GeoLayer: A Geo-Distributed Graph Storage Framework

• GeoLayer is a geo-distributed graph storage framework that decomposes graphs into latency-aware layers to optimize replica placement and pattern routing.
• It formulates joint optimization as a constrained binary integer programming problem minimizing storage, read, write, and association penalty costs.
• Empirical studies demonstrate 1.34×–3.67× online response speedups and significant offline performance improvements across heterogeneous data centers.

GeoLayer is a geo-distributed graph storage framework that jointly optimizes graph replica placement and pattern request routing in geo-distributed cloud storage systems. It was introduced for graph-structured data whose topology-driven access patterns make conventional replica placement and routing strategies inadequate, particularly when workloads span multiple data centers with heterogeneous wide-area latencies. Its central mechanism is a latency-aware layered graph architecture that decomposes the original graph into local and cross-data-center layers, then uses overlap-centric replica placement and stepwise layered routing to reduce response time and coordination cost for both online graph pattern requests and offline graph analysis (Yao et al., 2 Sep 2025). In a broader and looser sense, recent geospatial and vision literature also uses “GeoLayer-like” ideas to denote layered geographic representations, geometry-aware feature modules, or standardized embedding access layers, but those are analogous constructions rather than the named graph-storage framework itself (Fang et al., 19 Jan 2026).

1. Problem setting and scope

GeoLayer targets geo-distributed cloud storage systems that store graph-structured data across multiple data centers. The workloads of interest include online graph pattern requests, which require low-latency interactive queries and updates, and offline graph analysis, including PageRank and SSSP. The framework is motivated by the observation that graph workloads are not well served by distance-only replication policies or job-driven placement schemes, because graph access is topology-driven, highly diverse, overlapping, and latency heterogeneous; under such conditions, the slowest participating data center becomes a straggler and dominates end-to-end latency (Yao et al., 2 Sep 2025).

The framework formalizes the storage setting using the set of data items $I$, the set of data centers $D$, and the set of read request patterns $P$. Replica placement is encoded by $\delta_{xd}$, item-level serving decisions by $\sigma_{xyd}$, and pattern routing by $\rho_{pyd}$. Read and write rates are represented by $R_{xy}$, $W_{xy}$, and $R_{py}$. This notation is not incidental: it expresses the paper’s core claim that placement and routing cannot be optimized independently, because replica placement changes the feasible routing space, while routing decisions change both latency and cost (Yao et al., 2 Sep 2025).

A recurring misconception is to treat “GeoLayer” as a generic synonym for any layered geographic model. Strictly speaking, the formal term names the geo-distributed graph storage framework of 2025. Other papers describe layered geospatial abstractions, but often explicitly note that “GeoLayer” is not introduced as a formal term and is instead an interpretive label for a standardized embedding layer or layered geographic representation (Fang et al., 19 Jan 2026).

2. Formal optimization model

GeoLayer formulates joint replica placement and routing as a constrained binary integer programming problem. The objective is to minimize the total cost

$$\min_{\delta_{xd},\ \sigma_{xyd},\ \rho_{pyd}} \left(C^{(S)} + C^{(R)} + C^{(W)} + C^{(A)}\right),$$

where $D$0 is storage cost, $D$1 is read request cost, $D$2 is write cost, and $D$3 is an association penalty cost specific to graph workloads (Yao et al., 2 Sep 2025).

The association penalty is the framework’s most graph-specific term. It penalizes both spreading one pattern across multiple data centers and latency imbalance among the participating sites:

$D$4

with

$D$5

The first term encodes locality, and the second encodes straggler sensitivity by measuring latency heterogeneity within a routed pattern (Yao et al., 2 Sep 2025).

The optimization is constrained so that an item can be served only from a data center that stores it, each request is served exactly once, and routed patterns can use only data centers that hold the required items. GeoLayer also imposes an average latency bound and a per-pattern latency bound:

$D$6

and

$D$7

The paper identifies the resulting problem as a binary integer programming problem and therefore NP-hard (Yao et al., 2 Sep 2025).

3. Latency-aware layered graph architecture

The defining structural idea of GeoLayer is the layered graph. The original graph is decomposed into a base local layer and a hierarchy of bridge layers aligned with wide-area latency structure. The local layer is

$D$8

where each $D$9 is the local subgraph of a data center. Above this base layer, the framework constructs a bridge graph from boundary vertices and cross-partition edges. Boundary vertices are

$P$0

and the bridge graph aggregates these cross-data-center dependencies into $P$1 (Yao et al., 2 Sep 2025).

To make WAN heterogeneity explicit, each bridge edge is assigned a latency value and partitioned into latency layers by thresholds

$P$2

with $P$3 when the edge latency falls in the interval $P$4. This yields latency-homogeneous bridge-edge sets $P$5 and corresponding bridge-vertex sets $P$6. The result is a hierarchy of clusters in which lower layers capture tighter locality and upper layers capture broader, more expensive cross-site connectivity (Yao et al., 2 Sep 2025).

This decomposition serves two explicit purposes. First, it reduces the decision space by turning one large global optimization problem into smaller subproblems associated with layers and bridge subgraphs. Second, it mitigates the effect of network heterogeneity by grouping edges with similar latency characteristics. In the paper’s formulation, the layered graph is therefore both a systems abstraction and an optimization device (Yao et al., 2 Sep 2025).

4. Overlap-centric replica placement and directed heat diffusion

GeoLayer’s replica placement scheme is overlap-centric. Rather than treating each query pattern as an isolated unit, it explicitly models overlapping regions shared by multiple patterns. These shared substructures experience superposition of access frequency, so their effective “heat” can exceed that of non-overlapping regions (Yao et al., 2 Sep 2025).

Placement begins with pattern sinking. A pattern is placed in the lowest layer that satisfies its latency constraint; if the requirement cannot be met at a given layer, the pattern is replicated downward to a lower cluster or layer. For a pattern $P$7 placed at bridge subgraph $P$8, the replication gain is

$P$9

If $\delta_{xd}$0, full replication is selected; otherwise, the pattern is decomposed into overlapping regions that compete for placement (Yao et al., 2 Sep 2025).

The competition mechanism is the directed heat diffusion model. GeoLayer interprets vertices or data items as thermal masses, access frequency as heat, and edges as heat-conduction pathways. Each item $\delta_{xd}$1 is initialized with heat

$\delta_{xd}$2

and heat transfer from $\delta_{xd}$3 to $\delta_{xd}$4 at step $\delta_{xd}$5 is defined as

$\delta_{xd}$6

The vertex update is

$\delta_{xd}$7

At equilibrium, the system satisfies

$\delta_{xd}$8

The paper states that if

$\delta_{xd}$9

then there exists a unique non-trivial steady-state solution (Yao et al., 2 Sep 2025).

Placement then assigns a region $\sigma_{xyd}$0 to the bridge subgraph with maximum equilibrium heat:

$\sigma_{xyd}$1

GeoLayer also pre-caches high-heat regions by identifying vertices whose initial or steady-state heat exceeds a threshold $\sigma_{xyd}$2, thereby proactively storing frequently accessed regions before requests arrive (Yao et al., 2 Sep 2025).

5. Stepwise layered routing and empirical behavior

For online workloads, GeoLayer uses bottom-up progressive expansion. Routing starts at the requesting data center in $\sigma_{xyd}$3, searches the linked bridge subgraph in the next layer for missing items, chooses the data center holding the largest subset of requested items, and escalates only if necessary. The design goal is to reduce cross-data-center fanout, minimize the number of participating sites, and avoid unnecessary high-latency traversals (Yao et al., 2 Sep 2025).

For offline workloads, the framework uses top-down localization followed by bottom-up assembly. The top-down phase maps required items to candidate data centers while deferring redundancy resolution. The bottom-up phase decides which data centers should participate in computation and which data should migrate. Migration is triggered when the expected communication overhead exceeds a layer-adjusted threshold, after which bridge subgraphs redistribute migrated data by hashing it to retained data centers in the same cluster. This turns offline routing into a global-to-local planning phase followed by local-to-global execution (Yao et al., 2 Sep 2025).

The evaluation is conducted on Alibaba Cloud across eight geo-distributed data centers—Beijing, Seoul, Singapore, Sydney, Frankfurt, London, Virginia, and Mexico City—using 100 ms latency intervals. The datasets are LDBC-SNB-SF10, UK-2005, and Twitter-2010. Online workloads include LDBC interactive queries and 3-hop random walk patterns; offline workloads use the RAGraph engine for PageRank and SSSP (Yao et al., 2 Sep 2025).

The abstract reports that, relative to state-of-the-art replica placement and routing schemes, GeoLayer achieves a $\sigma_{xyd}$4–$\sigma_{xyd}$5 improvement in response times for online graph pattern requests and a $\sigma_{xyd}$6–$\sigma_{xyd}$7 speedup in offline graph analysis performance (Yao et al., 2 Sep 2025). In the detailed experiments, average online speedups are $\sigma_{xyd}$8 over Random-3, $\sigma_{xyd}$9 over Top-3, $\rho_{pyd}$0 over ADP, and $\rho_{pyd}$1 over DCD, while the average reduction in total objective cost is $\rho_{pyd}$2, $\rho_{pyd}$3, $\rho_{pyd}$4, and $\rho_{pyd}$5, respectively. Under increasing network heterogeneity, GeoLayer remains strongest, with average speedups of $\rho_{pyd}$6, $\rho_{pyd}$7, and $\rho_{pyd}$8, and up to $\rho_{pyd}$9. For offline execution, it achieves average speedups of $R_{xy}$0 over RAGraph, $R_{xy}$1 over RAGraph+, and $R_{xy}$2 over GrapH, while reducing WAN transfer cost by $R_{xy}$3, $R_{xy}$4, and $R_{xy}$5, respectively (Yao et al., 2 Sep 2025).

6. Broader uses of “GeoLayer-like” layered geography

Outside geo-distributed graph stores, several papers describe layered geographic mechanisms that are closely related in spirit but are not the same framework. One line of work treats geospatial embeddings as a reusable substrate. “Earth Embeddings as Products: Taxonomy, Ecosystem, and Standardized Access” formalizes a three-layer taxonomy—Embeddings (Data), Analysis Frameworks (Tools), and Downstream Applications (Value)—and extends TorchGeo with a unified API so that embedding products can be loaded and queried as first-class geospatial datasets. The paper explicitly states that “GeoLayer” is not introduced as a formal term, but it effectively sketches a standardized geospatial embedding access layer (Fang et al., 19 Jan 2026).

Another line uses layered spatial hierarchies for geolocation itself. “Spatial Language Representation with Multi-Level Geocoding” represents the Earth with hierarchical S2 cells, trains a CNN-based text geocoder with a multi-level loss over levels $R_{xy}$6–$R_{xy}$7, and combines coarse-to-fine evidence at inference so that fine cells are supported by their parent cells. The paper describes this as balancing “specificity at finer levels and aggregation/smoothing at coarser levels,” and characterizes the method as a GeoLayer-like layered geographic representation for text-to-coordinate prediction (Kulkarni et al., 2020). “HierLoc” similarly organizes geography into country, region, subregion, and city entities, embeds them in hyperbolic space, and performs hierarchical beam search over about $R_{xy}$8k entity embeddings instead of over $R_{xy}$9 million image embeddings; in that sense, it instantiates a geography-layered model for visual geolocation rather than a general GeoLayer framework (Gadi et al., 30 Jan 2026).

A third cluster of work uses geometry-aware layers to align modalities or impose structure. “Image-based Geolocalization by Ground-to-2.5D Map Matching” introduces projection-based pixel-to-point fusion between 2D map tiles and 2.5D structural maps, which the paper presents as a geometry-aware alignment mechanism for cross-view retrieval (Zhou et al., 2023). “GeoDTR+” uses an enhanced Geometric Layout Extractor together with geometric modulation $W_{xy}$0 to disentangle layout from appearance in cross-view geo-localization, and describes that module as the closest analogue to a geometry-aware layer (Zhang et al., 2023). “GeoLayout” predicts dominant planar depth maps and derives room layout from plane intersections, while GAL formulates outdoor scene interpretation in terms of sky, support, planar, porous, and solid geometric layers; both papers show that layered geometric structure can be more informative than direct edge, corner, or local appearance prediction alone (Zhang et al., 2020, Ren et al., 2016).

A more explicitly infrastructural example is LIGHTHOUSE, whose acronym expands to “Layered Iterative Geospatial Hierarchical Terrain-Oriented Unified Search Engine.” It separates a global $W_{xy}$1-meter coastline data layer from a hierarchical query-serving layer based on $W_{xy}$2 degree tiles, BallTrees, Haversine distance, spherical Voronoi routing, and caching, thereby realizing a layered geospatial search stack for distance-to-shoreline queries (Beukema et al., 23 Jun 2025).

Taken together, these papers suggest that “GeoLayer” has acquired a broader descriptive role for architectures that separate geospatial problems into explicit layers—data layers, hierarchy layers, geometry layers, or routing layers—even though only the geo-distributed graph store paper formalizes GeoLayer as the name of a specific system (Yao et al., 2 Sep 2025).