{"id":"60970207-64a8-4117-ab23-93eb05491184","arxiv_id":"2506.19760","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CORMO-RAN jointly optimizes which near-RT RIC nodes stay on and which migration technique to use for stateful xApps, cutting modeled cluster energy by up to 64% versus the OpenShift default scheduler.","lead":"An O-RAN orchestrator called CORMO-RAN decides when to turn off compute nodes and move AI apps, called xApps, between nodes to save energy. It compares two migration techniques on a real OpenShift testbed and reports up to 64% energy savings in a model-based evaluation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 64% energy saving rests on allowing SM migration downtime up to 300 s, contradicting the paper's own 1 s near-RT RIC deadline; under the stated deadline SM is infeasible and the headline saving is not supported.","rationale":"I focused on the central claim that CORMO-RAN 'minimizes energy consumption of the RIC cluster, yielding up to 64% energy saving' while 'ensuring xApp availability during migration.' For that claim to hold, the optimized solution must respect the near-RT RIC timing constraints that the paper itself uses as a yardstick. The most load-bearing weak point is the mismatch between the experimentally established SM downtime and the deadline used in the optimization. The paper's Sec. V-B Finding 2 states that even the best SM variant (SM-MD) has downtime around 5 s and 'is incompatible with the near-RT RIC control loop deadline of 1 s.' Yet Sec. VII sets T_max_Dk = 300 s in constraint (20), an 'arbitrary' threshold, and all reported SM energy gains and feasibility regions are generated with that threshold. This means the up-to-64% saving is not demonstrated for any configuration that satisfies the paper's own availability requirement. The reader's weakest assumption about extrapolating the fitted linear model to hundreds of xApps is a valid secondary concern, but it is not the primary problem: even granting the model, the optimization is over the wrong feasible set. The NP-hardness proof in Theorem 1 is also logically invalid (reducing SAL to general MIQP shows only that SAL is no harder than MIQP), but it does not affect the numerical results. The paper has genuine strengths: real testbed, reproducible measurements, and clear findings for SDL. The right remedy is to re-run the optimization with the correct deadline, or to clearly restrict the SM claim to scenarios where a 300 s downtime is acceptable and to then show that 'lossless' and 'availability' are not violated. I therefore keep the reader's conditional verdict; the condition should be made explicit and testable.","tokens_in":24215,"tokens_out":7520,"duration_ms":79170,"concrete_test":"Re-run the SAL optimization in Sec. VII with T_max_Dk = 1 s (the near-RT RIC deadline used in Sec. V-B) instead of 300 s for both SM-MR and SM-MD, and recompute the energy gains in Fig. 14 and the feasibility regions in Fig. 16. If no SM solution is feasible for any configuration, the 64% claim is unsupported under the paper's own timing requirement; if some SM solutions remain feasible, report the maximum gain under the 1 s constraint and compare it with the SDL-only gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Finding 2 (Sec. V-B) states that SM migration downtime is incompatible with the 1 s near-RT RIC control-loop deadline, with SM-MD downtime on the order of 5 s even for the smallest state size. Nevertheless, the SAL evaluation in Sec. VII sets T_max_Dk = 300 s in constraint (20), calling it an 'arbitrary maximum stateful migration downtime that can be tolerated.' The energy gains in Fig. 14 for SM-MR and SM-MD (up to 64%) and the SM feasibility regions in Fig. 16 are all computed under this relaxed 300 s threshold. If the actual near-RT deadline of 1 s is enforced, constraint (20) makes SM infeasible for every configuration reported, so the headline 64% saving cannot be achieved by SM while 'ensuring xApp availability during migration' as claimed in the abstract. The paper's own experiments therefore contradict the regime in which the headline result is obtained. This is an internal inconsistency, not a question of extrapolation: even if every linear fit in Sec. VI is exact, the optimization is over a feasible set that permits availability violations the paper itself documents.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CORMO-RAN, an rApp-based orchestrator that jointly decides which compute nodes of a near-RT RIC cluster to activate and how to migrate stateful xApps among them, using three strategies: stateful migration minimizing resources (SM-MR), stateful migration minimizing downtime (SM-MD), and the O-RAN Shared Data Layer (SDL). The authors build a testbed on Red Hat OpenShift with an open-source near-RT RIC, a DRL-based xApp, an E2 emulator, Prometheus/Kepler monitoring, and etcd as the SDL backend. They measure migration downtime/duration, resource usage, energy consumption, and etcd maintenance disruptions, then fit linear models to these measurements. These fits are used in a mixed-integer optimization problem (SAL) solved with Gurobi. The paper reports up to 64% energy saving compared with an OpenShift-style baseline and provides feasibility regions for the three migration strategies.","tokens_in":24589,"tokens_out":10278,"duration_ms":114636,"significance":"The experimental characterization of SM versus SDL migration in a realistic O-RAN cluster is a useful and timely contribution, and the paper is careful in several respects: measurements are averaged over 50 repetitions with 95% confidence intervals, the testbed uses commercial components and open-source tools, and the feasibility analysis explicitly highlights the scalability limits of SDL. If the model-based energy results are properly qualified, the framework could be a valuable aid for operators choosing migration strategies. However, the headline quantitative claim is not an end-to-end measured saving; it is the optimum of an objective built from the same fitted data, and it is obtained under a migration-downtime threshold that conflicts with the paper's own near-RT deadline finding. These issues must be resolved before the central claim can be accepted.","major_comments":[{"comment":"Finding 2 states that SM-MD downtime is on the order of 5 s even for the smallest state size and is 'incompatible with the near-RT RIC control loop deadline of 1 s.' Nevertheless, Section VII sets T_max_Dk = 300 s in constraint (20), calling it an 'arbitrary maximum stateful migration downtime that can be tolerated,' and Figs. 14 and 16 compute SM energy gains and feasibility under this relaxed threshold. If the paper's own 1 s near-RT deadline is enforced, constraint (20) makes SM infeasible in every reported configuration, so the headline up-to-64% saving is obtained in a regime that violates the availability requirement asserted in the abstract. The authors must either justify a 300 s downtime as an explicit availability trade-off (and revise the abstract accordingly) or re-evaluate the optimization with a per-xApp downtime constraint of 1 s, in which case SM cannot contribute to the reported energy saving.","section":"Sec. V-B (Finding 2) and Sec. VII (constraint (20))"},{"comment":"The 'up to 64% energy saving' is not an end-to-end measured result. The objective Es in Eq. (11) and the KPI constraints in Eqs. (1)-(4) and (20)-(22) use slopes and intercepts fitted from the experimental measurements in Figs. 6-12, and Fig. 14 is obtained by minimizing that same fitted objective with Gurobi. The reported saving therefore reduces to the fitted linear coefficients rather than providing independent confirmation of the system's performance. Moreover, the linear fits are measured for at most 15 migrated xApps (Fig. 6) and 50 xApps (Fig. 7), while Fig. 14 evaluates up to 400 xApps and Figs. 15-16 up to 150 xApps; the linear extrapolation to these scales is an unsupported assumption. Please reframe the 64% figure as a scenario analysis under the fitted model, and either validate the linearity at the evaluated scales or restrict the claims to the measured range.","section":"Sec. VI-B (Eq. (11)) and Sec. VII (Fig. 14)"},{"comment":"The SDL energy model in Eq. (10) uses slopes delta_SDL_E,k from Table II that are negative for all xApp classes (e.g., -0.18 W for class A). For large N_k this makes the term (delta_SDL_E,k * N_k + b_SDL_E,k) negative, so the objective rewards hosting additional SDL xApps with negative energy. For example, with 400 class-A xApps the per-class term becomes -0.18*400 + 32.35 = -39.65, and Eq. (10) then contributes a negative energy cost. This is an artifact of the etcd saturation behavior described in Sec. V-C and should not enter the optimization as a physical cost; the authors should use nonnegative fitted models or explicitly restrict the load range over which the linear fit is valid.","section":"Sec. VI-A (Eq. (10)) and Table II"},{"comment":"The proof of Theorem 1 is logically inverted. The paper argues that because SAL is an MIQP and the general decision version of MIQP is NP-complete, SAL is NP-hard by reduction. A reduction from SAL to the general MIQP only shows that SAL is no harder than the general MIQP; NP-hardness requires a polynomial-time reduction from a known NP-hard problem to SAL. Please replace this with a valid reduction (e.g., from bin packing or partition) or remove the theorem.","section":"Sec. VI-C (Theorem 1)"}],"minor_comments":[{"comment":"The bullets after Fig. 14 cite constraint (20) for resource usage and constraint (19) for the maximum downtime; these should refer to constraint (18) and constraint (20), respectively.","section":"Sec. VII (Fig. 14 discussion)"},{"comment":"The text says constraint (20) limits 'the downtime due to xApps being migrated to s,' but Eq. (1) defines T^tau_Dk,s in terms of the xApps leaving source s. The wording should match the equation.","section":"Sec. VI-B (constraint (20))"},{"comment":"The paper states that all results have 95% confidence intervals from 50 repetitions, but most figures do not show error bars or confidence bands; please indicate where the intervals were omitted and why.","section":"Sec. V (global)"},{"comment":"The term 'lossless migration' is used for both SM (state-preserving but with service downtime) and SDL (zero-downtime by design); define it explicitly at first use to avoid conflating state preservation with service availability.","section":"Abstract and Sec. III"},{"comment":"The baseline 'OpenShift default scheduler' is described only verbally as resource-based load balancing; specify whether the energy gain in Fig. 14 is computed against a modeled baseline or against a baseline measured on the testbed.","section":"Sec. VII (baseline definition)"}],"recommendation":"major_revision","confidential_remarks":"The experimental work is solid and likely publishable after major revision. The main risks are overclaiming the energy saving as a measured result, the 300 s downtime assumption contradicting the near-RT deadline, the unphysical negative SDL energy terms, and the invalid NP-hardness proof. None of these appear to be driven by bad faith; they are correctable presentation and modeling issues, but they affect the central quantitative claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has real experimental content worth engaging with, but the headline 64% energy saving is not supported. The stress-test note is correct. Finding 2 states SM migration downtime is incompatible with the 1 s near-RT RIC deadline, yet Sec. VII relaxes T_max_Dk to 300 s and reports up to 64% savings for SM. Under the actual 1 s deadline, SM is infeasible in every configuration they tested, and the 64% claim collapses.\n\nWhat is actually new and good: the first experimental characterization I know of comparing stateful migration (SM-MR/SM-MD) against the O-RAN Shared Data Layer on a real OpenShift cluster, including etcd compaction/defragmentation downtime and its effect on xApp service disruption. The measurement methodology is solid: 50 repetitions, 95% confidence intervals, a real DRL xApp plus an emulator for scale. The linear fits in Eqs. (1)-(4) are clearly derived from those measurements, and the SAL formulation is a reasonable extension of the authors' earlier PAM/MOSE work to include node activation and O-RAN timing constraints. The feasibility maps in Figs. 15-16 are useful.\n\nSoft spots, in order of severity. (1) The internal inconsistency is load-bearing. Setting T_max_Dk=300 s makes SM feasible in the optimization, but the paper itself documents that even a single stateful xApp experiences several seconds of downtime, violating the 1 s near-RT deadline. The abstract's \"ensuring xApp availability during migration\" is therefore inaccurate for the SM results. (2) The energy gain in Fig. 14 is the optimum of the SAL objective evaluated with the same fitted coefficients that enter the model. That is not a measured system saving; it is circular to the extent that the model is assumed correct. An end-to-end testbed implementation of the SAL solution is needed to validate the claimed gains. (3) The NP-hardness proof in Theorem 1 is logically invalid: being an MIQP does not imply NP-hardness, and reducing your problem to the general MIQP formulation proves nothing. This is fixable and not central to the contribution, since they solve with Gurobi anyway. (4) Extrapolating linear fits from one DRL xApp and a four-node cluster to hundreds of arbitrary AI xApps is a concern, but secondary.\n\nBottom line: the paper deserves a serious referee, but the quantitative claim needs major rework. I would send it to peer review with a demand for measured end-to-end evaluation under real O-RAN deadlines and a correction to the NP-hardness argument. The experimental comparison is the part worth keeping.","headline":"Genuinely useful SM-vs-SDL measurements for O-RAN xApps, but the headline 64% energy saving is an artifact of a downtime budget that contradicts the paper's own near-RT deadline.","tokens_in":25096,"tokens_out":5152,"would_cite":true,"duration_ms":55220,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CORMO-RAN is a data-driven orchestrator that jointly decides which RIC compute nodes stay on and how to migrate xApps without losing state, and the paper claims it cuts cluster energy by up to 64 percent against a standard load balancer.","keywords":["Open RAN","xApps","stateful migration","shared data layer","RAN Intelligent Controller","energy efficiency","near-RT RIC","etcd"],"falsifier":"Run the CORMO-RAN optimizer end-to-end on a production RIC cluster with more than a hundred real, non-emulated stateful xApps of different AI architectures and state sizes, then compare measured energy against a load-balanced always-on baseline; if the saving falls far short of 64% or any migration exceeds the 1-second near-real-time loop deadline, the central claim fails. A cheaper check is to refit the linear slopes from a second, architecturally different xApp and see whether the SAL solution still predicts measured energy within a small error.","tokens_in":24065,"feed_emoji":"⚡","tokens_out":14034,"duration_ms":124559,"temperature":0.7,"pith_summary":"The paper asks how to keep the compute cluster behind an O-RAN RAN Intelligent Controller from wasting energy at low-traffic times, when few xApps (AI-based applications that monitor and control the RAN) are needed but many nodes stay on. It argues that the answer is to treat xApp migration and node activation as one optimization problem, and it builds CORMO-RAN, an application on the non-real-time controller that turns off underused nodes and migrates stateful xApps to the remaining nodes while preserving their internal state. The central claim is that this joint control yields up to a 64% reduction in cluster energy consumption compared with the platform's native load-balancing scheduler, and the paper supports the claim with measurements from a four-node testbed and an optimization-based evaluation. CORMO-RAN also decides which of two migration techniques—container state transfer or a shared data layer—is feasible under the near-real-time 1-second control deadline.","feed_headline":"Lossless 5G app migration saves up to 64% control energy","feed_subtitle":"CORMO-RAN turns off idle compute nodes and shifts stateful xApps, cutting RIC cluster power use by up to 64 percent.","key_machinery":"The load-bearing machinery is the SAL (Server Activation and Lossless migration) problem, a mixed-integer quadratic program whose objective sums per-server energy from idle power, per-xApp load, and migration-strategy overhead. The models that feed it are linear fits derived from the testbed measurements: migration downtime and duration grow linearly with the number of xApps for both SM variants, SDL migration duration is independent of state size and xApp class, the SDL backend's maintenance downtime grows linearly with total offered load and state size, and per-xApp CPU and power consumption are linear in the number of xApps per class. These fitted slopes and intercepts convert the engineering trade-off between SM's downtime and SDL's backend cost into numeric parameters that the optimizer can respect while enforcing the near-real-time 1-second deadline constraints.","core_discovery":"CORMO-RAN's central contention is that the two established ways to move a stateful xApp—Stateful Migration (SM), which moves the container with its memory, and the O-RAN Shared Data Layer (SDL), which keeps state in an external key-value store such as etcd—have complementary failure modes: SM violates the 1-second near-real-time control deadline whenever downtime matters, while SDL's zero-downtime migration is bounded by backend database maintenance outages that grow linearly with load and state size. The paper's contribution is a working reconciliation: it fits linear models of downtime, migration duration, maintenance outage, and per-xApp resource and energy use from testbed experiments, encodes them as a mixed-integer quadratic program called the SAL (Server Activation and Lossless migration) problem, and solves it with branch-and-bound to choose both which nodes to power on and which migration strategy to use. Solving this problem in under a second for up to about 120 xApps, CORMO-RAN is claimed to achieve up to 64% energy savings over a load-balanced baseline while keeping xApp state intact and migration downtime within permitted bounds.","pith_inferences":["The same SAL machinery could be applied beyond xApps, for example to rApp placement or to any stateful microservice on a Kubernetes cluster, since the fitted KPI/resource models are only tied to O-RAN through the deadline constraints.","Because the fitted models come from a single reinforcement-learning xApp and an emulator, the 64% figure is best read as an upper bound; refitting with heterogeneous AI xApps would likely lower achievable savings in practice.","The feasibility maps imply a hybrid policy the authors do not explore: use SDL for small-state xApps during peak hours for zero-downtime migration, and switch to SM for large-state xApps at night when whole nodes can be freed.","Operators could test sensitivity by varying the backend maintenance period online: the model predicts that rarer maintenance cuts SDL energy but shrinks the feasible region, so an adaptive maintenance schedule might capture most of the energy gain without risking deadline violations."],"forward_implications":["Operators can power down RIC nodes during low-traffic periods and move stateful xApps between nodes without dropping near-real-time control loops, shifting cluster energy use to track actual xApp demand.","SDL is the right choice only when state is small and load moderate; for large state sizes or high load, the paper's feasibility analysis shows SDL cannot meet the 1-second deadline, so SM becomes the fallback even though its downtime exceeds the deadline.","The SAL problem, though NP-hard, is solvable optimally in under a second for realistic cluster sizes (up to about 120 xApps), so hourly re-optimization from the non-real-time RIC is practical.","Under the paper's model, energy gains approach zero when all nodes must stay on; at high traffic the system matches the baseline, so the benefit is concentrated in off-peak hours.","The best migration strategy depends on the dominant xApp class, state size, and backend maintenance period, and the paper supplies a feasibility map for choosing a viable strategy before solving the allocation."],"supporting_citations":[{"why":"Defines the O-RAN architecture and the Shared Data Layer that CORMO-RAN uses as one migration strategy.","marker":"[3]"},{"why":"Shows xApp energy consumption scales with their number, motivating the energy-aware orchestration.","marker":"[4]"},{"why":"Provides the analytical model of stateful migration downtime and duration as functions of memory and dirty-page rate.","marker":"[5]"},{"why":"The orchestration framework used to implement the SM variants (SM-MR and SM-MD).","marker":"[24]"},{"why":"Compares key-value store implementations for shared data and supports the choice of a strongly consistent backend for SDL.","marker":"[31]"},{"why":"The key-value store used as the SDL backend; its compaction and defragmentation drive the maintenance-outage model.","marker":"[36]"},{"why":"The open-source near-RT RIC platform used in the testbed.","marker":"[40]"},{"why":"The pre-trained DRL xApp used in the experiments, providing the reference AI workload for fitting the linear models.","marker":"[49]"},{"why":"Establishes that the decision version of mixed-integer quadratic programming is NP-complete, used to prove SAL is NP-hard.","marker":"[52]"}],"fun_headline_variants":["Lossless xApp migration cuts RIC energy by 64%","Stateful xApp migration saves 64% in O-RAN control power","CORMO-RAN: Shift xApps losslessly, cut cluster energy 64%","Zero-downtime xApp migration shrinks RIC power by 64%","Turn off idle nodes, move xApps safely, save 64% energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that linear relationships measured with one reinforcement-learning xApp and an emulated RAN interface on a four-node cluster continue to hold when hundreds of diverse AI xApps run on a larger production cluster; the 64% saving is computed from those extrapolated fits, not from a measured deployment.","fun_headline_variants_meta":{"raw":{"variants":["Lossless xApp migration cuts RIC energy by 64%","Stateful xApp migration saves 64% in O-RAN control power","CORMO-RAN: Shift xApps losslessly, cut cluster energy 64%","Zero-downtime xApp migration shrinks RIC power by 64%","Turn off idle nodes, move xApps safely, save 64% energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2783,"prompt_tokens":1073,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1607}},"tokens_in":689,"tokens_out":1710,"duration_ms":11893,"temperature":1.0,"reasoning_tokens":1607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:25:24.901067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the CORMO-RAN optimizer end-to-end on a production RIC cluster with more than a hundred real, non-emulated stateful xApps of different AI architectures and state sizes, then compare measured energy against a load-balanced always-on baseline; if the saving falls far short of 64% or any migration exceeds the 1-second near-real-time loop deadline, the central claim fails. A cheaper check is to refit the linear slopes from a second, architecturally different xApp and see whether the SAL solution still predicts measured energy within a small error.","supporting_citations":[{"cited_title":"Under- standing O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges,","cited_arxiv_id":null,"evidence_quote":"Defines the O-RAN architecture and the Shared Data Layer that CORMO-RAN uses as one migration strategy."},{"cited_title":"ScalO-RAN: Energy-aware Network Intelligence Scaling in Open RAN,","cited_arxiv_id":null,"evidence_quote":"Shows xApp energy consumption scales with their number, motivating the energy-aware orchestration."},{"cited_title":"Design, Mod- eling, and Implementation of Robust Migration of Stateful Edge Mi- croservices,","cited_arxiv_id":null,"evidence_quote":"Provides the analytical model of stateful migration downtime and duration as functions of memory and dirty-page rate."},{"cited_title":"Mose: A novel orchestration framework for stateful microservice migration at the edge,","cited_arxiv_id":null,"evidence_quote":"The orchestration framework used to implement the SM variants (SM-MR and SM-MD)."},{"cited_title":"Enabling efficient collection and usage of network performance metrics at the edge,","cited_arxiv_id":null,"evidence_quote":"Compares key-value store implementations for shared data and supports the choice of a strongly consistent backend for SDL."},{"cited_title":"A distributed, reliable key-value store for the most critical data of a distributed system,","cited_arxiv_id":null,"evidence_quote":"The key-value store used as the SDL backend; its compaction and defragmentation drive the maintenance-outage model."},{"cited_title":"RIC Platform GitHub Repository,","cited_arxiv_id":null,"evidence_quote":"The open-source near-RT RIC platform used in the testbed."},{"cited_title":"Mixed-integer quadratic programming is in np,","cited_arxiv_id":null,"evidence_quote":"Establishes that the decision version of mixed-integer quadratic programming is NP-complete, used to prove SAL is NP-hard."}],"review_version":2}