Sensor-Driven Load Migration

• Sensor-Driven Load Migration is a mechanism that leverages real-time sensor data (e.g., location, resource usage, queue states) to determine when and where to shift workloads.
• It applies across diverse settings such as mobile fog networks, IoT aggregators, edge object detection, and electricity demand management, achieving measurable improvements in latency and efficiency.
• Optimization frameworks employ closed-loop controllers, heuristics, and techniques like Particle Swarm Optimization to balance task completion time, resource utilization, and economic cost.

Sensor-driven load migration denotes a class of mechanisms in which monitored signals from devices, networks, infrastructures, markets, or environments determine when and where load should be relocated. In the current literature, the migrated entity may be a migratory module in a mobile fog network, a work package produced by a mobile sensor, a live industrial control service, an inference request in a heterogeneous edge gateway, a dominant data interval associated with pervasive tasks, an EV charging demand across aggregators, or a fraction of electricity demand across locations. Across these settings, sensing is not merely observational: user location history, queue state, scene complexity, request-density structure, state-of-charge, resource utilization, and price signals are incorporated into decision rules whose objectives include lower task completion time, lower downtime, improved utilization, reduced cloud offloading, lower cumulative utility loss, or lower procurement cost (Seraj et al., 2024, Shivaraman et al., 2024, Alqahtani et al., 16 Mar 2026, Boulougaris et al., 2024, Gundall et al., 2022, Mollaeivaneghi et al., 2023).

1. Scope and domain variants

The term spans both cyber load migration and physical or electrical load migration. In cyber settings, the relocated object is computation, application state, data, or a service instance; in energy settings, the relocated object is a flexible demand that can be shifted geographically across locations or aggregators. The literature therefore treats “load” as a workload, a service, a data interval, or an electricity demand, depending on the substrate and control objective (Seraj et al., 2024, Chang et al., 2018, Rocha et al., 2019, Shivaraman et al., 2024, Mollaeivaneghi et al., 2023, Boulougaris et al., 2024, Alqahtani et al., 16 Mar 2026, Gundall et al., 2022, Abdelwahab et al., 2016).

Domain	Migrated entity	Principal signals
Mobile fog networks	Migratory modules $T_1 \ldots T_n$	User location history, CPU, memory, bandwidth, $D_{pd}$, $\sigma$
IoT-enabled aggregators	Device charging load and device migration	SoC, remaining demand, deadline, charging modes, $S_a(t)$, $q_a(t)$
Multi-location electricity consumers	Fraction $\alpha$ of location demand	$D_n^{base}$, $E_p^{gen}$, $\pi_{g,b}$, $\pi_{g,s}^n$
Edge object detection	Per-request inference workload	Scene complexity group $D_{pd}$0, queue length $D_{pd}$1
Pervasive edge data management	Dominant interval $D_{pd}$2	Request endpoints, density contours, OCSVM inliers, frequency threshold $D_{pd}$3
Industrial real-time services	Control-service instance	Timestamped duplicated inputs, output deviation $D_{pd}$4, $D_{pd}$5, $D_{pd}$6

A recurring architectural pattern is a distributed hierarchy. LIMO adopts a three-tier architecture with IoT edge devices, fog nodes, and the cloud (Seraj et al., 2024). ABEONA likewise uses edge, fog, and cloud layers with metrics probes, a fog manager, and a global controller (Rocha et al., 2019). EPIoT and REM place Dynamic Code Execution on mobile sensors, fog nodes, mist nodes, and cloud nodes (Chang et al., 2018). Industrial real-time migration separates source and destination hosts and further separates real-time traffic from management traffic (Gundall et al., 2022). Flock generalizes the setting to VMs hosted on heterogeneous edge and conventional clouds (Abdelwahab et al., 2016).

A common misconception is that load migration is synonymous with time-shifting. The EV aggregator literature states that existing demand response solutions “cannot achieve significant improvements based only on time-shifting the loads without considering the device properties such as charging modes and movement capabilities to enable geographic migration” (Shivaraman et al., 2024). In the same way, fog and edge systems do not merely defer execution; they relocate execution to maintain proximity, reduce latency, or rebalance utilization.

2. Sensed state and observability

Sensor-driven migration is defined by the signals it monitors. In LIMO, the Monitor stage records user location history $D_{pd}$7, coordinates $D_{pd}$8 per second, fog-node CPU, memory, and bandwidth usage, available bandwidth at target $D_{pd}$9, propagation delay $\sigma$0, migration time derived from state size and bandwidth, and the standard deviation of fog-node loads. The location stream follows real vehicular traces with granularity $\sigma$1 s, and the Analyze stage predicts the next position $\sigma$2 from the time-based location history. LIMO explicitly focuses on location and resource utilization; energy, queue length, and RTT are not explicitly modeled (Seraj et al., 2024).

In the aggregator setting, sensing is both device-centric and infrastructure-centric. Each device provides state-of-charge $\sigma$3, remaining demand $\sigma$4, deadline $\sigma$5, charging mode capabilities $\sigma$6, location and feasible travel options with travel time $\sigma$7 and movement cost $\sigma$8, and a mobility flag $\sigma$9. Each aggregator publishes capacity $S_a(t)$0, spare capacity $S_a(t)$1, and queue length $S_a(t)$2. Updates occur in discrete-time slots of duration $S_a(t)$3, and no future arrival information is assumed (Shivaraman et al., 2024).

In heterogeneous edge-based object detection, the key sensed variables are request intensity and scene complexity. The gateway maintains a current scene complexity estimate $S_a(t)$4 by reusing the previous frame’s detected object count, while request intensity is observed through instantaneous queue length $S_a(t)$5 for each device–model pair via HAProxy. These signals drive a per-request decision rather than a periodic batch decision (Alqahtani et al., 16 Mar 2026).

In proactive dominant data migration, the sensor-driven signal is the observed access pattern over request endpoints. Autonomous edge receptors collect a sliding window of interval requests $S_a(t)$6, compute request density over the two-dimensional endpoint space with MKDE, train an OCSVM on boundary points, and maintain a frequency map $S_a(t)$7 that records how often requestor $S_a(t)$8 offloads to receptor $S_a(t)$9 (Boulougaris et al., 2024).

Industrial and edge-to-cloud systems extend observability to systems telemetry. ABEONA’s probes constantly monitor power, CPU, memory, network, and lifecycle events, with power measured by PowerSpy devices and stored in InfluxDB (Rocha et al., 2019). Downtime-optimized real-time migration monitors one-way latency, jitter, packet loss, CPU and GPU load, thermal state, control-task execution time versus deadline, and output deviation between source and destination controllers (Gundall et al., 2022). REM and EPIoT rely on local and remote Service Description Metadata containing CPU benchmark or core count, RAM size, disk read or write speeds, and measured link throughput to peers (Chang et al., 2018).

3. Control logic and optimization formulations

A large part of the literature instantiates sensor-driven migration as a closed-loop controller. LIMO is the clearest case: it uses MAPE, with Monitor, Analyze, Plan, and Execute stages, and the Plan stage contains two periodic sub-policies, mobility management and load balancing. Migration is triggered if $q_a(t)$0 meters, in which case a candidate fog node is sought and otherwise the task is migrated to the cloud. Load balancing is triggered when the standard deviation $q_a(t)$1 of fog-node loads exceeds a threshold. The optimization objective couples task completion time and utilization:

$q_a(t)$2

$q_a(t)$3

The weighting is adaptive: if the system is balanced or almost balanced, $q_a(t)$4 and $q_a(t)$5; if it is unbalanced, $q_a(t)$6 and $q_a(t)$7. Particle Swarm Optimization then searches over discrete task-to-node assignments subject to CPU, memory, and bandwidth constraints (Seraj et al., 2024).

The EV aggregator literature formalizes a different controller: a mixed-integer nonlinear problem, then an online distributed heuristic. The heuristic computes the accumulated progress $q_a(t)$8 of each device, derives a piecewise priority score $q_a(t)$9 from remaining demand and time-to-deadline, allocates charging modes in descending priority, and then allows unserved mobile devices to compare stay-loss against migration-loss. The migration score includes predicted deadline loss, mobility penalty, and a spare-capacity term $\alpha$0. The decisive feature is that geographic migration and multi-mode charging are optimized jointly rather than separately (Shivaraman et al., 2024).

In heterogeneous edge object detection, the decision rule is explicitly two-stage. First, accuracy-aware filtering defines the candidate set

$\alpha$1

where $\alpha$2. Second, expected latency is estimated as

$\alpha$3

and the final score is

$\alpha$4

The selected device–model pair is the minimizer of $\alpha$5 over the accuracy-feasible set. This formulation makes scene complexity a filter and queue state a load-balancing penalty (Alqahtani et al., 16 Mar 2026).

A different formalism appears in multi-location electricity consumption. There, migration capability is parameterized by $\alpha$6, which bounds the deviation of location demand:

$\alpha$7

DER owners strategically set prices and the MLEC solves a cost-minimization problem subject to migration bounds. When $\alpha$8, simulations place offered prices near the grid selling price; as $\alpha$9 grows, offered prices decline toward the grid buying price. The paper does not explicitly reference sensors, but it also describes a sensor-driven adaptation in which real-time telemetry updates $D_n^{base}$0, $D_n^{base}$1, price signals, and $D_n^{base}$2 in a receding-horizon policy (Mollaeivaneghi et al., 2023).

Two additional optimization idioms are influential. REM performs greedy assignment by repeatedly allocating one more data object to the worker with the smallest current estimated completion time, using a latency model that combines packing, transmission, unpacking, processing, output packing, output transfer, and output unpacking (Chang et al., 2018). Flock uses asynchronous threshold-based migration among clouds:

$D_n^{base}$3

with regularization $D_n^{base}$4 discouraging crowding and with convergence to a Nash Equilibrium under the paper’s assumptions (Abdelwahab et al., 2016). These works show that “sensor-driven” need not imply a centralized controller; it may also mean autonomous, locally informed migration.

4. Migration enactment and execution semantics

The execution side of load migration ranges from stateless request steering to stateful live migration. LIMO’s Execute stage performs fog-to-fog or fog-to-cloud migration and models the migration cost as $D_n^{base}$5 plus propagation delay $D_n^{base}$6. Module state is transferred atomically to preserve consistency (Seraj et al., 2024). In the object-detection setting, by contrast, requests are stateless per frame, so forwarding a frame to a different device–model pair from one request to the next does not require checkpoint transfer or preemption (Alqahtani et al., 16 Mar 2026).

EPIoT and REM implement migration by composing “work packages” that contain code, dependencies, metadata, and a subset of data objects. The EPIoT host discovers peer DCE servers, measures local baseline timings, partitions data into multiple work packages, and invokes fog, mist, or cloud resources. The prototype is implemented in Node.js, with Android execution via Termux, and execution occurs in isolated sandboxes with restricted APIs to local resources (Chang et al., 2018).

ABEONA moves closer to an orchestrated edge–fog–cloud control plane. Metrics probes feed telemetry to a controller that includes a metrics analyzer, a migration manager, and a global scheduler; the fog layer also includes a custom fog manager. The prototype fog cluster is Kubernetes-based, with three Raspberry Pi 3B+ worker nodes, PowerSpy devices for power measurements, and InfluxDB for telemetry storage. The paper does not detail CRIU-based stateful container migration or VM live migration; migration is described as task re-placement and re-scheduling across layers (Rocha et al., 2019).

The industrial real-time literature is the most exacting about execution semantics. The proposed concept is a downtime-optimized variant of Parallel Process Migration. In Option 1, the destination buffers timestamped inputs, starts a “state-zero” container, receives the control state vector $D_n^{base}$7, replays buffered input packets, and hands over only when output deviation $D_n^{base}$8 falls below thresholds $D_n^{base}$9. In Option 2, a redundant control container is started on the source, checkpointed with CRIU, restored on the destination, and then driven to convergence by replay. Time synchronization is provided by TSN and linuxptp, management traffic is separated from real-time traffic through dual NICs, and handover is aligned with the cyclic control schedule (Gundall et al., 2022).

At the VM level, Flock assumes the existence of live migration mechanisms across heterogeneous edge and conventional clouds but abstracts away from implementation details. The protocol is primarily concerned with decision logic and stability, not with compatibility or state-transfer internals (Abdelwahab et al., 2016).

5. Empirical results and application domains

The literature reports measurable gains across several substrates, although the reported metrics differ substantially.

System	Setting	Reported outcome
LIMO (Seraj et al., 2024)	MobFogSim; Val-de-Marne traces; scenarios include 15 FNs; 25, 50, 100, and 200 tasks	Network resource utilization with 10% improvement; task migration to cloud reduced by more than 15%; request response time reduced by 18%
IoT-enabled mobile-device heuristic (Shivaraman et al., 2024)	Synthetic data and ACN-Data	Loss improvement of at least 57.23%; SDLM loss $E_p^{gen}$0 vs EDF $E_p^{gen}$1 and highest-power $E_p^{gen}$2; heuristic computed schedules in $E_p^{gen}$3 minute total, $E_p^{gen}$4 s per slot
Multi-objective edge object detection (Alqahtani et al., 16 Mar 2026)	417 JPEG frames; 5 devices; concurrency varied via Locust users $E_p^{gen}$5	Halves energy consumption; more than 80% reduction in average response time relative to HA; up to 10% decrease in detection accuracy relative to an accuracy-centric baseline
Downtime-optimized industrial migration (Gundall et al., 2022)	3 mini PCs; 5 ms control update interval	Sub-millisecond downtime; $E_p^{gen}$6–$E_p^{gen}$7 ms; Option 1 around $E_p^{gen}$8 ms when the image is not pre-placed and around $E_p^{gen}$9 ms when the image is already on DST; Option 2 below $\pi_{g,b}$0 s
Proactive dominant data migration (Boulougaris et al., 2024)	$\pi_{g,b}$1 nodes; trips $\pi_{g,b}$2; $\pi_{g,b}$3	Migration on $\pi_{g,b}$4 of requests for $\pi_{g,b}$5 and $\pi_{g,b}$6 for $\pi_{g,b}$7; average number of migrations $\pi_{g,b}$8 and $\pi_{g,b}$9; $\pi_{g,s}^n$0 is $\pi_{g,s}^n$1–$\pi_{g,s}^n$2 higher than DBSCAN-DM
DER pricing with migration capability (Mollaeivaneghi et al., 2023)	$\pi_{g,s}^n$3 and $\pi_{g,s}^n$4 multi-location cases	Reported MLEC bill reductions: $\pi_{g,s}^n$5 for $\pi_{g,s}^n$6 and $\pi_{g,s}^n$7 for $\pi_{g,s}^n$8

The ABEONA report is preliminary and intentionally cautious. It states that horizontal scaling from one to two and three fog nodes produces a “considerable decrease in the runtime and energy consumption as the number of nodes increase,” but exact Joules, percentages, latency distributions, and variance are not published (Rocha et al., 2019). REM likewise emphasizes relative superiority: it reports that REM consistently outperformed naïve equal partitioning and mono-cloud migration across the tested scenarios because it adapts to both network and compute heterogeneity (Chang et al., 2018).

These evaluations also reveal the breadth of the field. Vehicular edge traces, EV charging, real-time control loops, visual inference, taxi-trip-driven request generation, and local energy markets all appear under a common migration vocabulary, yet each domain uses different observables, constraints, and success criteria.

6. Limitations, misconceptions, and open directions

Several limitations recur. In LIMO, task deadlines are not modeled, and energy, queue length, RTT, and multi-tenant interference are not explicitly modeled; the paper also notes that threshold choice can produce oscillations and that rapid mobility may trigger “thrashing” if hysteresis is not applied (Seraj et al., 2024). In the EV heuristic, no future arrival information is assumed, exact real-time MINLP solving is impractical, and migration effectiveness falls when travel energy and long charging times dominate (Shivaraman et al., 2024). In the object-detection system, offline profiling may drift with firmware or ambient changes, the score uses queue length directly rather than an explicit stochastic queueing model, and tie-breaking or hysteresis are not specified (Alqahtani et al., 16 Mar 2026).

A second limitation concerns scope. Some papers are explicitly sensor-driven, while others are only adaptable to sensor-driven operation. The DER pricing paper states that it “does not explicitly reference sensors,” and its sensor-driven implementation is presented as an adaptation in which telemetry updates demand, generation, and prices (Mollaeivaneghi et al., 2023). Flock similarly relies on regularized latency measurements and stochastic approximation rather than on a stated sensor stack; any augmentation must preserve reciprocity and monotonicity to retain the paper’s convergence and Price of Anarchy guarantees (Abdelwahab et al., 2016). Proactive dominant data migration does not formulate latency, bandwidth, energy, or storage costs explicitly, and return or eviction policies are not specified (Boulougaris et al., 2024).

Privacy and security are also unevenly handled. LIMO identifies security and privacy as future work (Seraj et al., 2024). ABEONA discusses local preprocessing, planned use of Intel SGX and ARM TrustZone, and the need for TLS and access control, but its preliminary evaluation is centered on energy-aware orchestration rather than secure migration (Rocha et al., 2019). Industrial real-time work assumes controlled networking and synchronized hosts, which is appropriate for TSN-enabled plants but narrower than open Internet deployments (Gundall et al., 2022).

The literature also corrects a broader misconception: migration is not always about minimizing one metric. Industrial control prioritizes downtime $\pi_{g,s}^n$9 over total migration time $D_{pd}$00 (Gundall et al., 2022); Flock regularizes latency with crowding penalties and analyzes equilibrium behavior (Abdelwahab et al., 2016); the object-detection balancer constrains accuracy before trading latency against energy (Alqahtani et al., 16 Mar 2026); and the DER model shows that migration flexibility changes market power rather than merely changing utilization (Mollaeivaneghi et al., 2023). This suggests that sensor-driven load migration is best understood not as a single algorithmic family but as a family of control problems in which monitored signals alter feasible actions, objective weights, and even the economic structure of the system.