Dynamic Allocation Scheme (DAS)

• Dynamic Allocation Scheme (DAS) is a design pattern that reallocates constrained resources online in response to runtime state changes such as workload, channel conditions, and fairness constraints.
• It is applied across heterogeneous domains including SoC scheduling, wireless communications, traffic engineering, and dynamic mechanism design, each with tailored allocation policies.
• Empirical studies reveal DAS trade-offs between decision overhead and quality, achieving notable improvements in speed, energy efficiency, and system throughput.

Dynamic Allocation Scheme (DAS) denotes a family of state-adaptive allocation procedures in which a constrained resource budget is redistributed online or stage by stage in response to observed workload, channel conditions, queue states, interference, or fairness requirements. In the literature represented here, the term does not identify a single canonical algorithm. Rather, it appears in heterogeneous systems scheduling, wireless resource control, traffic engineering, and dynamic mechanism design, while closely related formulations extend the same allocation logic to LLM context compression, static memory planning, and GPU object allocation (Goksoy et al., 2021, Yu et al., 2021, Song et al., 2020, Sroka et al., 2021, Fallah et al., 2024, Arquam et al., 2024, Chen et al., 17 Feb 2025).

1. Terminological scope and recurring problem structure

Across these works, DAS refers to a policy that maps a runtime state to an allocation of processors, power, spectrum, bandwidth, auction outcomes, or compressed representation capacity. The runtime state varies by domain: input data rate and processor availability in DSSoCs, channel and cache states in Fog-RAN, queue-state information and energy-state information in RAN slicing, QoS measurements in communication networks, residual fairness requirements in dynamic auctions, and perplexity- or attention-derived relevance scores in LLM compression (Goksoy et al., 2021, Yu et al., 2021, Song et al., 2020, Arquam et al., 2024, Fallah et al., 2024, Chen et al., 17 Feb 2025).

A common misconception is that DAS denotes a standardized protocol. The cited literature instead uses the same label for several mathematically distinct mechanisms. This suggests that DAS is best understood as a design pattern: a limited budget is reallocated dynamically under explicit constraints and a task-specific objective.

Formulation	Allocated resource	Objective or governing constraint
DSSoC DAS	scheduler choice between $F$ and $S$	makespan, energy, EDP (Goksoy et al., 2021)
Fog-RAN dynamic allocation	RRH transmit power $P_n(t)$	total network power cost with QoS (Yu et al., 2021)
RAN-slicing DAS	subchannels and power control	weighted-sum rate under delay constraints (Song et al., 2020)
Traffic-allocation DAS	per-class bandwidth $A_i(t)$	throughput with latency/jitter/loss bounds (Arquam et al., 2024)
Platooning DAS	TVWS frequency and power	max–min SINR with DTT protection (Sroka et al., 2021)
Dynamic fair allocation	round-by-round item allocation and payments	discounted revenue with fairness shares (Fallah et al., 2024)
DAST	soft tokens per context chunk	context-aware compression budget split (Chen et al., 17 Feb 2025)

Related systems work addresses the same allocation problem class without always using the exact DAS acronym. "Futureproof Static Memory Planning" studies dynamic storage allocation as offset assignment under lifetime overlap constraints (Lamprakos et al., 7 Apr 2025), and DynaSOAr develops a lock-free GPU allocator organized around block-level density and fragmentation control (Springer et al., 2018).

2. Mathematical forms of DAS

The mathematical structure of DAS depends on whether the underlying problem is combinatorial, convex, stochastic-control, or recursive-dynamic in nature.

In LLM compression, the closely related Dynamic Allocation of Soft Tokens (DAST) partitions a context into $N$ contiguous chunks and reallocates a global budget of $M$ soft tokens according to local perplexity and global attention. The local score is

$$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$

the normalized perplexity is $\tilde P_i = P_i/\sum_{k=1}^N P_k$, and the combined score is

$S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$

After a softmax,

$$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$

so that $S$0. The only hyperparameter is $S$1, and sensitivity analysis shows robust performance across $S$2, with $S$3 fixed in practice (Chen et al., 17 Feb 2025).

In heterogeneous SoCs, DAS is a policy-selection problem rather than a direct resource-partition problem. The system switches per task between a fast lookup-table scheduler $S$4 and a slow Earliest-Task-First scheduler $S$5. The slow scheduler has complexity $S$6 per decision, and its latency is modeled as $S$7, where $S$8 is the number of ready tasks. The switching decision is learned offline and implemented online by a depth-2 decision tree over two features: the input data rate $S$9 and the earliest availability time of the “big” CPU cluster $P_n(t)$0 (Goksoy et al., 2021).

In 5G and beyond RAN slicing, DAS is formulated as an infinite-horizon average-reward constrained Markov decision process. The global state is

$P_n(t)$1

and the optimization maximizes the long-run weighted sum rate under per-slice average delay constraints. Because exact value iteration suffers from the curse of dimensionality, the global Q-factor is approximated by a sum of per-slice Q-factors,

$P_n(t)$2

with online stochastic updates for both Q-factors and Lagrange multipliers (Song et al., 2020).

In Fog-RAN, the dynamic allocation problem is posed as power minimization with QoS and delivery constraints. The backhaul term uses an $P_n(t)$3-norm surrogate,

$P_n(t)$4

followed by majorization–minimization and a reduced convex problem. In the special case $P_n(t)$5, the delay constraint is tight and yields closed-form “water-filling-type” policies; in the general case $P_n(t)$6, the remaining problem is solved through KKT conditions or time discretization (Yu et al., 2021).

In platooning, DAS is a max–min interference-management problem. For platoon $P_n(t)$7, the decision variables are a TVWS center frequency $P_n(t)$8 and transmit powers $P_n(t)$9, and the objective is

$A_i(t)$0

subject to power bounds and DTT protection constraints derived from a Radio Environment Map (Sroka et al., 2021).

In dynamic mechanism design, DAS is expressed through residual fairness states $A_i(t)$1, seller value functions $A_i(t)$2, and buyer continuation utilities $A_i(t)$3. The seller solves a Bellman recursion over feasible allocations, and the optimal round-$A_i(t)$4 rule compares shifted virtual values, where the shift depends on future-utility differences for both buyers and seller (Fallah et al., 2024).

3. DAS in computing systems and allocation infrastructure

In domain-specific SoCs, the central issue is the mismatch between nanosecond-scale task runtimes and scheduler overhead. DAS addresses this by combining a very low-overhead LUT-based fast scheduler with a more sophisticated ETF scheduler. On Arm Cortex-A53 at $A_i(t)$5 GHz, the fast scheduler overhead is $A_i(t)$6 cycles $A_i(t)$7 ns with energy $A_i(t)$8 nJ, whereas the slow scheduler reaches measured peak latency $A_i(t)$9 ns and energy $N$0 nJ under heavy load. The preselection classifier is a depth-2 decision tree with $N$1 accuracy, $N$2 KB code, and $N$3 ns runtime, while the critical path has zero added latency because the required features are prefetched (Goksoy et al., 2021).

The empirical results show that, across 40 workloads, DAS achieves on average $N$4 speedup and $N$5 lower EDP compared to the sophisticated scheduler at low data rates, and $N$6 speedup and $N$7 lower EDP than the fast scheduler when workload complexity increases. Scheduler use shifts from $N$8 $N$9 at the lowest rates to $M$0 $M$1 at the highest rates, with average scheduling overhead of $M$2 ns $M$3 nJ$M$4 at low/medium loads and $M$5 ns $M$6 nJ$M$7 at heavy loads (Goksoy et al., 2021).

Related allocation infrastructure highlights a neighboring systems interpretation of “dynamic allocation.” In static memory planning, the dynamic storage allocation problem assigns offsets $M$8 to buffers with sizes $M$9 and lifetimes $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$0 so as to minimize

$$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$1

under non-overlap constraints induced by lifetime intersections. The lower bound is the max load

$$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$2

and fragmentation is $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$3. The idealloc implementation targets low fragmentation, high throughput, and scalability to millions of buffers, with each boxing/unboxing iteration running in $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$4 time in typical inputs and space $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$5, where $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$6 is the number of overlaps (Lamprakos et al., 7 Apr 2025).

GPU allocation research reaches similar concerns from a different angle. DynaSOAr is a CUDA-only, lock-free object allocator for Single-Method Multiple-Objects applications. It organizes the heap into fixed-size blocks, each storing many objects of a single C++ type in Structure-of-Arrays form, and uses lock-free hierarchical bitmaps for free, allocated, and active block tracking. Its benchmarks report application-code speedups of up to $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$7 over state-of-the-art allocators and allow up to $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$8 larger problem sizes with the same amount of memory (Springer et al., 2018). This suggests that, in systems work, DAS is closely linked to low-overhead decision making, fragmentation control, and state-aware reuse of existing capacity.

4. DAS in wireless communications and networking

In intelligent Fog-RAN for high-speed railway communication, the dynamic allocation variable is the RRH power process $$P_i \;=\; -\sum_{l=1}^{L} q(x_{i,l}) \,\log p\bigl(x_{i,l}\mid x_{i,<l}\bigr),$$9. The total cost is the sum of radio-access transmission cost and backhaul cost, subject to instantaneous delay, total file-delivery, and per-RRH average-power constraints. Caching enters through $\tilde P_i = P_i/\sum_{k=1}^N P_k$0: if $\tilde P_i = P_i/\sum_{k=1}^N P_k$1, the backhaul cost for file $\tilde P_i = P_i/\sum_{k=1}^N P_k$2 at RRH $\tilde P_i = P_i/\sum_{k=1}^N P_k$3 vanishes. The study analyzes the trade-off among total network cost, delay, and delivery content size, and simulation shows that dynamic power allocation yields up to $\tilde P_i = P_i/\sum_{k=1}^N P_k$4–$\tilde P_i = P_i/\sum_{k=1}^N P_k$5 lower total cost than invariant allocation, especially under RndC or NonC; under PopC the gain is $\tilde P_i = P_i/\sum_{k=1}^N P_k$6–$\tilde P_i = P_i/\sum_{k=1}^N P_k$7. The MM-based Algorithm 1 converges in $\tilde P_i = P_i/\sum_{k=1}^N P_k$8 iterations, PopC outperforms RndC by $\tilde P_i = P_i/\sum_{k=1}^N P_k$9–$S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$0 in cost, and both outperform NonC by $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$1–$S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$2 (Yu et al., 2021).

In uplink RAN slicing, the dynamic allocation variables are subchannel indicators $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$3 and fractional power controls $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$4. The objective is the infinite-horizon weighted-sum rate, with delay approximated through Little’s law with dropping, $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$5. The proposed DAS decomposes the Q-factor per slice and updates the Q-factors and Lagrange multipliers by two-timescale stochastic approximation. Its complexity scales linearly in $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$6, $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$7, $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$8, and the Q/E state sizes rather than exponentially. In simulation with $S_i \;=\;\alpha\,A_i\;-\;(1-\alpha)\,\tilde P_i.$9 slices, $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$0 UEs per slice, and $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$1 subchannels, DAS outperforms the random equal slicing and heuristic CSI+QSI baselines, yields $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$2–$$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$3 gain over the heuristic baseline under tight energy budgets, maintains QoS with lower dropping probability, and converges within $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$4 iterations of Q-learning and LM updates (Song et al., 2020).

In communication-network throughput optimization, the Dynamic Traffic Allocation Scheme operates on four traffic classes—VoIP, video streaming, web browsing, and file download. At each epoch it measures latency, jitter, packet loss, bandwidth demand, total resource demand, priority score, QoS index, and current load, then forms normalized weights $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$5, $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$6, $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$7, $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$8, $$r_i \;=\;\frac{\exp(S_i)}{\sum_{k=1}^N \exp(S_k)},\qquad s_i \;=\; M r_i,$$9, and $S$00. The raw allocation fraction is

$S$01

followed by normalization and allocation $S$02. In a 50-node mesh with link capacities drawn from $S$03 Mb/s, the average throughput increases from $S$04 Mb/s for static allocation to $S$05 Mb/s for dynamic DAS, an improvement of $S$06. The reported throughput gain over static load balancing is $S$07–$S$08, VoIP and video constraints are met $S$09 of the time, and the scheme self-stabilizes within $S$10–$S$11 s of large traffic shifts (Arquam et al., 2024).

In autonomous platooning, DAS jointly allocates TVWS frequencies and powers using REM information. Every $S$12 s, a central spectrum manager evaluates candidate frequency tuples, computes ACIR matrices, derives the maximum allowable per-vehicle power under DTT protection, estimates PU-to-vehicle and vehicle-to-vehicle interference, and selects the tuple that maximizes the minimum platoon SINR. In simulation, leader CACC reception exceeds $S$13 under DAS in both traffic densities; without power control, up to $S$14 of empirical SIR measurements fall below $S$15, whereas with DAS power control nearly $S$16 remain above threshold, apart from a few outliers due to shadowing. Channel-switching overhead is $S$17–$S$18 changes per run without power control versus $S$19–$S$20 with power control (Sroka et al., 2021).

5. Fairness-constrained DAS in mechanism design

In dynamic mechanism design, DAS refers to a sequence of allocation and payment rules that maximize discounted seller revenue while guaranteeing a minimum discounted average allocation share for each of two buyer groups. The fairness constraint is imposed ex ante at each round $S$21, and the state variable is the residual minimum allocation requirement $S$22 after the history of earlier rounds (Fallah et al., 2024).

The static case $S$23 already exhibits the core structure. Under regularity, the fair allocation rule modifies Myerson’s virtual-value mechanism by two forms of subsidization: a universal subsidy $S$24 that lowers the effective reserve for group $S$25, and a group-specific subsidy $S$26 that shifts the virtual-value comparison boundary. If $S$27 and $S$28, then the good is allocated to group 1 when

$S$29

and to group 2 symmetrically, with $S$30. The multipliers are chosen so that $S$31 for $S$32, subject to $S$33 (Fallah et al., 2024).

For $S$34, the seller’s recursion is

$S$35

with analogous continuation utilities $S$36 for buyers. The optimal per-round rule remains a shifted virtual-value comparison, but the shift now depends on future-utility differences $S$37 and $S$38. Payments include two parts: a “participation bonus,” $S$39, paid only when the group wins, and an “entry fee,” $S$40, charged to all participants. The paper states that the seller commits to a participation bonus to incentivize truth-telling and charges an entry fee for every round (Fallah et al., 2024).

Because exact recursion is exponential in $S$41, the paper gives two approximation schemes. The early-stopping approximation fixes $S$42, solves the first $S$43 rounds exactly, and uses the unconstrained second-price auction afterward. It satisfies each fairness share approximately at level $S$44 and has complexity $S$45 oracle calls. The bucketed-discount approximation targets $S$46, groups rounds into buckets so that $S$47, and yields at least a factor $S$48 approximation to both revenue and fairness shares, with overall polynomial complexity in $S$49 (Fallah et al., 2024).

6. Empirical patterns, trade-offs, and common misconceptions

A first recurring trade-off is overhead versus decision quality. In DSSoCs, the fast LUT scheduler has $S$50 ns latency but can be suboptimal under heavy load, whereas ETF can make better decisions at the price of $S$51 ns peak latency; DAS exists precisely to move between these regimes (Goksoy et al., 2021). In network throughput optimization, DAS explicitly avoids a heavy LP or convex optimizer in favor of a one-shot closed-form rule (Arquam et al., 2024). In platooning, worst-case vehicle-to-vehicle interference is approximated by the nearest interferer to reduce complexity with negligible performance loss (Sroka et al., 2021).

A second trade-off is exactness versus tractability. The 5G slicing problem begins from an average-reward Bellman equation but then uses linear value-function approximation and per-slice Q-factor decomposition to escape exponential complexity (Song et al., 2020). The Fog-RAN problem uses $S$52-norm approximation, MM, and convex reduction rather than direct combinatorial treatment of backhaul activation (Yu et al., 2021). The fair-allocation mechanism relies on approximation schemes because the exact state space can reach $S$53 residual states (Fallah et al., 2024).

A third trade-off is adaptivity versus scope of validation. DAST is evaluated on LongBench document and example compression tasks with LLama-2-7B and Qwen-2-7B backbones, where it achieves averages of $S$54 and $S$55, compared with Beacon at $S$56 and $S$57, and on the MSC benchmark suffers only a $S$58 relative drop at $S$59 compression compared to Beacon’s $S$60 drop. At the same time, the paper states that evaluation has been conducted only at the $S$61B-parameter scale, that scaling studies on $S$62B+ models are open, and that the impact on hallucination rates or catastrophic forgetting remains unexplored (Chen et al., 17 Feb 2025).

A fourth trade-off concerns model coverage. The traffic-allocation DAS assumes the per-class path is fixed and does not model link-failure rerouting; extension to DAS+SDN and multi-tenant slices is left for future work (Arquam et al., 2024). idealloc notes that bootstrapping and $S$63-tuning incur a small constant factor, even though early stopping keeps worst-case time bounded (Lamprakos et al., 7 Apr 2025). DynaSOAr requires all types to be known at compile time, does not support virtual functions, and can be less advantageous when an application touches only one or two fields of many objects (Springer et al., 2018).

Taken together, these results suggest a unifying interpretation of DAS: dynamic allocation is most effective when the chosen state variables are strong proxies for marginal utility or constraint tightness. In the cited literature, those proxies include perplexity and attention in LLM compression, input data rate and $S$64 in heterogeneous scheduling, queue and energy states in network slicing, cache state and channel evolution in Fog-RAN, measured QoS indices in traffic engineering, and residual fairness requirements in repeated auctions (Chen et al., 17 Feb 2025, Goksoy et al., 2021, Song et al., 2020, Yu et al., 2021, Arquam et al., 2024, Fallah et al., 2024).