Cross-Modal Data Co-Locality for Heterogeneous Storage Systems

Cross-Modal Data Co-Locality for Heterogeneous Storage Systems

arXiv:2610.34595v1 [cs.DC] 25 Oct 2026

Abdullah Al-Mamun Augusta University aalmamun@augusta.edu

Abstract

is largely orthogonal to ANN search efficiency. Consider a multimodal query that first searches text or EHR vectors and subsequently accesses the corresponding image representations. The two modality-specific ANN searches can each be efficient, yet their results may reside on different machines, NUMA domains, memory tiers, or storage devices. The system must then move data across the network, NUMA hierarchy, CPU–GPU boundary, or storage hierarchy to assemble the final result. Thus, a vector system can optimize its individual ANN searches while still paying a substantial physical data-movement cost at the multimodal query level. The fundamental problem is therefore not only where vectors should be searched, but also where heterogeneous representations required by the same query should physically meet. Recent vector systems increasingly optimize physical locality, but at different levels. Distributed ANN systems such as CoTra and RED-ANNS optimize partitioning, communication, and distributed execution [8, 37], while d-HNSW explores vector search over disaggregated memory [12]. Storageaware vector systems exploit locality within vector indexes to reduce I/O [20], and heterogeneous-memory systems study locality-aware placement and migration across memory tiers [13, 24, 38]. Meanwhile, vector database architectures are increasingly spanning heterogeneous memory and storage resources [18, 21, 22]. These systems address important forms of index, communication, or resource locality, but do not make cross-modal co-consumption a first-class physical placement objective. Table 1 summarizes the fundamental limitations of current representative locality-aware systems. In summary, multimodal queries induce a distinct form of physical locality: representations that are repeatedly consumed together should preferably be physically close, even when they belong to different indexes and modalities. Semantic similarity alone is insufficient because semantically related representations need not be accessed together, while observed co-access can be sparse for emerging workloads. The placement problem must therefore combine workload co-access with semantic and temporal evidence while accounting for heterogeneous hardware costs.

Distributed vector stores increasingly serve multimodal workloads whose representations are jointly accessed but independently partitioned and placed. As a result, cross-modal physical locality remains largely unoptimized. Multimodal queries can incur remote accesses and cross-tier data movement even when each modality’s ANN search remains efficient. We present CoLoc, a storage-management layer that makes cross-modal access locality a first-class data-layout signal. A Cross-Modal Access Graph combines semantic affinity, co-access, and temporal behavior. A movement-aware optimizer places frequently co-accessed vectors under capacity and topology constraints. A migration manager adapts placement as workloads shift by weighing locality gains against migration cost. A confidence-gated prefetcher hides remaining slow-tier latency. CoLoc operates above existing ANN indexes without modifying embeddings, search, or retrieval semantics. We deploy CoLoc on a multi-tier AWS testbed spanning GPU, NUMA-DRAM, NVMe, and S3. CoLoc substantially reduces data movement and query latency while leaving ANN retrieval unchanged. Under a scheduled workload shift, its migration manager matches the locality of periodic global reorganization while moving 5.9× less data and completing 3.9× faster.

1

Introduction

Multimodal applications increasingly represent the same logical object using multiple modality-specific embeddings, such as text, images, clinical records, and sensor data [7, 9, 10, 36]. Medical workloads, for example, increasingly combine electronic health records (EHRs), medical images, laboratory measurements, and other clinical modalities [11], motivating new systems support for multimodal data storage and retrieval [17]. These representations are often independently embedded and indexed, allowing each modality to use an appropriate ANN structure but also allowing related representations to become physically separated. This separation creates a physical data-layout problem that 1

Table 1: Positioning CoLoc against representative locality-aware vector, memory, and storage systems. System CoTra [37] RED-ANNS [8] d-HNSW [12] Turbocharging [20] M2T [38] PolyStore [19] LACX [13] CoLoc

Placement Scope Vector partitions ANN graph partitions ANN data ANN index layout Memory pages Storage objects Memory pages Cross-modal placement units

Locality Signal Search / vector similarity Graph / vector similarity ANN search locality Access locality Memory-tier access Storage workload Shared-page access Co-access + semantic + temporal

We present CoLoc, a storage-management layer that introduces cross-modal co-locality as a first-class placement objective without modifying embeddings, ANN indexes, distance metrics, or retrieval semantics. The Cross-Modal Access Graph (CAG) captures relationships among independently indexed representations using cross-modal co-access, semantic affinity, and temporal proximity. The Co-Locality Placement Optimizer (CPO) translates these relationships into capacity-, load-, and topology-aware placement decisions while accounting for physical movement cost. A topology-aware layer maps logical decisions onto GPU memory, NUMA DRAM, NVMe, and remote storage, while the Locality-Preserving Migration Manager (LPMM) adapts placement under workload drift only when the expected future locality benefit justifies migration cost. The central distinction is simple: existing vector systems optimize where vectors are searched; CoLoc optimizes where the heterogeneous representations required by a query physically meet. In summary, the paper makes the following contributions:

Cross-Modal – – – – – – – ✓

Movement-Aware – – – – ✓ ✓ ✓ ✓

Adaptive Migration – – – – ✓ – ✓ ✓

Retrieval quality (Recall@k) remains completely unaffected by placement throughout.

2

Related Work

Newer ANN serving and multimodal query systems. Recent large-scale and GPU-accelerated ANN systems – DistVS [32], GIDS [1], and VecFlow [3] – continue the searchexecution focus already discussed in §1; none reasons about placement across independently-managed indexes. Closer in spirit is ARCADE [30], a real-time engine for hybrid, continuous queries across diverse data modalities. ARCADE addresses query execution across modalities within one system; it does not address the physical co-location of independentlyindexed representations across heterogeneous storage tiers, which is where CoLoc operates – the two are complementary rather than competing. General cost-aware data placement. Beyond the heterogeneous storage/memory placement model of §3.3, similar movement-vs-reorganization trade-offs appear in learned directory placement [34], Markov-clustering-based locality for MapReduce [25], and multicloud placement [5]. This work is domain-agnostic; CoLoc specializes the same trade-off to independently-indexed multimodal vectors, where the placement unit and migration trigger are shaped by cross-modal query structure (§3.1) rather than a generic access trace. Industry movement toward object-storage-integrated vector search [6] and filtered ANN search at scale [2] reflects the same trend motivating CoLoc: vector workloads increasingly span storage tiers by default. Multimodal retrieval quality and clinical fusion. A separate line of work improves what gets embedded and how modalities are combined for retrieval quality: generative multimodal search [15] and noisy-correspondence-robust cross-modal alignment [4, 27, 28, 33]. In the clinical domain, DrFuse [31] and missing-modality EHR embedding [16] extend the fusion challenge §1 raised via MedMod. All of this determines embedding and retrieval semantics, which CoLoc explicitly does not touch (§3); better embeddings increase the value of correct placement without changing where independentlyindexed vectors physically end up relative to each other – the gap CoLoc fills. VDBMS engineering maturity. Work characterizing reliability gaps [26], real production bugs [29], and fundamental system concepts [23] in vector database management re-

• Cross-modal co-locality abstraction. We introduce cross-modal co-locality as a physical data-layout primitive for independently indexed multimodal representations. • Movement-aware placement. We develop CAG and CPO to translate cross-modal co-access, semantic and temporal evidence into topology- and capacity-aware placement decisions that minimize physical data movement. • Adaptive migration. We develop LPMM to adapt placement under workload changes while explicitly trading future movement savings against migration cost. • System evaluation. We implement CoLoc above existing ANN indexes and evaluate it on a real multitier AWS deployment spanning GPU, NUMA-DRAM, NVMe, and S3. CoLoc eliminates cross-tier movement entirely (73%→0% of multimodal queries) and reduces end-to-end latency by up to 25% over independent placement at high multimodal query intensity. LPMM adapts to workload shift while moving 5.9× less data and completing 3.9× faster than periodic global reorganization; a CAG-driven prefetcher further reduces latency by 8.5% by overlapping remote fetches with query processing. 2

flects growing engineering rigor in this space. Physical placement across heterogeneous tiers remains comparatively underexamined in that line of work; CoLoc treats it as a first-class, independently evaluable concern (§4).

3

reduction against reorganization during a full re-solve, such as initial placement or periodic reoptimization. LPMM’s continuous per-group test instead uses an expected query horizon H, estimated from each group’s observed rate rather than a fixed weight. The optimization unit is neither an ANN neighborhood nor an index structure, but the set of heterogeneous representations that a workload repeatedly consumes together, at whatever granularity can safely move.

Cross-Modal Co-Locality Design

CoLoc is a storage-management layer above existing modality-specific embedding and ANN indexes. It does not modify embeddings, ANN search algorithms, distance functions, index structures, or retrieval semantics. Instead, it adapts the physical placement of heterogeneous representations according to their joint consumption by multimodal queries. Cross-modal accesses are asynchronously collected by the Cross-Modal Access Graph (CAG), which combines observed co-access with semantic and temporal evidence through incremental updates. The Co-Locality Placement Optimizer (CPO) uses this signal to determine movement-aware placement through an iterative group-partitioning heuristic and maps the resulting groups onto topology- and tier-aware resources. The Locality-Preserving Migration Manager (LPMM) adapts placement when workload changes justify migration. A CAGdriven prefetcher (§3.6) complements these mechanisms by warming a bounded cache ahead of predicted accesses without changing placement. Together, these components form a closed control loop with no synchronous coordination on the query critical path.

3.1

3.2

with tunable α, β, γ. Co-access provides the primary evidence, semantic affinity provides cold-start evidence, and temporal proximity captures short-lived correlations. This weighted fusion provides a practical locality score rather than a probabilistic model. CAG uses an incremental maintenance policy. Each new access event marks its object dirty, and a rebuild recomputes Auv /Suv /Tuv only for pairs touching a dirty object. Since embeddings are static, Suv is cached permanently. All other edges remain unchanged: co-access pairs form only within one object, and Tuv depends only on the two endpoints’ own lastaccess times. One caveat is that the normalization denominator of Auv is a global statistic, computed from a robust percentile of coaccess counts to reduce sensitivity to individual high-traffic pairs. Consequently, the statistic can change even when the updated workload does not directly involve a particular object. When such a change invalidates the existing normalization, the implementation falls back to a full rebuild to avoid using stale co-access weights. Thus, the benefit of incremental maintenance is most pronounced under frequent, small workload updates and progressively diminishes toward full-rebuild cost for large, infrequent workload shifts. The Control-Plane Overhead evaluation in §4 examines both regimes rather than assuming that the incremental bound holds unconditionally.

Query-Induced Cross-Modal Locality

Existing vector systems [8, 12, 18, 20–22, 37] reason about locality at the level of individual vectors, ANN neighborhoods, or storage blocks – not multimodal queries whose results span several independently managed indexes. CoLoc instead defines locality at the level of joint query consumption. Let V = {v1 , . . . , vN } be modality-specific representations and H = {h1 , . . . , hM } available heterogeneous resources (GPU, NUMA, NVMe, remote). A placement unit is the smallest independently movable storage extent exposed by the underlying layer. Let p(v) denote the resource holding v’s placement unit. A multimodal query induces Qi = {vi1 , . . . , viki }, possibly spanning multiple modalities and indexes. The physical separation cost of a query is Csep (Qi , p) =

∑ D p(u), p(v) ,

(1)

u,v∈Qi u<v

where D(ha , hb ) is the physical cost (seconds) of moving data between ha , hb (§3.4). Over a workload Q with frequencies ρi , the expected movement cost is Cmove (p) = ∑ ρiCsep (Qi , p).

3.3 Movement-Aware Cross-Modal Placement Existing graph-partitioning and vector-placement techniques [8, 12, 19, 20, 35, 37] optimize partition balance, index locality, or resource utilization, but not the physical cost of separating representations consumed together by a multimodal query. CoLoc instead uses cross-modal movement as the placement objective. CPO consumes CAG and topology information to determine where placement units reside. For v at h0 , the reduction

(2)

Qi ∈Q

CoLoc minimizes recurring movement while limiting reorganization: min C (p) + ηC (p, p ), (3) p

move

mig

Cross-Modal Access Graph

Semantic affinity alone does not establish that two representations will be consumed together [7, 9–11, 36], while sparse observations can be insufficient for emerging relationships. CoLoc therefore combines observed co-access with semantic and temporal evidence in a weighted graph GA = (V, EA ,WA ), where each vertex is a placement unit. For each pair u, v, CAG maintains co-access Auv , semantic affinity Suv , and temporal proximity Tuv , fused as wuv = αAuv + βSuv + γTuv , (4)

0

subject to per-resource capacity. Here p0 is the current placement, Cmig is its change cost (§3.5), and η trades movement 3

Algorithm 1 CPO: Movement-Aware Group Placement

when supplied. High-traffic groups claim fast-tier capacity first, leaving data structurally forced onto slower tiers disproportionately cold. This ordering does not change ∆C(K, h) itself. When a group cannot fit on one resource, CPO splits it iteratively along its weakest locality edges and returns the resulting sub-groups to the queue. CPO never rewrites ANN graph edges, vector codes, or index structures; it only relocates placement units, preserving each index’s internal optimizations.

Require: CAG GA = (V, EA ,WA ), resources H with capacities, vector sizes, current placement p0 , optional queryfrequency f (·) Ensure: placement p 1: K ← ClusterViaLabelPropagation(GA ) ▷ seeded, deterministic 2: if f given then
3: sort K by ∑v∈Ki f (v), IncidentWeight(Ki ) , descending 4: else 5: sort K by IncidentWeight(Ki ), descending 6: Q ← queue initialized with K in sorted order 7: while Q not empty do 8: Ki ← Q.pop() 9: size ← ∑v∈Ki size(v) 10: h⋆ ← ⊥; best ← (∞, ∞) 11: for h ∈ H with remaining capacity ≥ size do 12: cost ← ∑v∈Ki ∑(v,u)∈EA , u already placed wvu · D(h, p(u)) 13: if (cost, RefCost(h)) < best then ▷ RefCost = mean D to every other resource 14: h⋆ ← h; best ← (cost, RefCost(h)) 15: if h⋆ ̸= ⊥ then 16: p(v) ← h⋆ for all v ∈ Ki ; deduct size from h⋆ ’s remaining capacity 17: else if |Ki | ≤ 1 then ▷ no resource had room for even one unit 18: h⋆ ← lowest-RefCost resource with any room; else highest-remaining-capacity resource 19: p(v) ← h⋆ for the sole v ∈ Ki 20: else 21: (v1 , . . . , vk ) ← Ki sorted by CAG degree, ascending 22: Q.push({v1 , . . . , vk−1 }); Q.push({vk }) ▷ peel off the highest-degree vertex 23: return p

3.4

Topology- and Tier-Aware Placement

Generic locality abstractions are insufficient on heterogeneous systems, where NUMA-remote, CPU–GPU, NVMe, and remote-node costs differ substantially [19, 35]. CoLoc therefore represents the platform as a topology-aware cost matrix in time units, allowing it to compose with ρi in Eq. 2: D(hi , h j ) = λL L̂i j + λB

1 ˆ i j + λG Ĝi j , + λI IO ˆ ij BW

with normalized latency, bandwidth, I/O, and GPU-transfer terms weighted by λL , λB , λI , λG . This allows CPO to distinguish same-NUMA from cross-NUMA access, GPU transfer from NVMe fetches, and local from remote movement. Thus, the same CAG relationship can yield different placements under different topologies without restructuring any ANN index, since only the backing extent moves.

3.5

Locality-Preserving Migration

Static placement [13, 24, 38] degrades as access patterns evolve, while continuous reorganization can outweigh the resulting locality benefits [13, 35]. LPMM monitors CAG changes and evaluates whether units or groups should move. Migrating v from h0 to h costs Cmig (v, h) =

size(v) , ˆ (h0 , h) BW

(7)

commensurate with ∆C(v, h). LPMM computes B(v, h) = ∆Cfuture (v, h) −Cmig (v, h),

in future movement cost from placing it at h is ∆C(v, h) = Cfuture (v, h0 ) −Cfuture (v, h).

(6)

(8)

and migrates only when B(v, h) > 0 . For an expected horizon H, this is equivalent to

(5)

CPO operates on groups K of high-connectivity units, so ∆C(K, h) = ∑v∈K ∆C(v, h). The resulting constrained graph-partitioning problem is NP-hard, so CPO uses a lightweight heuristic (Protocol 1). It forms groups using seeded asynchronous label propagation, orders them by incident co-locality weight, and assigns each group to a feasible resource that maximizes ∆C(K, h) under capacity. Ties are resolved using the resource’s mean cost to other resources rather than incidental enumeration order. Incident weight measures connectivity but not query volume. CoLoc therefore optionally accepts per-vector query frequency and uses aggregate query volume to order groups

H · ∆Cquery (v, h) > Cmig (v, h),

(9)

with ∆Cquery = ∆Cfuture /H the expected per-query benefit and H estimated from v’s observed rate. Migration is therefore an amortized decision rather than an unconditional response to every locality change. LPMM re-optimizes only units or groups whose relationships changed enough to matter, avoiding repeated reorganization of cold data. Migration follows a copy–verify–swap protocol (Protocol 2): the destination is populated and verified, the placement map is updated atomically, and the old extent is released only after success. A failed migration therefore leaves the previous placement valid. 4

Table 2: Movement-Aware Placement.

Algorithm 2 LPMM: Threshold-Gated Migration Require: candidates C (dirty units flagged since the last check), CAG GA , current placement p0 , margin (default 0) Ensure: placement p, migrated bytes 1: N ← C ∪ {u : (v, u) ∈ EA , v ∈ C} ▷ candidates + their CAG neighbors 2: GN ← subgraph of GA induced on N 3: p′ ← CPO(GN , . . . , current placement = p0 |N , pinned = N \C) ▷ Alg. 1, restricted to N old ← 4: Cmove ∑(u,v)∈EGN wuv · D(p0 (u), p0 (v)) new ← 5: Cmove ∑(u,v)∈EG wuv · D(p′ (u), p′ (v))

Metric Cross-tier access rate (%) Bytes on remote tier (%) End-to-end latency (ms) Max tier utilization (%) Capacity constraint violated?

Multimodal frac. 0.1 0.3 0.5

N

73.04 34.9 12.52 166.7 Yes

0.00 38.8 11.67 100.0 No

Indp (ms)

CAG+CPO (ms)

Reduction

10.52 11.64 13.27

9.65 9.73 9.98

8.3% 16.4% 24.8%

Table 4: Retrieval Neutrality and End-to-End Cost.

8: if ∆C > Cmig + margin then ▷ Eq. 3.5, generalized 9: for v with p′ (v) ̸= p0 (v) do ▷ copy–verify–swap 10: copy v to p′ (v) 11: if verify(v at p′ (v)) succeeds then 12: p(v) ← p′ (v) atomically; release v’s copy at

Configuration Independent Access-Aware CAG + CPO F. Sys F. Sys + Prefetch

p0 (v) else

Recall@10

ANN Lat. (ms)

E2E Lat. (ms)

1.0000 1.0000 1.0000 1.0000 1.0000

8.72 8.72 8.72 8.72 8.72

12.52 20.67 11.67 11.76 10.75

p(v) ← p0 (v) ▷ never released – p0 remains

valid for v 15: else 16: p ← p0 17: return p

3.6

CAG+CPO

Table 3: Latency vs. Query Intensity.

old −Cnew 6: ∆C ← Cmove move ˆ (p0 (v), p′ (v)) 7: Cmig ← ∑v : p′ (v)̸= p0 (v) size(v) / BW

13: 14:

Independent

Since the worker pool has bounded capacity, dispatch order matters when more candidates clear the threshold than the pool can serve. CoLoc uses a confidence-ordered priority queue rather than FIFO submission, so the pool serves the highest-confidence outstanding candidate first.

▷ decline: below threshold, no migration

3.7

CAG-Driven Prefetching

Placement and migration change where data resides, but cannot eliminate the fraction structurally forced onto the slowest tier under tight capacity (see end-to-end performance evaluation in §4). CoLoc instead hides the cost of reaching that data. When a query accesses u, its CAG neighbors become candidates for background retrieval into a bounded cache, overlapping the slow tier’s round trip with subsequent computation. The cache is maintained separately from the placement map p(·): a hit changes where a fetch is served from, but not CPO’s or LPMM’s record of where v is placed. Raw co-access counts do not provide calibrated confidence: two-of-two and two-hundred-of-two-hundred reflect different levels of evidence that Auv alone cannot distinguish. CoLoc therefore scores each candidate v of accessed vertex u under a Beta-Binomial model. Given suv observed co-accesses and fuv observed accesses of u without v following, P(v | u) =

suv + a0 suv + fuv + a0 + b0

Operational Overhead and Isolation

All optimization mechanisms stay off the query critical path: execution generates only lightweight access observations, while CAG updates, CPO, topology evaluation, LPMM migration, and prefetching (§3.6) run asynchronously in the background – a prefetch cache miss simply falls through to the vector’s normal placed resource at no extra cost. The design also avoids global reorganization: CPO evaluates only affected groups, and LPMM migrates only units whose savings justify their cost, so migration decisions scale with the amount of workload change, not dataset size. Incremental CAG maintenance (§3.2) is a second, narrower reduction: it bounds rebuild cost by workload drift only under frequent, small updates, where the global co-access normalization stays stable between calls. A single large, infrequent shift – such as the workload shift in Adaptive Migration evaluation in §4 – can still trigger a full rebuild. CoLoc therefore bounds control-plane overhead by workload drift in the frequent-update regime specifically, not unconditionally. The Evaluation section reports both regimes rather than assume the bound holds everywhere convenient.

(10)

under a Beta(a0 , b0 ) prior that shrinks estimates toward the prior for sparse history and converges to the empirical rate with increasing evidence. A candidate is prefetched only above an explicit confidence threshold, and only if v is on a slow tier and not already cached.

4

Evaluation

We implement the prototype in Python for the proposed crossmodal co-locality framework as a management layer deployed 5

Table 5: Adaptive Migration (LPMM). Metric

Static

Periodic Reorg

LPMM

Migrated data (MB) Migration time (s) Post-shift cross-tier rate (%) Post-shift latency (ms)

0.0 0.0 42.02 11.16

30.19 54.65 0.00 10.82

5.10 14.03 0.11 10.15

with multimodal query intensity. The reduction grows from 8.3% at 10% multimodal queries to 16.4% at 30% and 24.8% at 50%, indicating that cross-modal co-locality becomes increasingly beneficial as more queries access multiple modalities. ANN Quality, Retrieval Neutrality, and Prefetch. Table 4 shows that Recall@10 (1.0000) and ANN latency (8.72,ms) remain unchanged across all configurations, demonstrating that CoLoc does not alter the ANN retrieval path. Full System+Prefetch further reduces end-to-end latency from 11.76,ms to 10.75,ms (8.5%) relative to Full System, showing that confidence-gated prefetching can hide remaining remoteaccess latency without affecting retrieval quality.

Table 6: Incremental CAG Overhead, Real AWS. Manysmall-batches regime (Sweep D). Metric Full-rebuild fallback triggered Incremental total time (s) Full-rebuild-equivalent time (s) Speedup

Value 1 / 19 batches 0.756 5.235 6.92×

Adaptive Migration. Table 5 shows that LPMM adapts to the workload shift with substantially lower migration cost than periodic reorganization. LPMM moves only 5.10,MB and completes migration in 14.03,s, compared with 30.19,MB and 54.65,s for periodic reorganization—a 5.9× reduction in data movement and 3.9× faster migration. At the same time, LPMM nearly restores the locality of periodic reorganization, reducing post-shift cross-tier access from 42.02% to 0.11% and achieving the lowest post-shift latency (10.15,ms). These results demonstrate that LPMM effectively adapts placement while substantially reducing the cost of global reorganization.

on AWS, above an existing vector-search stack. The prototype separates the ANN/search path from the placementmanagement path, allowing us to evaluate physical data placement and data movement without modifying the underlying retrieval procedure. Experimental Setup. We deploy CoLoc on one AWS EC2 instance per tier: g5.xlarge (GPU), r6i.4xlarge (NUMADRAM), i4i.2xlarge (NVMe), and Amazon S3 (remote, measured 24.2 ms round-trip, 86.4 MB/s – one to two orders of magnitude slower than every other tier). Tier capacities are sized relative to dataset size (15%/20%/30%/50% for GPU/NUMA/NVMe/remote) so no strategy can trivially avoid multi-tier use. We evaluate 50,000 objects (100,000 vectors, 768 dimensions) from MIMIC-CXR [14] (377k radiographs, 228k reports). CAG weights use α = β = γ = 1.0 (Eq. 4, uniform weighting). We compare five configurations: (i) Independent: modality-specific representations are placed independently without cross-modal coordination; (ii) AccessAware: placement uses observed cross-modal access only, isolating the contribution of workload co-access from semantic and temporal signals. (iii) CAG+CPO: cross-modal relationships are constructed by CAG and translated into capacityand topology-aware placement by CPO. (iv) Full System: CAG+CPO plus LPMM for adaptive migration. (v) Full System+Prefetch: the complete system with confidence-gated background prefetching. End-to-End Performance and Movement. Table 2 shows that CAG+CPO eliminates cross-tier access (73.04%→0.00%) and reduces end-to-end latency by 6.8% (12.52→11.67,ms), while satisfying the tier-capacity constraints. Although CAG+CPO places slightly more data on the remote tier (38.8% vs. 34.9%), it achieves lower latency by co-locating frequently accessed representations. In contrast, Independent placement exceeds its capacity allocation (166.7%), whereas CAG+CPO remains within the limit (100%). These results demonstrate that movementaware placement improves cross-modal locality and query performance under constrained heterogeneous storage. Table 3 shows that the latency benefit of CAG+CPO increases

Control-Plane Overhead. Table 6 shows that incremental CAG maintenance reduces construction time from 5.235,s to 0.756,s, providing a 6.92× speedup in the many-smallupdates regime. The mechanism triggers a full rebuild only once across 19 batches, demonstrating that incremental maintenance substantially reduces control-plane overhead when workload changes are small and frequent.

5

Conclusion and Future Work

We presented CoLoc, a storage-management layer that treats cross-modal co-locality as a first-class data-layout objective for distributed vector stores and comprises three tightly coupled components. CAG captures cross-modal access, semantic, and temporal locality. CPO translates these signals into topology- and capacity-aware placement. Finally, LPMM adapts placement under workload shifts. A confidence-gated prefetcher complements these components by reducing remaining slow-tier latency. CoLoc operates above existing ANN indexes without altering retrieval semantics. Our AWS evaluation demonstrates that cross-modal placement reduces data movement and end-to-end latency, while LPMM recovers locality after workload shifts with substantially less migration than global reorganization. Future work includes online adaptation of placement and prefetching parameters, evaluation across larger and geographically distributed deployments, and integration with multitenant vector serving. 6

References

on Health, Inference, and Learning, volume 287, pages 781–803, 2025.

[1] Design and optimization of an anns framework based on gids. High-Confidence Computing, 2025.

[12] Fei Fang, Yi Liu, and Chen Qian. d-HNSW: A highperformance vector search engine on disaggregated memory. arXiv preprint arXiv:2603.13591, 2026.

[2] Scalable metadata filtering for vector search. In Proceedings of the International Conference on Learning Representations / OpenReview, 2025.

[13] H. Jeong, B. Song, M. Jo, and H. Jo. LACX: Localityaware shared data migration in NUMA + CXL tiered memory. Electronics, 14(21):4235, 2025.

[3] Vecflow: Accelerating filtered approximate nearest neighbor search on gpus. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2025.

[14] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. volume 6, page 317, 2019.

[4] Learning with noisy cross-modal correspondences. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026.

[15] Sungyeon Kim, Xinliang Zhu, Xiaofan Lin, Muhammet Bastan, Douglas Gray, and Suha Kwak. Genius: A generative framework for universal multimodal search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19659–19669, 2025.

[5] John Agyekum, Somnath Mazumdar, and Christoph Scheich. Adapt: An effective data-aware multicloud data placement framework. Cluster Computing, 28:825, 2025. [6] Kyle Bader. Closer than you think: Adding nearest neighbor search to ceph. Ceph Blog, March 2025.

[16] Kwanhyung Lee, Soojeong Lee, Sangchul Hahn, Heejung Hyun, Edward Choi, Byungeun Ahn, and Joohyung Lee. Learning missing modal electronic health records with unified multi-modal data embedding and modalityaware attention. In Proceedings of the 8th Machine Learning for Healthcare Conference, volume 219 of Proceedings of Machine Learning Research, pages 423– 442. PMLR, 2023.

[7] Davide Caffagni, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Recurrence-enhanced vision-and-language transformers for robust multimodal document retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9286–9295, 2025. [8] Yue Chen, Kai Zhang, Sipeng Chen, Shihai Xiao, Xiaomin Zou, Ren Ren, Yinan Jing, X. Sean Wang, Li Cao, and Mingxiang Wan. RED-ANNS: An RDMA-enabled distributed framework for graph-based approximate nearest neighbor search. Proceedings of the VLDB Endowment, 19(3):399–412, 2025.

[17] Yihao Lu and Hao Tang. Multimodal data storage and retrieval for embodied ai: A survey. arXiv preprint arXiv:2508.13901, 2025. [18] James Jie Pan, Jianguo Wang, and Guoliang Li. Survey of vector database management systems. The VLDB Journal, 33:1591–1615, 2024.

[9] Yuhao Cui, Xinxing Zu, Wenhua Zhang, Zhongzhou Zhao, and Jinyang Gao. Incorporating dense knowledge alignment into unified multimodal representation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 29733–29743, 2025.

[19] Yujie Ren, David Domingo, Jian Zhang, Paul John, Rekha Pitchumani, Sanidhya Kashyap, and Sudarsun Kannan. PolyStore: Exploiting combined capabilities of heterogeneous storage. In 23rd USENIX Conference on File and Storage Technologies (FAST 25), pages 539– 555. USENIX Association, 2025.

[10] Siyuan Duan, Yuan Sun, Dezhong Peng, Zheng Liu, Xiaomin Song, and Peng Hu. Fuzzy multimodal learning for trusted cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20747–20756, 2025.

[20] Joobo Shim, Jaewon Oh, Hongchan Roh, Jaeyoung Do, and Sang-Won Lee. Turbocharging vector databases using modern ssds. Proceedings of the VLDB Endowment, 18(11):4710–4722, 2025.

[11] Shaza Elsharief, Saeed Shurrab, Baraa Al Jorf, Leopoldo Julian Lechuga Lopez, Krzysztof J. Geras, and Farah E. Shamout. Medmod: Multimodal benchmark for medical prediction tasks with electronic health records and chest x-ray scans. In Proceedings of the sixth Conference

[21] Yitong Song, Xuanhe Zhou, Christian S. Jensen, and Jianliang Xu. Vector search for the future: From memory-resident, static heterogeneous storage, to cloudnative architectures. arXiv preprint arXiv:2601.01937, 2026. 7

[22] Chandan Suri and Gursifath Bhasin. Vextra: A unified middleware abstraction for heterogeneous vector database systems. arXiv preprint arXiv:2601.06727, 2026.

[32] Zhen Yin et al. Distvs: Large-scale vector search. In Proceedings of the 23rd USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2026.

[23] Toni Taipalus. Vector database management systems: Fundamental concepts, use-cases, and current challenges. arXiv preprint arXiv:2309.11322, 2023.

[33] Quanxing Zha, Xin Liu, Shu-Juan Peng, Yiu-ming Cheung, Xing Xu, and Nannan Wang. Recon: Enhancing true correspondence discrimination through relation consistency for robust noisy correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 29680– 29689, 2025.

[24] Wenda Tang, Tianxiang Ai, and Jie Wu. Tiresias: Optimizing NUMA performance with CXL memory and locality-aware process scheduling. In ACM Turing Award Celebration Conference (ACM-TURC), 2024.

[34] et al. Zhang. An optimized learning-based directory placement policy with two-rounds selection in distributed file systems. Future Generation Computer Systems, 154:235–250, 2024.

[25] S. Vengadeswaran, S. R. Balasundaram, and P. Dhavakumar. Idaps – improved data-locality aware data placement strategy based on markov clustering to enhance mapreduce performance on hadoop. Journal of King Saud University Computer and Information Sciences, 36(3):101973, 2024.

[35] Lei Zhang, Reza Karimi, Irfan Ahmad, and Ymir Vigfusson. Optimal data placement for heterogeneous cache, memory, and storage systems. Proceedings of the ACM on Measurement and Analysis of Computing Systems, 4(1):1–27, 2020.

[26] Shenao Wang, Yanjie Zhao, Yinglin Xie, Zhao Liu, Xinyi Hou, Quanchen Zou, and Haoyu Wang. Towards reliable vector database management systems: A software testing roadmap for 2030. arXiv preprint arXiv:2502.20812, 2025.

[36] Xin Zhang, Yanzhao Zhang, Wen Xie, Mingxin Li, Ziqi Dai, Dingkun Long, Pengjun Xie, Meishan Zhang, Wenjie Li, and Min Zhang. Bridging modalities: Improving universal multimodal retrieval by multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9274–9285, 2025.

[27] Wuyuqing Wang, Zeyuan Gu, and Erkun Yang. Noisy correspondence learning with modality gap direction correction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 10163–10171, 2026.

[37] Xiangyu Zhi, Meng Chen, Xiao Yan, Baotong Lu, Hui Li, Qianxi Zhang, Qi Chen, and James Cheng. Towards efficient and scalable distributed vector search with RDMA. In Proceedings of the ACM on Management of Data (SIGMOD/PACMMOD), volume 4, 2025.

[28] Yu Wang, Fengxia Han, and Jianyu Wang. Boosting noisy correspondence discrimination via dynamic neighborhood semantic verification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 26706–26714, 2026.

[38] Yuhong Zhong, Daniel S. Berger, Carl A. Waldspurger, Ryan Wee, Ishwar Agarwal, Rajat Agarwal, Frank Hady, Karthik Kumar, Mark D. Hill, Mosharaf Chowdhury, and Asaf Cidon. Managing memory tiers with CXL in virtualized environments. In 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), pages 37–56, 2024.

[29] Yinglin Xie, Xinyi Hou, Yanjie Zhao, Shenao Wang, Kai Chen, and Haoyu Wang. Toward understanding bugs in vector database management systems. arXiv preprint arXiv:2506.02617, 2025. [30] Jingyi Yang, Songsong Mo, Jiachen Shi, Zihao Yu, Kunhao Shi, Xuchen Ding, and Gao Cong. Arcade: A realtime data system for hybrid and continuous query processing across diverse data modalities. 2025. [31] Wenfang Yao, Kejing Yin, William K. Cheung, Jia Liu, and Jing Qin. Drfuse: Learning disentangled representation for clinical multi-modal fusion with missing modality and modal inconsistency. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16416–16424, 2024. 8

Related documents

• Cross-Modal Data Co-Locality for Heterogeneous Storage Systems#1457749
• Architectural Support for Coherent Architecturally Visible Storage in Instruction Set Extensions#1100332
• Technoeconomic Analysis of Small Modular Reactors for Hyperscale Data Center Applications#124687
• Technoeconomic Analysis of Small Modular Reactors for Hyperscale Data Center Applications#124686
• CVE-2026-93088 — CRITICAL 9.8#1067256