{"id":"5be34533-d9ce-403e-94d6-c9becf275da0","arxiv_id":"1908.01153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MAPO is a multi-objective genetic algorithm that places IoT application components in a fog network, trading off completion time, energy, and cost, and reports large gains over two prior placement methods.","lead":"MAPO places IoT app components on nearby fog computers while balancing three goals at once: speed, energy, and cost. It uses a genetic search to find trade-off solutions and picks one based on latency needs, tested on medical apps in simulation and on Raspberry Pi hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy objective is incomplete: Eq. (4) adds an undefined static-energy term Es, and Section VII reports testbed energy without power metering; the 23–68% energy savings are not empirically supported.","rationale":"The strongest claim is MAPO's triple improvement (7.3x completion time, 23-68% energy reduction, 27% cost reduction) over FSPP and EW. Of these, the energy claim is the least secured. The analytic model is incomplete because Es in Eq. (4) is never defined, the real testbed has no power instrumentation, and the simulation tables contradict a universal energy advantage for MAPO. I agree with the reader's conditional verdict rather than moving to acceptance or rejection: the placement method, comparison setup, and other objectives are described in enough detail to be tested, and the paper can be repaired by defining Es and adding real power measurements. But the paper cannot stand as-is with a headline energy number that is not reproducible and not empirically grounded. The proposed power-meter experiment is a feasible, decisive check. The verdict therefore remains CONDITIONAL; no change from the reader's assessment is needed, but the stated condition should explicitly include a complete definition of Es and validated energy measurements.","tokens_in":12272,"tokens_out":4681,"duration_ms":45144,"concrete_test":"Add a calibrated power meter to the Section VII.B testbed (one per Raspberry Pi and on the CDC host), rerun the three placement methods for the mental health care application at INSTR=1000 MI and Data=1 Mbit, and compare measured per-placement energy to the values predicted by Eqs. (4)-(7). If the measured MAPO advantage over EW and FSPP falls outside the reported 23-68% range, the energy model is not empirically validated and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim includes 'decreasing the energy requirements by 23-68%'. This claim rests on the energy model in Section III.D.2, but that model is incomplete. Eq. (4) defines E(mi,rj) = Ep + Em + Es; Eqs. (5)-(6) define Ep and Em, and Tables II-III list values for ϱp_j, ϱm_j, and ϵ_j. The static term Es(mi,rj) is never defined, and no value is provided for any device. Consequently, the energy objective f2 in Eq. (10) cannot be computed by a reader, and the reported energy values in Tables IV-IX and Figures 6-7 are not reproducible from the manuscript. The real-testbed evaluation in Section VII does not cure this: Section VII.B describes Raspberry Pi devices, Docker, tc, and nc, but no power meter or energy measurement instrumentation. The energy plots in Figures 6-7 therefore appear to be evaluations of the same analytic formula, not direct measurements. If Es is large, MAPO's placement decisions could change; if Es is omitted, the claimed energy savings may shift. Additionally, Section VI.C reports that EW consumes nearly 61% less energy than MAPO in the data-size experiments and 55% less in the workload experiments, so the 23-68% reduction in the abstract comes from a selective reading of the real-testbed comparisons only. The energy dimension of the central claim is load-bearing and currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAPO, a Pareto-based NSGA-II approach for placing IoT applications on fog devices, optimizing three conflicting objectives: completion time, energy consumption, and economic cost. Applications are modeled as finite state machines, and the approach is evaluated on three medical IoT case studies (augmented reality, insulin pump, mental health care) in both a simulated fog environment and a small real-world testbed based on Raspberry Pi devices. MAPO is compared against two state-of-the-art methods, FSPP and Edge-ward. The claimed contributions are up to 7.3x lower completion time, 23-68% energy savings, and up to 27% lower economic cost.","tokens_in":12633,"tokens_out":4768,"duration_ms":47397,"significance":"If the evaluation were fully supported, MAPO would be a useful contribution to the IoT-fog placement literature, particularly because it applies a genuine multi-objective Pareto search instead of a weighted sum, and because it reports both simulation and real-testbed evaluations. The state-machine application model and the formulation of the three objectives are clearly presented and follow standard practice. The strengths are the two-tier evaluation, the comparison against two relevant baselines, and the use of the jMetal framework. However, the energy model is incomplete (the static energy term Es is never defined), the energy results are not backed by physical measurements on the real testbed, and the abstract's headline energy and completion-time claims are selectively supported or contradicted by the paper's own tables. These issues are load-bearing for the central claims, but they are addressable in a revision.","major_comments":[{"comment":"The static energy term Es(mi,rj) is introduced in Eq. (4) but never defined, and no numerical value or formula is provided for any device class. Since the energy objective f2 in Eq. (10) is the sum of Eq. (4) over all components, the energy objective cannot be computed by a reader, and the energy entries in Tables IV-IX and Figures 6-7 are not reproducible from the manuscript. Either define Es as a device idle power scaled by time or provide its constant value; in either case, its contribution can change the Pareto front and the reported energy tradeoffs.","section":"Section III.D.2, Eq. (4)"},{"comment":"The real-testbed experiment is described with Raspberry Pi devices, Docker, tc, and nc, but no power meter or energy measurement instrumentation is mentioned. The reported 'energy consumption' in the real-world evaluation therefore appears to be computed from the same analytic model in Section III rather than measured directly. Since Section VII.A states the goal is to 'validate the simulation results', the energy results should either be based on direct power measurements or explicitly labeled as model-based estimates; the current text does not distinguish these.","section":"Section VII.B, Figures 6-7"},{"comment":"The abstract's claim of 'decreasing the energy requirements by 23-68%' is not supported by the simulation results. In the data-size experiments, Table V and Table VI show EW consuming roughly 61% less energy than MAPO (e.g., 25 kJ vs 50-64 kJ for the insulin pump, and 25.1-26.3 kJ vs 62.5-65 kJ for mental health care). In the CPU-workload experiments, Table IX shows EW consuming 55% less energy than MAPO (e.g., 25 kJ vs 63 kJ at 2000 MI). The 23-68% range appears to be drawn only from the real-testbed comparison against FSPP, which is a selective reading. The energy claim should be revised to reflect the full set of reported results or explicitly qualified to the specific FSPP comparison.","section":"Abstract, Sections VI.C.1 and VI.C.2"},{"comment":"The headline 'up to 7.3 times' improvement in completion time is not backed by any result in the paper. Section VII.C.1 reports at most 6.9 times improvement compared to EW and 3 times compared to FSPP; Section VI.C.1 reports reductions of up to 70% (approximately 3.3 times). The number 7.3 appears in the abstract and conclusion but does not correspond to any table or figure. Please correct the number or provide the specific result that supports it.","section":"Abstract, Sections VI.C, VII.C, and VIII"},{"comment":"The automated decision making module (ADM) that selects a single placement from the Pareto set is described only as 'extends on a simple and computationally efficient a-priori method [17]', without specifying the selection rule, weights, thresholds, or any parameter values. Because all reported objective values are those of the selected solution, the lack of an explicit ADM specification makes the evaluation not reproducible. Please provide the ADM decision formula and its parameter settings, or point to the exact equations in reference [17] that are used.","section":"Section IV and Section VI.A"}],"minor_comments":[{"comment":"In Eq. (1), the communication delay term appears as 'Datai BW k,j' without a division sign; it should be Datai / BWk,j, as is clear from the context and Eq. (8).","section":"Section III.D.1, Eq. (1)"},{"comment":"The paper states that objective results are 'averaged over 1000 runs for statistical significance', but no standard deviation, confidence interval, or error bars are reported anywhere. Please include variance information in the tables or an appendix to substantiate the statistical claim.","section":"Section VI.A, Tables IV-IX"},{"comment":"The power constants ϱp_j and ϱm_j in Tables II and III, as well as the hardware constant ϵ_j, are presented without a source or derivation. Please cite the reference for these values or state explicitly that they are assumed, since they directly determine the energy objective.","section":"Tables II and III"},{"comment":"The real testbed uses an Intel Core i7-7700 VM as the CDC and Raspberry Pi 3 B+ as MEs, but Table X lists CPU capacities in MIPS without explaining how they were obtained or mapped from the simulated configuration; a short description of the benchmark or mapping method would improve reproducibility.","section":"Section VII.B, Table X"}],"recommendation":"major_revision","confidential_remarks":"The core approach is publishable in principle, but the energy model is under-specified and the abstract overstates the energy and completion-time improvements relative to what the paper's own tables show. The authors should be asked to fix these issues and to either add direct power measurements or explicitly present energy as a model-based estimate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing here is the packaging: MAPO gives a concrete multi-objective formulation for fog application placement — FSM application model, three explicitly conflicting objectives (completion time, energy, cost), NSGA-II search, and a decision module on top of the Pareto front. The paper also does more evaluation than most in this subfield: simulation plus a real Raspberry Pi testbed, three medical case studies, and two baselines. The completion-time story is credible; up to 7.3x is plausible given that MAPO places components near the data sources while the baselines favor cloud or gateway.\n\nThe soft spot is energy. The abstract's 23–68% reduction is not what the simulation shows: Edge-ward beats MAPO on energy by 55–61% in most simulated configurations. The real-testbed figures are presented as the basis, but Section VII never mentions a power meter. The energy plots appear to be evaluations of the same analytic model rather than direct measurements. And that model cannot be reproduced from the paper: Eq. (4) includes a static term Es that is never defined, and the hardware constant epsilon_j in Eq. (6) never appears in any table. So the central energy claim is unsupported. I don't read this as deliberate overclaiming — the authors openly state that MAPO often loses to Edge-ward on energy in simulation — but the abstract reads as if the range applies to both baselines, which it does not.\n\nMinor issues: ADM parameters are undisclosed, NSGA-II hyperparameters are incomplete, and there are no error bars despite '1000 runs.' Missing comparison with the closest multi-objective baseline, the improved NSGA-II in [24], weakens the novelty positioning. These are fixable.\n\nWho this is for: people working on fog/edge placement who want a concrete multi-objective model and a testbed template; they will find the architecture clear and the case studies useful. Prior to acceptance, the authors need to define Es, give epsilon values, report either measured energy or clearly label computed values as model output, and rewrite the abstract to match the results. A serious referee should engage with it, not desk-reject it, but the referee should make the energy evidence the center of the revision.","headline":"A credible multi-objective placement method with a completion-time result that holds up, but the headline energy savings are not supported by the evidence as presented.","tokens_in":13135,"tokens_out":3498,"would_cite":false,"duration_ms":35642,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MAPO claims that Pareto-based multi-objective placement in fog computing can cut IoT application completion time by up to 7.3 times, energy consumption by 23-68%, and economic cost by up to 27% compared with single-objective methods on…","keywords":["fog computing","IoT application placement","multi-objective optimization","NSGA-II","Pareto frontier","energy consumption","economic cost","completion time"],"falsifier":"Run the mental health care application on the real-world testbed described in Section VII at the stated workloads and data sizes, measuring actual device power draw with a power meter at each Fog device, and compare measured energy to MAPO's predicted $E(A,R)$; a discrepancy exceeding the claimed improvement margin would refute the energy objective and the headline energy savings.","tokens_in":12102,"feed_emoji":"⚡","tokens_out":9361,"duration_ms":80375,"temperature":0.7,"pith_summary":"This paper argues that treating IoT application placement in a fog environment as a three-objective Pareto optimization problem performs better than optimizing a single objective or a weighted sum. The proposed system, MAPO, models an application as a finite state machine of lightweight components and searches for non-dominated placements that trade off completion time, energy consumption, and economic cost. In simulated and small real testbeds with medical applications (augmented reality, insulin pump, mental health care), MAPO claims reductions in completion time by up to 7.3 times, energy by 23-68%, and economic cost by up to 27% relative to FSPP and Edge-ward. If true, this gives fog operators a principled way to place latency-sensitive IoT services near the data source instead of relying on Cloud-heavy heuristics.","feed_headline":"Pareto placement cuts fog app time 7.3x and energy 68%","feed_subtitle":"Multi-objective search finds near-data placements that beat single-objective and weighted-sum methods on medical IoT workloads.","key_machinery":"The load-bearing mechanism is the combination of a finite-state-machine application model with three analytic objective functions and a non-dominated sorting genetic search. An application is a state machine whose states are lightweight components; placing component $m_i$ on device $r_j$ assigns each component a device, and the objectives are computed from the chain of transitions: completion time $T(A,R)$ from Eqs. (1)-(3) adds computation time (instructions over CPU speed) to communication time (data size over bandwidth); energy $E(A,R)$ from Eqs. (4)-(7) adds computation power $\\varrho^p_j$, communication power $\\varrho^m_j$, and a hardware constant $\\epsilon_j$; cost $C(A,R)$ from Eqs. (8)-(9) sums processing, storage, and ingress communication charges. NSGA-II evolves a population of full placements, ranks them by Pareto dominance, and returns a frontier of non-dominated solutions; an automated decision module then selects one low-latency placement for deployment. The key identity is Pareto dominance itself: one placement beats another only if it is no worse on every objective and strictly better on at least one.","core_discovery":"The paper's central claim is that the optimal placement of an IoT application in a fog hierarchy is not a single-objective scheduling problem but a three-way tradeoff among completion time, energy consumption, and economic cost, and that a Pareto search over these three objectives finds placements that dominate single-objective or weighted-sum heuristics. In the medical IoT case studies considered, MAPO reports completion time reductions of up to 7.3 times versus FSPP and Edge-ward, energy reductions of 23-68% versus FSPP in most settings, and cost reductions of up to 27%; the main tradeoff is that MAPO can consume up to about 54% more energy than Edge-ward, a tradeoff the Pareto formulation makes explicit rather than hides. The paper presents these numbers as evidence that near-data placement on low-capability mobile edge devices, chosen by multi-objective search, outperforms Cloud-centric or gateway-centric placement rules for latency-sensitive workloads.","pith_inferences":["Because the objective functions and search procedure are application-agnostic, the same Pareto placement machinery should transfer to other multi-tenant fog workloads (e.g., industrial control or smart retail); only the state machine and device constants need replacing.","The automated decision module is effectively a policy knob: switching from the low-latency rule to a cost-minimizing rule would select a different point on the same Pareto frontier, letting the same search serve different service-level objectives without re-optimization.","A cleaner test of the Pareto benefit would compare MAPO against a weighted-sum multi-objective baseline with the same three objectives, not only single-objective baselines; that would isolate whether the gains come from multi-objective search or from simply considering three criteria.","The analytic energy model could be calibrated per device with actual power measurements and then used not only at placement time but as a runtime scheduling signal; if calibration reveals large discrepancies, the reported 23-68% energy savings would need to be revised."],"forward_implications":["For latency-sensitive IoT applications, placing components on mobile edge devices close to sensors can cut completion time by 3-7x compared with Cloud- or gateway-oriented placements.","Operators who optimize only one objective (e.g., cost or latency) forgo placements that are better on the other two; the Pareto frontier makes the tradeoff explicit and lets a decision rule choose the operating point.","In the tested workloads, communication latency, not data volume, dominates completion time, so placement algorithms should prioritize device proximity over bandwidth or Cloud power.","Fog devices are the energy-efficient choice for small applications, while Cloud resources become competitive at high CPU workloads; a placement optimizer needs both tiers to navigate this crossover.","The search reaches stable solution quality (measured by hypervolume) around 12,500 evaluations and scales to 30 components with a modest increase in optimizer runtime."],"supporting_citations":[{"why":"Supplies the NSGA-II evolutionary algorithm that performs the Pareto search and ranks placements by dominance.","marker":"[9]"},{"why":"Defines the Fog Service Placement Problem (FSPP) baseline, a linear integer-programming single-objective method MAPO is compared against.","marker":"[22]"},{"why":"Defines the Edge-ward hierarchical best-fit baseline that places the final component on the ISP gateway and provides the second comparison method.","marker":"[13]"},{"why":"Provides the optimization framework in which MAPO is implemented, giving the genetic algorithm and evaluation harness.","marker":"[10]"},{"why":"Supplies the energy model with the power constants and hardware term used in Eqs. (4)-(6), grounding the energy objective.","marker":"[8]"},{"why":"Provides the hypervolume metric used to measure the quality of MAPO's Pareto frontiers in Section VI.","marker":"[25]"},{"why":"Supplies the low-latency a-priori decision strategy used to select a single placement from the Pareto set.","marker":"[17]"}],"fun_headline_variants":["Multi-objective fog placement tri-optimizes time, energy, cost","Pareto search finds fog app placements with 7.3x faster time","MAPO: Pareto placement beats single-goal heuristics in fog","Fog placement optimization: 7.3x time, 68% energy cuts","Pareto-based fog placement cuts cost 27%, energy up to 68%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy savings are computed from the analytic model in Eqs. (4)-(6) using listed power constants, not measured on the real testbed; if those constants are inaccurate, the claimed 23-68% energy reductions are not empirically validated.","fun_headline_variants_meta":{"raw":{"variants":["Multi-objective fog placement tri-optimizes time, energy, cost","Pareto search finds fog app placements with 7.3x faster time","MAPO: Pareto placement beats single-goal heuristics in fog","Fog placement optimization: 7.3x time, 68% energy cuts","Pareto-based fog placement cuts cost 27%, energy up to 68%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3695,"prompt_tokens":921,"completion_tokens":2774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2671}},"tokens_in":537,"tokens_out":2774,"duration_ms":16800,"temperature":1.0,"reasoning_tokens":2671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:21:47.761545+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the mental health care application on the real-world testbed described in Section VII at the stated workloads and data sizes, measuring actual device power draw with a power meter at each Fog device, and compare measured energy to MAPO's predicted $E(A,R)$; a discrepancy exceeding the claimed improvement margin would refute the energy objective and the headline energy savings.","supporting_citations":[{"cited_title":"Towards qos-aware fog service placement","cited_arxiv_id":null,"evidence_quote":"Defines the Fog Service Placement Problem (FSPP) baseline, a linear integer-programming single-objective method MAPO is compared against."},{"cited_title":"ifogsim: A toolkit for modeling and simulation of resource man- agement techniques in the internet of things, edge and fog computing environments","cited_arxiv_id":null,"evidence_quote":"Defines the Edge-ward hierarchical best-fit baseline that places the final component on the ISP gateway and provides the second comparison method."},{"cited_title":"jmetal: A java framework for multi- objective optimization","cited_arxiv_id":null,"evidence_quote":"Provides the optimization framework in which MAPO is implemented, giving the genetic algorithm and evaluation harness."},{"cited_title":"First hop mobile offloading of dag computations","cited_arxiv_id":null,"evidence_quote":"Supplies the energy model with the power constants and hardware term used in Eqs. (4)-(6), grounding the energy objective."},{"cited_title":"A faster algorithm for calculating hypervolume","cited_arxiv_id":null,"evidence_quote":"Provides the hypervolume metric used to measure the quality of MAPO's Pareto frontiers in Section VI."},{"cited_title":"Multi- objective service oriented network provisioning in ultra-scale systems","cited_arxiv_id":null,"evidence_quote":"Supplies the low-latency a-priori decision strategy used to select a single placement from the Pareto set."}],"review_version":1}