Shared-Resource (SR) Overview

• Shared-Resource (SR) is a technical abstraction for concurrently accessed resources, emphasizing coordinated access under interference, contention, and reuse constraints.
• SR spans applications from wireless spectrum sharing and shared memory in multicore systems to fair scheduling and quantum entanglement reuse.
• Research on SR focuses on domain-specific resource models, optimization criteria, and security implications to ensure effective and robust allocation strategies.

Shared-Resource (SR) is a recurring technical abstraction for a resource, state, or infrastructure element that is concurrently accessed, scheduled, protected, or reused by multiple agents. In the cited literature, SR denotes shared cellular spectrum and uplink resource blocks, non-partitionable DRAM, cache, and interconnect resources, divisible online capacity, repeated single-item allocation pools, heap-resident program objects, globally shared multiagent resource endowments, switch-resident network-function pools, and reusable entangled states (Han et al., 2014, 1711.01866, Song et al., 2018, Damerius et al., 2023, Schill et al., 2014, Bulling et al., 2015, Du et al., 2018, Roy et al., 2019). Across these settings, the common problem is not the material form of the resource but the coordination of access under interference, contention, depletion, or reuse.

1. Scope and conceptual structure

In contemporary research usage, SR is not a single standardized object. In wireless systems it can mean a shared radio band or a shared region of uplink resource blocks. In multicore and MPSoC work it can mean DRAM controller queues, memory channels and banks, NoC routers and links, cache ports, MSHRs, fill and writeback buffers, memory bandwidth, and shared interconnect. In scheduling theory it can mean a single divisible resource of total capacity $1$, or a single item allocated in each round. In concurrency and verification it can mean shared heap objects or global resource endowments. In quantum information it can mean a shared entangled state (1711.01866, Song et al., 2018, Damerius et al., 2023, Schill et al., 2014, Bulling et al., 2015, Roy et al., 2019).

Domain	Meaning of SR	Representative source
Cellular and wireless systems	shared spectrum, shared RBs, shared power, shared infrastructure	(Han et al., 2014, 1711.01866, Carrasco et al., 2017)
Computer architecture and clouds	DRAM, LLC, NoC, memory bandwidth, cache, shared interconnect	(Song et al., 2018, Subramanian, 2015, Monaco et al., 2023)
Scheduling and allocation theory	divisible capacity, repeated resource pools, multi-resource server capacity	(Damerius et al., 2023, Fikioris et al., 2023, Khamse-Ashari et al., 2016)
Correctness and security	heap objects, shared ownership graphs, shared caches and pages	(Schill et al., 2014, Yan et al., 2018)
Formal verification	single or multiple globally shared resources in ATL-style models	(Bulling et al., 2015)
Quantum information	reusable shared entangled state	(Roy et al., 2019)

This suggests that SR is best understood as a family of models rather than a single mechanism. The cited work repeatedly specifies four ingredients: the resource model, the admissible access semantics, the contention or interference mechanism, and the optimization or correctness criterion. The resulting objectives vary sharply—spectral efficiency, QoS protection, slowdown estimation, fairness, race freedom, decidability, architecture extraction, or entanglement reuse—but the shared mathematical structure is constrained access to a common pool.

2. Wireless, spectrum, and networked infrastructure

In cellular systems, SR appears first as spectrum refarming: an OFDMA system reuses the spectrum licensed to a legacy CDMA system during a transition period in which the CDMA load is below its designed capacity. The underlay model in "Spectrum Refarming: A New Paradigm of Spectrum Sharing for Cellular Networks" places OFDMA and CDMA in the same wideband channel and at a common BS site, with OFDMA exploiting the interference margin tolerable by CDMA. For the CDMA matched-filter receiver, the asymptotic SINR is

$$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$

and the interference margin is derived from the gap between actual CDMA load and the maximum supportable load at target SINR. The OFDMA uplink resource-allocation problem is then solved by dual decomposition, yielding a water-filling-like power law, while requiring only system-level parameters from CDMA and no upgrade of the legacy CDMA system (Han et al., 2014).

A second SR meaning in wireless networking is in-band underlay reuse of uplink RBs by device-to-device pairs. In the combined shared/dedicated framework for D2D communication, the shared mode lets D2D pairs reuse RBs allocated to cellular users, while the dedicated mode isolates D2D traffic from CUE–D2D interference. Reuse eligibility is governed by the D2D-side SINR threshold $\gamma_j^{s,r} \ge \gamma_{\min}$ and by a strict power restriction that keeps DUE-T interference at the eNB ten times below $NI_e$ when $\tau_{\mathrm{DUE}}=10$. Graph construction, decomposition into per-CUE subgraphs, and maximal-clique enumeration then determine which D2D pairs may simultaneously reuse the same shared RBs. Under the reported settings, combined use of shared and dedicated regions increases sum D2D capacity by $1.67$ to $2.5$ times relative to schemes that select only one mode (1711.01866).

SR is also institutionalized in Licensed Shared Access and dense small-cell control. In the SPEED-5G architecture, centralized RRM in the edge cloud manages shared LSA spectrum, shared RBs, shared transmit power, and shared infrastructure under MOCN and network-slicing assumptions. The cRRM interacts with the LSA controller and repository, enforces incumbent protection thresholds, maintains radio coverage maps through cooperative sensing, and reconfigures carriers, power masks, ABS patterns, and bearer mappings at TTI granularity. The formalization includes assignment variables $x_{k,m}$, power $p_{k,m}$, SINR constraints, LSA availability $B_s(t)$, and interference-to-incumbent threshold $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$0 (Carrasco et al., 2017).

An adjacent networking meaning of SR appears in the Resource Pooling Switch Architecture. There the global shared resource pool is the Network Function Pool: compute resources hosting NFIs are not permanently bound to a fixed line-card pipeline. Service Function Chains are realized through this pool, and the resulting non-uniform traffic motivates the Balanced Service Capacity based FIRM scheduler. BSC-FIRM changes FIRM only in the first request phase, where each input points to the VOQ with minimum service-capacity status, and the paper reports lower queuing delay and packet-loss rate than FIRM and iSLIP under hotspot and burst traffic (Du et al., 2018).

3. Shared compute and memory subsystems

In heterogeneous MPSoCs, SR denotes non-partitionable memory-system components that require instant response and cannot be managed effectively by slow partitioning. SARA identifies DRAM controller and command queues, memory channels and banks, NoC routers and links, and shared cache ports, MSHRs, and buffers as the relevant SR set. Each DMA engine computes a core-specific Normalized Performance Indicator and maps it to a discrete priority. For example,

$$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$1

and the system responds by priority-based arbitration in routers and the DRAM controller, with aging and an optional row-buffer-aware policy. In the reported evaluation, SARA keeps NPIs for critical cores at or above $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$2, while QoS-RB delivers DRAM bandwidth only $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$3 lower than FR-FCFS and still maintains QoS (Song et al., 2018).

A closely related SR literature addresses slowdown control in multicore systems. One line of work treats DRAM bandwidth and the shared last-level cache as the dominant interference channels and proposes mitigation, estimation, and control mechanisms. BLISS blacklists applications that receive too many consecutive memory services and then prioritizes non-blacklisted applications, achieving higher performance and fairness than prior ranking schedulers with much lower critical-path and area cost. MISE estimates memory-interference-induced slowdown through request service rates,

$$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$4

for memory-bound phases, and augments this with a stall fraction $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$5 for non-memory-bound phases. ASM extends slowdown estimation to shared-cache interference by using cache-access rate and an auxiliary tag store. Reported average estimation errors are $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$6 for MISE and $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$7 for ASM, enabling soft guarantees, fairness controllers, and slowdown-aware cache and memory partitioning (Subramanian, 2015).

In RT-cloud orchestration, SR is elevated to a first-class scheduling object in Kubernetes. The proposed architecture models memory bandwidth, caches, and shared bus or interconnect as node-local shared resources, monitors them via PMU-based Linux mechanisms, abstracts them into resource bands, and exports them as extended resources with strict-versus-loose annotations. Placement can then be constrained by

$$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$8

where $$\gamma^{a({\rm MF})} = \frac{q}{\alpha q + \overline{\sigma_n^2} + \sigma^2}, \qquad \overline{\sigma_n^2} = \frac{1}{N}\sum_{n=1}^N \sum_{k=1}^K p_{k,n} g_{k,n},$$9 is the number of bands of resource $\gamma_j^{s,r} \ge \gamma_{\min}$0 on node $\gamma_j^{s,r} \ge \gamma_{\min}$1. Enforcement is delegated to local resource managers that throttle containers exceeding their band budgets and report status to a global resource manager for rebalancing or rescheduling (Monaco et al., 2023).

A more recent observability-oriented formulation defines SR bottlenecks across CPU, LLC, memory bandwidth, I/O, network, and accelerators. Buoyancy combines application-level KPI slack with resource scores in $\gamma_j^{s,r} \ge \gamma_{\min}$2, using

$\gamma_j^{s,r} \ge \gamma_{\min}$3

with $\gamma_j^{s,r} \ge \gamma_{\min}$4 in the reported experiments. The limiting resource is $\gamma_j^{s,r} \ge \gamma_{\min}$5. On the evaluated workloads, buoyancy provides a $\gamma_j^{s,r} \ge \gamma_{\min}$6 better indication of bottlenecks than traditional heuristics on average and is presented as a drop-in replacement for conventional performance metrics in orchestration loops (Larsson et al., 26 Feb 2026).

4. Scheduling, fairness, markets, and pooling

In scheduling theory, SR can mean a single divisible resource of total capacity $\gamma_j^{s,r} \ge \gamma_{\min}$7 that is fractionally allocated over time. In "Improved Scheduling with a Shared Resource," each job $\gamma_j^{s,r} \ge \gamma_{\min}$8 has a processing volume $\gamma_j^{s,r} \ge \gamma_{\min}$9 and a resource requirement $NI_e$0, and a feasible schedule assigns $NI_e$1 with $NI_e$2 and $NI_e$3. The online makespan problem admits an optimal $NI_e$4-competitive water-filling algorithm with runtime $NI_e$5, while offline total completion time admits a $NI_e$6-approximation via a continuous linear program, primal–dual geometry, and a greedy fallback (Damerius et al., 2023).

A different SR literature studies repeated allocation over rounds. Dynamic Max-min Fairness allocates the item in each round to the requester with the least number of allocations received to date, normalized by fair share $NI_e$7, and in large markets yields a $NI_e$8 fraction of ideal utility under mild distributional assumptions. Credit Fairness strengthens sharing incentives by introducing credit balances $NI_e$9 that record whether an agent has lent or borrowed effective capacity over time, proves that no mechanism can be anonymous, strategyproof, credit fair, and Pareto efficient simultaneously, and proposes LendRecoup as a credit-fair, Pareto-efficient, online-strategyproof mechanism. Robust Equilibria in Shared Resource Allocation shows that a natural threshold strategy can be both a Bayes–Nash equilibrium and robust, with equilibrium utility factor $\tau_{\mathrm{DUE}}=10$0 and robust factor $\tau_{\mathrm{DUE}}=10$1, via a strengthened Border-type feasibility argument (Fikioris et al., 2023, Zahedi et al., 25 Jan 2026, Lin et al., 16 May 2025).

SR also appears as multi-resource fair allocation across heterogeneous servers. PS-DSF replaces a pooled dominant share by a per-server virtual dominant share $\tau_{\mathrm{DUE}}=10$2, where $\tau_{\mathrm{DUE}}=10$3 is the number of tasks user $\tau_{\mathrm{DUE}}=10$4 could execute if it monopolized server $\tau_{\mathrm{DUE}}=10$5. The mechanism satisfies single-resource fairness, bottleneck fairness, envy freeness, and sharing incentive, and under TDM it is Pareto optimal and strategyproof (Khamse-Ashari et al., 2016). In queueing-based resource pooling, partial server pooling allows provider $\tau_{\mathrm{DUE}}=10$6 to contribute $\tau_{\mathrm{DUE}}=10$7 servers to a common pool. Under cancel-on-complete redundancy, full pooling is the only Pareto-optimal configuration for mean response time; under cancel-on-start, the Pareto frontier is conjectured to lie on boundary points where at least one provider fully shares (Mete et al., 2019).

An explicitly economic SR model is SC-Share, where autonomous small clouds lend and borrow VMs. Each cloud chooses a maximum share level $\tau_{\mathrm{DUE}}=10$8, faces public-cloud price $\tau_{\mathrm{DUE}}=10$9 and federation price $1.67$0, and optimizes a utility that combines cost reduction and utilization change. The coupled stochastic performance model and repeated best-response market are reported to converge empirically to efficient sharing decisions, with different $1.67$1 regimes favoring utilitarian, proportional-fair, or max-min-fair outcomes (Lin et al., 2017).

5. Correctness, verification, and attack surfaces

In shared-memory concurrency, SR denotes heap-resident objects, arrays, and other data structures that can be accessed by multiple threads or processes. Shared ownership makes the protective relationship between synchronization and resource explicit through an acyclic ownership graph $1.67$2. Processes have no incoming edges, every resource has at least one incoming edge, and access privileges are determined by root ownership: $1.67$3 The operational model supports read, write, pass, share, release, spawn, and allocate, and the core theorem states that conflicting reducible accesses cannot coexist; this yields dynamic race detection when SOM statements are simulated alongside execution (Schill et al., 2014).

In formal verification, SR becomes a global endowment available to all agents in an ATL-style multiagent model. Resources can be consumed and produced up to any amount, and strategies are interpreted over resource-extended paths with opponents-first feasibility. The decisive complexity boundary is the number of unbounded shared resources: model checking is undecidable with two or more such resources, by reduction from two-counter machines, but decidable with exactly one unbounded shared resource by reduction to CTL model checking over alternating Büchi pushdown systems, using compact ABPDS to encode multi-unit consumption and production (Bulling et al., 2015).

The same sharedness that enables efficiency and reuse also creates attack surfaces. Cache Telepathy exploits shared LLCs and shared memory pages to recover DNN architecture information from tiled GEMM behavior through Prime+Probe and Flush+Reload. By mapping observed GEMM call structure and matrix dimensions to layer hyperparameters, the attack reduces the VGG-16 search space from more than $1.67$4 architectures to $1.67$5 candidates when OpenBLAS is used, and reduces the ResNet-50 search space to $1.67$6 candidates under OpenBLAS (Yan et al., 2018). A common misconception is therefore that SR is only a performance or utilization primitive; the security literature shows that the same contention and observability channels can expose confidential structure.

6. Shared quantum states and controlled reuse

In quantum information, SR denotes the entangled state shared between sender and receiver in teleportation. Standard projective Bell measurement maximizes information gain but destroys the shared entanglement. The cited work replaces the projective Bell measurement on $1.67$7 by a weak POVM,

$1.67$8

so that the teleportation round still yields quantum advantage while leaving a degraded but reusable entangled state. For a maximally entangled initial SR, the first-round fidelity is

$1.67$9

and the recycled SR is a Werner state with mixing parameter

$2.5$0

Across successive rounds, the Werner parameter multiplies, $2.5$1, and the $2.5$2-th round fidelity becomes $2.5$3 (Roy et al., 2019).

With the criterion $2.5$4 in every round, the maximal reattempt number for a maximally entangled initial SR is $2.5$5, and the same number persists for a finite neighborhood of nonmaximally entangled pure states, specifically $2.5$6 in the Schmidt-form family $2.5$7. The reusability number then decreases as initial entanglement decreases. Among the weakening strategies examined, Bell measurement admixed with white noise performs better than the alternative low-rank weak measurements, which either immediately produce separable recycled states or preserve entanglement over a narrower parameter range (Roy et al., 2019).

Taken together, these results show that SR is not merely a sharing arrangement but often a consumable asset whose degradation law matters. In radio systems the limiting quantity is interference margin; in multicore systems it is contention-induced slowdown; in online allocation it is fairness under repeated use; in formal methods it is decidability under unbounded production and consumption; in security it is side-channel observability; and in quantum information it is entanglement retained after weak measurement. The term therefore names a common structural problem—multiple claimants acting on a common substrate—but the operative mathematics is always determined by the domain-specific notion of admissible use, degradation, and protection.