SEM-O-RAN: Semantic and Flexible O-RAN Slicing for NextG Edge-Assisted Mobile Systems

I. INTRODUCTION

• 32 Gb/s of traffic would result from 100 connected vehicles sending frames from four cameras every 10 ms.The estimate assumes 100 KB images and real-time self-navigation.Source: §I · p. 1
• RAN slicing allocates networking and computational resources to virtual operators, and O-RAN supports fine-grained, real-time control of those resources.Source: §I · p. 1
• Existing slicing either does not support O-RAN or defines edge tasks monolithically; SEM-O-RAN is proposed as an O-RAN slicing framework for edge-assisted mobile applications.Source: §I · p. 1
• Combining semantic compression and flexible resource allocation improves allocated tasks by up to 169% over.Flexibility alone improves allocated tasks by up to 31% over.Source: §I · p. 1, Figure 1 · p. 2
• The formulation must capture nonlinear links among slicing, compression, class accuracy, and latency while handling the added complexity of flexible resource allocation.The authors state that no other work, to their knowledge, holistically tackles both aspects.Source: §I · pp. 1–2

Summary of Novel Contributions

• The SF-ESP maximizes admitted DL tasks under latency and accuracy guarantees while avoiding resource over-provisioning; it is NP-hard and solved with a greedy algorithm.The formulation captures nonlinear relations among slicing, compression, latency, and accuracy, and balances different resource types.Source: Summary of Novel Contributions · p. 2
• SEM-O-RAN improves allocated tasks by up to 169% and by 18% on average over Sl-EDGE.Evaluation includes numerical analysis and a Colosseum prototype spanning YOLOX object detection and BiSeNet V2 image segmentation.Source: Summary of Novel Contributions · p. 2

II. TWO KEY CONCEPTS IN SEM-O-RAN

• Different application classes tolerate different image-compression levels, leaving room for compression while maintaining acceptable average inference accuracy.Source: §II · p. 2
• Different RBG/GPU allocations can meet the same latency target, allowing resource choices that leave capacity for additional tasks.The example contrasts allocations of (6, 3) and (10, 2) for a 0.4-second requirement.Source: §II · p. 2, Figure 2 · p. 2

A. Background Notions on O-RAN

• The O-RAN separates RU, CU, and DU functions, with Non-real-time and Near-real-time RICs handling orchestration and fine-grained controls such as slicing.Third-party rApps and xApps can run in the respective RICs.Source: A. Background Notions on O-RAN · pp. 2–3

B. SEM-O-RAN: Functional Blocks and Interfaces

• SDLA runs as an rApp and SESM as an xApp; the former supplies latency and accuracy functions, while the latter selects tasks, compression, and radio and compute allocations.Task descriptions specify the service, model, and target classes; task requirements specify latency, accuracy, device count, and job rate.Source: B. SEM-O-RAN: Functional Blocks and… · p. 3

C. A Walk-through of SEM-O-RAN.

• The workflow sends task descriptions to SDLA, shares performance functions with SESM, and uses task requirements plus current radio and edge status to produce slices.Changed requests trigger recomputation, and previously running tasks may then be removed; compression settings are returned through VNOs because RIC apps cannot directly communicate with device applications under the described specifications.Source: C. A Walk-through of SEM-O-RAN · pp. 3–4, B. SEM-O-RAN: Functional Blocks and… · p. 3

A. System Model

• An application class specifies a DL service, target object classes, minimum expected accuracy, and maximum delay for the devices running it.A task is a periodic edge execution of a DL service over a device's inference-data stream, with results returned to that device.Source: A. System Model · p. 4
• Compression scaling lowers stream bitrate and can reduce bandwidth needs and latency, while lower scaling sacrifices data quality and inference accuracy.The model assumes the original stream size depends on application class and compression latency is constant across scaling factors.Source: A. System Model · p. 4
• Each task receives networking and compute resources subject to capacity limits, with resource-specific costs used to represent their differing costs.Expected accuracy depends on compression, while latency depends on compression and resource allocation; both are treated as input data-driven functions.Source: A. System Model · p. 4

B. SF-ESP Problem Formulation

• SF-ESP uses a binary admission variable, a resource allocation matrix, and a compression scaling vector that affects accuracy and latency.The scaling factor sets data quality; accuracy depends on it, while latency depends on both scaling and allocated resources.Source: B. SF-ESP Problem Formulation · p. 5
• The objective maximizes admitted tasks according to priority while minimizing allocated resources, and the formulation is a mixed-integer nonlinear problem.The model includes integer and continuous variables.Source: B. SF-ESP Problem Formulation · p. 5
• SF-ESP is NP-hard because a 0/1 multidimensional knapsack instance can be reduced to it in polynomial time.The reduction fixes compression and allocations and makes performance constraints always satisfied, leaving task selection with known values and resource weights.Source: B. SF-ESP Problem Formulation · p. 5
• The authors characterize SF-ESP as harder than 0/1 multidimensional knapsack and note that existing results do not provide a non-trivial approximation ratio for their greedy heuristic.Their comparison cites strongly correlated knapsack with variable weights and nonlinear constraints as part of the complexity.Source: B. SF-ESP Problem Formulation · p. 5

C. Greedy Algorithm for the SF-ESP

• The heuristic derives from primal effective gradients for 0/1 multidimensional knapsack and ranks tasks by relative value under resource-use penalties.Under a monotonic latency assumption, it selects the minimum compression scaling factor meeting the accuracy requirement.Source: C. Greedy Algorithm for the SF-ESP · p. 5
• Because several resource combinations can meet a task's latency and accuracy requirements, the algorithm seeks allocations that maximize admitted tasks rather than simply minimizing each task's resources.It balances consumption according to availability, for example substituting compute resources when radio resources are scarce.Source: C. Greedy Algorithm for the SF-ESP · pp. 5–6
• Because accuracy and latency are treated as generic functions, the algorithm enumerates the resource-allocation solution space to solve its optimization equations.The paper notes that efficient alternatives would require properties such as monotonicity or convexity.Source: C. Greedy Algorithm for the SF-ESP · p. 6
• The algorithm prunes tasks whose accuracy targets cannot be met, explores feasible resource allocations, and repeatedly admits the task with the highest maximum primal gradient.It initializes task admissions and allocations to zero and compression factors to one, then returns the admission, allocation, and scaling vectors.Source: C. Greedy Algorithm for the SF-ESP · p. 6

• SEM-O-RAN is evaluated through numerical analysis and practical experiments on the Colosseum network emulator.Source: §V · p. 6

A. Experimental Setup

• The study evaluates object detection with YOLOX on COCO and instance segmentation with BiSeNet v2 on Cityscapes, defining 10 object-detection tasks.Source: A. Experimental Setup · p. 6
• The five baselines include Sl-EDGE, semantics-only MinRes-SEM, flexible non-semantic FlexRes-N-SEM, aggressive HighComp, and static HighRes.HighComp compresses tasks to 10% of original size; HighRes allocates 20% of total resources to tasks.Source: A. Experimental Setup · pp. 6–7
• The Colosseum prototype extends SCOPE with uplink slicing, which the original framework did not support.Source: A. Experimental Setup · p. 7
• One of 20 SRNs hosts the O-RAN components and solver; 18 emulate a VNO requesting three object-detection slices, with 15 RBGs available for slicing.Up to 20 Tesla K40m GPUs are available for DNN execution.Source: A. Experimental Setup · p. 7, Figure 5 · p. 7

B. Numerical Results

• With more resource types, SEM-O-RAN allocates up to 169% more tasks than Sl-EDGE and 18,5% more on average, outperforming all schemes in the considered scenarios.This advantage is especially evident as task counts increase and requirements become more stringent.Source: B. Numerical Results · p. 7
• With 50 generated tasks under medium accuracy and high latency requirements, SEM-O-RAN allocates 20% more tasks than Sl-EDGE and FlexRes-N-SEM, and 402% more than HighRes.Source: B. Numerical Results · p. 7
• For high mAP/mIoU requirements, only SEM-O-RAN and MinRes-SEM allocate tasks that meet constraints; Sl-EDGE cannot meet the stated 0.55/0.70 thresholds.The evaluation varies resource counts (2 or 4), accuracy thresholds, and latency thresholds of 0.2 or 0.7 seconds.Source: B. Numerical Results · p. 7

C. Experimental Results

• The Colosseum experiments track end-to-end latency against task thresholds as frame-generation rates change every 25 seconds, alongside RBG and GPU allocations.The figure captions state that chosen compression rates are also reported.Source: C. Experimental Results · p. 7, Figure 7 · p. 8, Figure 6 · p. 8
• In the third and fourth periods, SEM-O-RAN allocates all three tasks; MinRes-SEM misses Animals in the first period because its 8-RBG request would push use to 16 RBGs.SEM-O-RAN instead requests 6 RBGs and 5 GPUs in that period.Source: C. Experimental Results · pp. 7, 9
• FlexRes-N-SEM’s non-semantic compression misses Bags’ mAP requirement and never admits Animals, while SEM-O-RAN and MinRes-SEM use class-specific compression to meet requirements.For Flat, FlexRes-N-SEM uses 18% compression rather than the 8% selected by the semantic approaches, consuming more RBGs in the latest period.Source: C. Experimental Results · p. 9

VI. RELATED WORK

• Related work situates MEC slicing within RAN-slicing research,, and latency-driven MEC,, and points readers to surveys,.Source: §VI · p. 9
• Prior MEC approaches address admission control or heterogeneous offloading, but cited frameworks are not O-RAN-compatible; DIRECT also omits co-located MEC and networking resources.Van Huynh et al. focus on admission control; Ndikumana et al. consider heterogeneous MEC resources, and Liu et al. propose DIRECT.Source: §VI · p. 9
• O-RAN studies select scheduling policies or DL models and execution locations, whereas this work focuses on slicing MEC resources under accuracy constraints.Flexible VNF allocation, does not consider application semantics and is not readily applicable to edge offloading.Source: §VI · p. 9
• The closest work, Sl-EDGE, instantiates heterogeneous edge slices but does not account for DL semantics.Source: §VI · p. 9

VII. CONCLUDING REMARKS

• SEM-O-RAN improves allocated tasks by up to 169% over the state-of-the-art work while meeting accuracy and delay constraints.The conclusion describes numerical and Colosseum experiments and proposes a greedy approximation algorithm for the NP-hard SF-ESP.Source: §VII · p. 9