratios, level capacities, and per-level run counts
“ElasticLSM” in CuLSM/Img2Particle	Lattice Spring Model	Image-driven particle-spring network for elastic and brittle-fracture behavior
IELSM	Lattice Spring Model	Adds a volumetric constraint term to represent arbitrary isotropic Poisson’s ratios in 2D

In the storage literature, ElasticLSM is an explicit design name. In the mechanics literature, the formal model name introduced in 2026 is Isotropic Elastic Lattice Spring Model (IELSM), whereas the earlier 2022 work presents a lattice spring framework that is explicitly characterized as an elastic-and-fracture LSM. This suggests that nomenclature is standardized in the storage case but still partly descriptive in the mechanics case.

2. ElasticLSM as a generalized Log-Structured Merge Tree

ElasticLSM in ArceKV starts from a critique of traditional LSM-tree assumptions: fixed structural parameters, hard-wired compaction patterns, and write stalls tied to structure (Liu et al., 5 Aug 2025). Classical systems such as LevelDB, RocksDB, Cassandra, and Pebble, as well as workload-aware variants such as Moose, Dostoevsky, and Ruskey, retain some predefined level capacities, size ratios, or run-count limits. Compactions are ordinarily triggered when a level becomes “full,” and write stalls are typically tied to the count of level-0 files.

ElasticLSM removes these constraints. There is no fixed ratio such as $\{T,T,T\}$, no enforced level capacity such as $\{TF,T^2F,\dots\}$, and no predetermined per-level run-count vector such as $\{K_1,K_2,\dots\}$ (Liu et al., 5 Aug 2025). Compactions become Any Time rather than “when full,” and candidates become Any Runs rather than only adjacent/full-level merges. What remains are two invariants: timestamp ordering across levels, expressed as “the smallest timestamp in $i$-th level must not be less than the largest timestamp in $(i-1)$-th level,” and the requirement that compactions move downward.

The on-disk state is formalized as a set of sorted runs $S$, with $s=|S|$. On top of that state, ElasticLSM admits two classes of actions. The first is global run-count-based write throttling:

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$

where $F$ is MemTable size, $B$ is block size, $\{TF,T^2F,\dots\}$0 is write-block time, $\{TF,T^2F,\dots\}$1 is the stall penalty, and $\{TF,T^2F,\dots\}$2 is a global run-count threshold. The second is generalized compaction, organized into three patterns: intra-level compaction of any subset of more than one run within one level; adjacent-level compaction of all runs from level $\{TF,T^2F,\dots\}$3 plus zero or more runs from level $\{TF,T^2F,\dots\}$4; and multilevel compaction of all runs from level $\{TF,T^2F,\dots\}$5 to level $\{TF,T^2F,\dots\}$6 plus zero or more runs from level $\{TF,T^2F,\dots\}$7, with output placed at level $\{TF,T^2F,\dots\}$8.

To reason about these actions, the design introduces count windows with workload tuple $\{TF,T^2F,\dots\}$9 for range lookups, updates, and point lookups. The corresponding point and range costs are

$\{K_1,K_2,\dots\}$0

and the total windowed cost from state size $\{K_1,K_2,\dots\}$1 over $\{K_1,K_2,\dots\}$2 windows is

$\{K_1,K_2,\dots\}$3

The central shift is therefore conceptual as much as structural: rather than transitioning toward a precomputed “optimal” LSM layout, ElasticLSM continuously evaluates actions against current state and workload.

3. Arce, ArceKV, and workload-driven compaction control

Arce is the decision engine built on top of ElasticLSM, and ArceKV is the full key-value store implemented atop RocksDB (Liu et al., 5 Aug 2025). Arce models each compaction by its size $\{K_1,K_2,\dots\}$4, run reduction $\{K_1,K_2,\dots\}$5, completion time in count windows $\{K_1,K_2,\dots\}$6, and starting run count $\{K_1,K_2,\dots\}$7, and defines the average cost of a compaction sequence as

$\{K_1,K_2,\dots\}$8

subject to

$\{K_1,K_2,\dots\}$9

The paper states that deciding the optimal sequence of $i$0 compactions is NP-hard, via reduction from Equal-Cardinality Partition.

Arce therefore uses a greedy score that combines short-term penalty and long-term benefit. The short-term penalty is

$i$1

the long-term benefit is

$i$2

and the resulting effectiveness score is

$i$3

Arce selects the candidate with maximum score and skips compaction if all scores are negative. The design also defines domination: a compaction with $i$4 dominates another with $i$5 if $i$6 and $i$7. Only Pareto-front, non-dominated candidates need be considered. The paper further states that there exists a sequence of parameter triples $i$8 such that greedy highest-score selection yields an average cost at most 2× the optimal value.

Compaction duration is estimated by equating foreground I/O time and compaction I/O time:

$i$9

Parameter search is simulation-based rather than RL-based. Every 1,000,000 operations, the foreground reports a current workload tuple $(i-1)$0. If workload mix or tree state changes by more than a threshold $(i-1)$1, taken as $(i-1)$2 in the reported experiments, Arce re-searches the parameter triple $(i-1)$3.

ArceKV integrates this logic into RocksDB through four components: a foreground workload statistics module, an adaptive write controller using the global run threshold $(i-1)$4, an ArcePicker implementing RocksDB’s CompactionPicker interface, and a background parameter searcher. Under concurrent foreground threads, the system adjusts compaction duration using Amdahl’s Law,

$(i-1)$5

with $(i-1)$6.

The reported results are centered on dynamic workloads. ArceKV is described as delivering around 3x faster performance in dynamic scenarios (Liu et al., 5 Aug 2025). On the specific dynamic sequences labeled Workload I and Workload II, the paper reports overall normalized throughput of 2.92× and 2.17×, respectively, relative to LazyLeveling. It also states that ArceKV “adapts rapidly to workload shifts, within 20 million operations, without significant latency spikes.” On YCSB workloads, the reported normalized throughputs are 4.18× on YCSB-A, 2.26× on YCSB-B, 2.34× on YCSB-C, 1.89× on YCSB-D, 5.47× on YCSB-E, and 4.21× on YCSB-F. Overhead measurements attribute approximately 2% of total compaction-plus-flush time to simulation and less than 1% to score computation and decision logic.

A common misconception is that ElasticLSM in ArceKV is simply another choice along the leveling-tiering continuum. The paper’s position is narrower and more specific: the novelty is not a new fixed structure, but the removal of fixed structural constraints themselves.

4. ElasticLSM as an image-driven lattice spring framework

In the mechanics literature, ElasticLSM denotes a lattice spring representation of elastic and brittle-fracture behavior in which a solid is modeled as a network of particles connected by springs (Chiang et al., 2022). The 2022 CuLSM/Img2Particle framework begins from a coarse-grained particle-spring network and emphasizes GPU execution and image-driven model generation. The total elastic potential energy is given as

$(i-1)$7

with the purely elastic regime represented by

$(i-1)$8

Forces follow from differentiation of the potential and act as central forces along the particle-particle line.

CuLSM implements only linear harmonic springs and distinguishes stiff–stiff, soft–soft, and stiff–soft interactions. Representative parameters reported in Table 2 are: stiff–stiff, $(i-1)$9, $S$0, $S$1; soft–soft, $S$2, $S$3, $S$4; and stiff–soft, $S$5, $S$6, $S$7. Boundary loading is displacement-controlled through explicit motion of boundary particles. For a mode-I tension test, top and bottom edge particles are displaced apart along the $S$8-axis at a prescribed engineering strain rate, for example $S$9.

The equations of motion are integrated with velocity Verlet:

$s=|S|$0

$s=|S|$1

$s=|S|$2

The framework validates elastic behavior by comparison both to the analytical solution of a simple harmonic oscillator and to LAMMPS in terms of energies, stress and strain fields, and fracture patterns.

A distinctive feature is the Img2Particle conversion pipeline. A 2D image or image stack provides grayscale values representing material phases; a regular 2D triangular packing is overlaid on the image domain; grayscale thresholds assign particle types; neighboring particles are connected by springs; spring parameters $s=|S|$3 are chosen from particle-type combinations; and boundaries or notches are introduced for mechanical characterization. This yields a particle-spring network that reflects image microstructure without constructing a conventional FEM mesh.

For continuum-like interpretation, the framework computes virial stress

$s=|S|$4

and Lagrangian strain

$s=|S|$5

Average stress-strain response is then used to infer effective elastic properties before fracture, with

$s=|S|$6

On the implementation side, CuLSM parallelizes directly over the spring list rather than via spatial decomposition. Each GPU thread handles one bond, computes force, and accumulates it atomically into the per-particle force array. Using image-derived composites with roughly 12,000 to 47,000 particles and 35,000 to 138,000 springs, the paper reports that CuLSM is one to two orders of magnitude faster than LAMMPS under large communication cutoff and frequent neighbor updates, remains up to 4.4× faster than LAMMPS on one CPU and about 1.5× faster than LAMMPS on four CPUs under optimized settings, and achieves about 2.5× speedup over its own CPU-only implementation.

5. IELSM and continuum-consistent isotropic elasticity

The 2026 Isotropic Elastic Lattice Spring Model (IELSM) addresses a specific limitation of classical lattice spring models: under isotropy on a square lattice with perimeter and diagonal axial springs, the effective Poisson’s ratio is fixed at $s=|S|$7 in plane stress and $s=|S|$8 in plane strain (Li et al., 3 May 2026). The paper derives this from the classical RVE strain energy and shows that the lattice cannot independently tune shear modulus and bulk modulus.

IELSM preserves the square-lattice spring skeleton but adds a volumetric constraint term:

$s=|S|$9

For plane strain, $$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$0. For plane stress, the volumetric contribution is written as

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$1

The total IELSM energy density becomes

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$2

Matching this with isotropic Hooke’s law yields explicit lattice-to-continuum mappings. The direct mappings reported are

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$3

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$4

The admissible Poisson’s-ratio ranges become $$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$5 in plane stress and $$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$6 in plane strain. At $$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$7, IELSM reduces to the classical fixed-$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$8 model.

A central implementation result is the decomposition of the volumetric constraint into standard discrete components. For a square RVE with four nodes and translational DOFs only, the volumetric energy is written as

$$U(s)=\frac{F}{B}\cdot I_w + k\cdot \mathbb{I}(s>c),$$9

This matrix is then decomposed into 10 sub-elements: six rod-type elements and four pseudo-rotational elements. The rod local stiffness is

$F$0

and the rotational spring local matrix is

$F$1

The paper interprets this as a realization of the additional bulk response through equivalent axial, shear, and rotational springs while retaining only translational nodal DOFs.

The numerical analysis emphasizes stability. Generalized eigenvalue comparison with CST and Q4 elements leads to three specific observations: CST exhibits a parasitic shear-locking mode, Q4 exhibits hourglass modes and rank deficiency for nearly incompressible materials, and IELSM has no zero-energy or excessively stiff modes over the admissible isotropic ranges. The paper therefore states that IELSM has better numerical stability than both the standard bilinear quadrilateral element and the constant strain triangular element.

Validation is reported on uniaxial tension, pure shear, stress concentration around a circular hole, a centre-cracked plate, and crucifix-shaped cracks. In uniaxial tension on a $F$2 plate with $F$3, IELSM reproduces analytical $F$4 and $F$5 to machine precision across $F$6. In pure shear, maximum $F$7 agrees with the analytical solution for $F$8. For a centre-cracked plate, the theoretical $F$9 MPa$B$0, while IELSM SIF estimates converge from 2.0092 at $B$1 to 2.5150 at $B$2, reducing error from −19.9% to +0.33%. For cruciform cracks, normalized SIFs are reported to match Cheung’s reference values with <1% error over most cases and maximum error of about 0.9%.

A likely misconception is that IELSM merely tunes spring constants in an existing axial lattice. The derivation is stronger than that: it adds an independent volumetric energy term and derives an exact small-strain isotropic mapping.

6. Relationship, limits, and research directions

The storage and mechanics uses of ElasticLSM share vocabulary but not substrate. In ArceKV, “elasticity” refers to the removal of rigid structural constraints so that compaction and stall actions adapt continuously to workload and current run state (Liu et al., 5 Aug 2025). In CuLSM and IELSM, “elastic” refers to material response represented through springs, either in an image-driven fracture-capable network or in a continuum-consistent isotropic lattice formulation (Chiang et al., 2022). This suggests that the commonality lies only in adaptive or elastic behavior within each field’s native meaning of LSM.

The limitations are correspondingly domain-specific. In ArceKV, the model assumes stable workload composition within count windows, approximates compaction completion at window boundaries, does not explicitly optimize worst-case latency or strict tail-latency SLAs, and is not designed primarily to minimize space amplification. The authors also identify future directions involving specialized hardware, multi-tenant settings, and integration with ML-based policies (Liu et al., 5 Aug 2025). In the CuLSM/Img2Particle framework, springs are linear and brittle, broken springs are assigned zero stiffness rather than deleted, the presented simulations are in 2D triangular packing, and there is no explicit damping term in the reported implementation (Chiang et al., 2022). In IELSM, the derivation assumes small strain, linear elasticity, isotropy, and 2D square topology; extension to 3D cubic networks, arbitrary quadrilateral cells, anisotropy, nonlinear constitutive laws, and dynamic fracture is explicitly identified as future work (Li et al., 3 May 2026).

Taken together, the literature supports two precise encyclopedia meanings of ElasticLSM. One is a workload-driven generalization of the Log-Structured Merge Tree that underpins ArceKV and expands compaction control through “AnyTime–AnyRuns” actions. The other is a family of elastic lattice spring formulations in computational mechanics, ranging from GPU-accelerated image-to-particle fracture simulation to continuum-consistent isotropic spring lattices with arbitrary Poisson’s ratio.