Adaptive Joint Scheduler

• Adaptive Joint Scheduler is a mechanism that jointly optimizes resource allocation and control parameters based on dynamic signals such as input data rate and load balance.
• It integrates multiple architectures—from decision tree-based policy selection to weighted multi-objective controllers—across systems like wireless networks, Kubernetes clusters, and robotic inference.
• Empirical studies report up to 1.29× speedup, 45% lower EDP, and reduced latency by over 80%, demonstrating significant performance gains in diverse domains.

Searching arXiv for papers on “adaptive joint scheduler” and closely related scheduler frameworks to ground the article. Search query: adaptive joint scheduler arXiv schedulers workload runtime adaptive joint scheduling An adaptive joint scheduler is a scheduling mechanism that adapts to workload, system, or channel state while jointly controlling multiple decision variables that would otherwise be optimized separately. In the literature, this jointness appears in several forms: runtime selection between heterogeneous scheduling policies, simultaneous control of load and capacity, coupled tuning of resource-allocation objectives, or co-adaptation of execution depth and latency budget. The concept spans distributed wireless MAC protocols, heterogeneous SoCs, Kubernetes edge clusters, LLM serving stacks, robotic diffusion policies, federated split learning, and high-throughput satellites (Garcia-Saavedra et al., 2014, Goksoy et al., 2021, Tang et al., 4 Jun 2025, Sidik et al., 29 Jan 2026, Sun et al., 27 May 2026, Ding et al., 23 Jun 2026, Zhuoya et al., 2024).

1. Concept and scope

Two properties recur across the literature. First, adaptivity means that scheduling decisions depend on observed state such as input data rate, queue structure, CPU load, resource balance, latency, gradient norm, or mobility/load context. Second, jointness means that the scheduler does not optimize a single scalar knob in isolation: it jointly selects among schedulers, balances multiple objectives, or co-controls several resources or control dimensions. DAS, for example, jointly chooses between a fast LUT scheduler $F$ and a slow ETF scheduler $S$ at runtime; LRScheduler jointly combines a layer-sharing score with the default Kubernetes score via a dynamic weight $\omega$; SANTS jointly predicts a stopping hazard and a relative noise-progression ratio; and the FSL scheduler jointly selects compression mode and synchronisation interval $\rho$ from latency (Goksoy et al., 2021, Tang et al., 4 Jun 2025, Sun et al., 27 May 2026, Ding et al., 23 Jun 2026).

System	Joint decision variables	Adaptation signal
DAS	scheduler choice $F$ or $S$	input data rate; earliest availability time of the Arm big cluster
LRScheduler	layer-sharing weight and node score	cached layer size; CPU load; CPU/memory balance
SANTS	stopping hazard and noise-progression ratio	video-state representation $z_k$; noise level $\sigma_k$
FSL scheduler	compression mode and $\rho$	EMA-smoothed latency $\hat{l}_t$

This breadth matters because the term does not denote a single algorithmic family. In some systems the scheduler is a selector over expert policies, as in DAS and ASA; in others it is a weighted multi-objective controller, as in LRScheduler and OFDMA scheduling; in others it is a trajectory controller over iterative inference, as in SANTS; and in others it is a networked control loop over communication and synchronisation, as in federated split learning (Goksoy et al., 2021, Wang et al., 7 Nov 2025, Taie et al., 2016, Sun et al., 27 May 2026, Ding et al., 23 Jun 2026).

2. Historical development across domains

Early formulations arose in wireless networking. In distributed opportunistic scheduling, the scheduler jointly adapted each station’s access probability $S$0 and transmission threshold $S$1 to maximize proportional fairness, formalized as $S$2 (Garcia-Saavedra et al., 2014). In IEEE 802.11e HCCA, AMTXOP jointly adapted TXOP sizing and polling overhead through piggybacked next-frame sizes and multi-polling, thereby coordinating multiple uplink VBR streams (Al-Maqri et al., 2016). In LTE/OFDMA downlink scheduling, an adaptive ML-based framework combined clustering, SVM classification, and a GA-based multi-objective scheduler to adjust throughput-vs-GBR weights on a per-transmission basis (Taie et al., 2016).

The concept then broadened into heterogeneous computing. XiTAO introduced a Performance Trace Table (PTT) to jointly choose leader core and resource width for mixed-mode DAG tasks, thereby adapting to both static and dynamic heterogeneity (Chen et al., 2019). DAS extended this direction by jointly using two schedulers—a $S$3 ns LUT policy and a higher-overhead ETF heuristic—while moving the policy-selection computation off the critical path via preselection (Goksoy et al., 2021). In meta-learning, ATS replaced uniform task sampling with a neural scheduler driven by query loss and support–query gradient similarity, trained to improve generalization to unseen tasks (Yao et al., 2021).

Recent work shifted the idea into cloud, edge, and learning systems. MultiTASC++ continuously adjusted per-device forwarding thresholds in multi-device cascade inference using SLO satisfaction rate feedback (Nikolaidis et al., 2024). Flexible satellite scheduling jointly manipulated beam-user mapping, beam geometry, and bandwidth assignment to match non-uniform traffic (Zhuoya et al., 2024). LRScheduler combined layer-aware placement with resource-adaptive weighting inside Kubernetes (Tang et al., 4 Jun 2025). EWSJF jointly optimized partitioning, routing, and prioritization for mixed-workload LLM inference (Sidik et al., 29 Jan 2026). SANTS moved the scheduler into the diffusion trajectory itself, choosing both when to stop and how far to move along the noise axis (Sun et al., 27 May 2026). Federated split learning added a latency-driven joint controller over quantisation and synchronisation (Ding et al., 23 Jun 2026). O-RAN work used an A2C scheduler to jointly activate or deactivate pre-trained xApps to mitigate conflict without retraining them (Cinemre et al., 9 Apr 2025).

3. Mechanisms and mathematical structures

A central mechanism is scheduler selection. In DAS, the operating system chooses a policy

$S$4

where $S$5 is a fast LUT scheduler and $S$6 is ETF. The classifier does not run on the critical scheduling path; instead, it precomputes the next scheduler choice from a small set of counters, notably input data rate and earliest availability time of the Arm big cluster (Goksoy et al., 2021). ASA generalizes the same pattern as a “Mixture-of-Schedulers”: a learned router classifies workload patterns and then selects one scheduler from a portfolio through a mapping table and time-weighted probability voting (Wang et al., 7 Nov 2025).

A second mechanism is joint weighted scoring. LRScheduler defines the final Kubernetes node score as

$S$7

where $S$8 switches between $S$9 and $\omega$0 according to cached layer size, CPU utilisation, and CPU/memory balance thresholds (Tang et al., 4 Jun 2025). The LTE/OFDMA framework uses an analogous weighted objective,

$\omega$1

with $\omega$2 representing throughput maximization and $\omega$3 capturing GBR-user demand satisfaction (Taie et al., 2016). In satellite scheduling, the joint optimization is expressed through the quadratic unmet-demand objective

$\omega$4

with joint decisions over beam-user mapping, beam center and radius, and per-beam bandwidth (Zhuoya et al., 2024).

A third mechanism is control-theoretic joint adaptation. In distributed opportunistic scheduling, the optimal threshold satisfies

$\omega$5

while proportional fairness implies

$\omega$6

ADOS implements these through two coupled feedback loops: one for access probability and one for threshold (Garcia-Saavedra et al., 2014). AMTXOP follows a similar control style at the MAC layer by adapting TXOP according to the piggybacked next-frame size while simultaneously reducing polling overhead through a single multi-poll frame (Al-Maqri et al., 2016).

A fourth mechanism is trajectory scheduling in iterative inference. SANTS uses a cumulative hazard model: $\omega$7 and jointly predicts a relative noise-progression ratio $\omega$8 through a Beta distribution, with

$\omega$9

The scheduler therefore decides both whether to stop denoising and how aggressively to advance along the noise trajectory (Sun et al., 27 May 2026). Receiver-side scheduling for interactive delivery uses a related but different time-domain controller: ADC maintains an adaptive offset $\rho$0 and schedules release at

$\rho$1

with asymmetric updates that track an upper envelope of recovery delay (Luby, 21 Nov 2025).

A fifth mechanism is joint optimization of training dynamics. The SGD scheduler derives a critical batch size

$\rho$2

and then jointly updates batch size and learning rate in stages based on the observed decay of the full gradient norm (Umeda et al., 7 Aug 2025). ATS uses a neural task scheduler driven by query loss and support–query gradient similarity,

$\rho$3

to sample meta-training tasks that improve generalization under noise and limited task budgets (Yao et al., 2021).

4. Learning, control loops, and systems integration

The training and control architecture varies sharply by domain. DAS uses a two-run offline oracle process: first, an instrumented execution compares LUT and ETF decisions; second, a slow-only run determines whether pending states should be labeled $\rho$4 or $\rho$5. The resulting classifier is a depth-2 decision tree using only two features, with $\rho$6 accuracy and $\rho$7 ns runtime on Arm Cortex-A53 @ 1.2 GHz (Goksoy et al., 2021). ATS is also bi-level, but its scheduler is learned directly against meta-validation performance via REINFORCE over task-sampling probabilities (Yao et al., 2021).

In online service systems, strategic and tactical loops are often separated. EWSJF uses Refine-and-Prune to construct performance-homogeneous request groups, Dynamic Queue Routing to place incoming requests, Density-Weighted Scoring to choose the next queue, and Bayesian Meta-Optimization to tune partitioning and scoring parameters from live feedback (Sidik et al., 29 Jan 2026). ASA splits workload recognition from hardware-specific control: a universal XGBoost-centered classifier produces workload-class probabilities, time-weighted voting smooths them, and a machine-specific mapping table selects a sched_ext scheduler (Wang et al., 7 Nov 2025).

Other systems are almost purely feedback-controlled. The federated split-learning scheduler maintains a per-client EMA

$\rho$8

then maps smoothed latency to $\rho$9, $F$0, or $F$1 (Ding et al., 23 Jun 2026). MultiTASC++ updates each device threshold according to

$F$2

with a multiplicative accelerator when SLO satisfaction exceeds target (Nikolaidis et al., 2024). The O-RAN xApp scheduler is trained separately from the xApps themselves: the xApps are frozen after offline A2C training, and the scheduler learns over context variables such as average user speed and mean data arrival rate to activate A2C or baseline xApps (Cinemre et al., 9 Apr 2025).

Integration points are correspondingly diverse. DAS plugs into a Linux-based DS3 environment (Goksoy et al., 2021). LRScheduler is implemented as a Kubernetes custom scheduler via the Scheduling Framework’s Score extension point and uses framework.Handle, NodeInfo, the API server, etcd, Kubelet, and the Docker HTTP API (Tang et al., 4 Jun 2025). EWSJF sits upstream of the vLLM execution scheduler (Sidik et al., 29 Jan 2026). Receiver-side scheduling is an independent module in the BitRipple Tunnel receive path (Luby, 21 Nov 2025). The xApp conflict mitigator resides in the Near-RT RIC and interacts over E2 and A1 (Cinemre et al., 9 Apr 2025).

5. Representative empirical behavior

Empirical results consistently show that joint adaptation is most valuable when a single static policy faces regime changes. DAS reports, over 40 workloads, a $F$3 speedup and $F$4 lower EDP than the sophisticated scheduler at low data rates, and a $F$5 speedup with $F$6 lower EDP than the fast scheduler when workload complexity increases (Goksoy et al., 2021). LRScheduler reduces overall download time by about $F$7 compared with the default Kubernetes scheduler under varying bandwidth and keeps resource imbalance much closer to the default scheduler than a static layer-sharing policy (Tang et al., 4 Jun 2025).

In LLM serving, EWSJF improves end-to-end throughput by over $F$8 and reduces average Time-To-First-Token for short requests by up to $F$9 compared to FCFS; in reported workloads it reaches gains such as $S$0, $S$1, and $S$2 in token throughput across different scales (Sidik et al., 29 Jan 2026). ASA, in operating-system scheduling, outperforms EEVDF in $S$3 of tested scenarios and places its chosen scheduler among the top three in $S$4 of all scenarios (Wang et al., 7 Nov 2025).

In robot control, SANTS reaches $S$5 overall success on RoboTwin 2.0 and $S$6 average success across seven real-robot tasks while reducing latency by $S$7 and $S$8, respectively, relative to full video denoising (Sun et al., 27 May 2026). In federated split learning for rainfall prediction, AUPRC varies only slightly across configurations—$S$9 to $z_k$0 in simulation and within $z_k$1 on Raspberry Pi—while the selected endpoint, int8 with $z_k$2, reduces activation upload payload by $z_k$3, synchronisation traffic by $z_k$4, and runtime jitter from $z_k$5 s to $z_k$6 s (Ding et al., 23 Jun 2026).

Wireless and radio-resource systems show similarly strong regime dependence. AMTXOP yields up to $z_k$7 delay reduction versus standard HCCA and up to $z_k$8 versus ATXOP, while preserving nearly identical throughput (Al-Maqri et al., 2016). The O-RAN scheduler mitigates up to $z_k$9 transmission-rate loss observed under conflicting xApps in high-load, high-mobility settings, and the richer four-xApp scheduler achieves the highest total transmission rate among the tested deployment scenarios (Cinemre et al., 9 Apr 2025). In satellites, BW-SR dominates alternatives in NQU and NU under homogeneous, wide hot-spot, and real traffic, while roughly halving runtime relative to SR in the 64-beam real-traffic case (Zhuoya et al., 2024).

6. Trade-offs, misconceptions, and open directions

A recurring trade-off is decision quality versus control overhead. DAS exists precisely because ETF quality is not worth its overhead at low data rates, but LUT quality is insufficient under heavier workloads (Goksoy et al., 2021). LRScheduler balances layer reuse against load balancing; its static-weight counterpart saves more disk but produces worse balance (Tang et al., 4 Jun 2025). Receiver-side scheduling reduces jitter but necessarily adds bounded waiting; the clamp $\sigma_k$0, neutral band $\sigma_k$1, and guard window $\sigma_k$2 express that latency–smoothness compromise explicitly (Luby, 21 Nov 2025). SANTS shows the same structure in a different domain: full denoising is often not the best action condition, so extra inference can be both costly and counterproductive (Sun et al., 27 May 2026).

A common misconception is that “joint” necessarily means simultaneous online execution of all candidate schedulers. DAS explicitly does not run both schedulers in parallel online; it trains with both but invokes only one at runtime (Goksoy et al., 2021). ASA likewise routes among expert schedulers rather than fusing them into a single monolithic policy (Wang et al., 7 Nov 2025). By contrast, LRScheduler is joint in the sense of weighted score composition, and SANTS is joint in the sense of co-controlling stopping and progression along a trajectory (Tang et al., 4 Jun 2025, Sun et al., 27 May 2026). The literature therefore uses the term for several architectures: selection, composition, co-allocation, and coupled control.

Another misconception is that adaptivity implies heavy online learning. Some systems use RL or Bayesian optimization, but many do not. ADOS is control-theoretic (Garcia-Saavedra et al., 2014); DAS uses a depth-2 decision tree (Goksoy et al., 2021); the FSL scheduler is a rule-based latency controller (Ding et al., 23 Jun 2026); and LRScheduler uses thresholded switching between two preset weights (Tang et al., 4 Jun 2025). What unifies these systems is not a learning paradigm but the combination of state dependence and multi-variable coordination.

The forward-looking agenda is correspondingly broad. DAS suggests more than two schedulers and online adaptation beyond the original training set (Goksoy et al., 2021). LRScheduler points to reinforcement learning, cloud-edge collaborative layer sharing, and more sophisticated dynamic weighting (Tang et al., 4 Jun 2025). Receiver-side scheduling highlights joint design with transport, automatic parameter tuning, and integration into TCP, QUIC, WebRTC, RTP, or SRT (Luby, 21 Nov 2025). EWSJF motivates multi-tenant and multi-GPU extensions (Sidik et al., 29 Jan 2026). SANTS identifies adaptive action-denoising, broader perception scheduling, and joint optimization of backbone and scheduler as natural next steps (Sun et al., 27 May 2026). Across these directions, the unifying problem remains the same: how to expose multiple coupled control levers, observe enough state to make them useful, and keep adaptation overhead below the gains it is supposed to deliver.