{"id":"e941f338-8d8e-4e9a-8006-99203d4f12f0","arxiv_id":"2411.16086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"HiDP hierarchically partitions DNN inference at both the cluster and the core level and reports large latency, energy, and throughput gains over re-implemented baselines on Jetson and Raspberry Pi edge clusters.","lead":"This paper presents HiDP, a two-tier strategy that splits DNN inference workloads across edge devices and then splits each device's assigned share across its CPU and GPU cores. The authors report 38% lower latency, 46% lower energy, and 56% higher throughput than three baseline distributed inference methods on commercial edge hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported latency/energy/throughput advantage may be inflated because the three baselines are re-implemented with HiDP's own partitioning modules rather than the original systems, making the comparison a partial self-comparison.","rationale":"The reader's weakest_assumption focuses on the latency model in Eqs. (5)-(6). That is a legitimate concern: those equations define Θ as a product of a vector γ and a scalar ω or σ, which is dimensionally inconsistent and not a validated latency predictor. However, the more load-bearing issue for the paper's central claim is the baseline construction. The reported numbers are the entire evidence for the headline advantage, and the text explicitly states that all three baselines are implemented using HiDP's own modules. This means the comparison is not against the original systems but against HiDP with certain capabilities disabled. Even a perfect latency model would not salvage the comparison if the baselines are strawmen. I therefore partially agree with the reader: the model is weak, but the baseline fidelity is a prior and more direct threat to the central claim. The recommended verdict remains CONDITIONAL, but the condition should explicitly require a fair, independent baseline evaluation (and ideally released code), not merely a refined latency model. The paper's idea of hierarchical partitioning with core-level awareness is valuable and could be correct; the current evidence, however, does not substantiate the claimed margins.","tokens_in":9136,"tokens_out":4050,"duration_ms":40433,"concrete_test":"Require the authors to run the same workloads and cluster with the original code or faithful independent re-implementations of MoDNN, OmniBoost, and DisNet (following their published algorithms, with custom schedulers and partitioning logic not derived from HiDP). If HiDP's average latency reduction against these faithful baselines falls below the claimed 38%, or if any baseline outperforms HiDP, the central claim must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that HiDP beats 'other relevant approaches' rests on the experiments in Section IV. In the 'Comparison w.r.t. state-of-the-art approaches' paragraph, the authors state: 'We implemented MoDNN using the data partitioning module of HiDP framework.' For DisNet: 'We used the data and model partitioning algorithm of HiDP to implement DisNet.' For OmniBoost: 'We implemented the throughput estimator of Omniboost using Gymnasium library and trained it on our target workloads.' These are not faithful re-implementations of the original systems. MoDNN is reduced to HiDP's data partitioner; DisNet is reduced to HiDP's global partitioning without local optimization; OmniBoost is reduced to a throughput estimator, omitting its Monte-Carlo tree search and scheduling components. Thus the baselines are effectively ablated versions of HiDP itself, sharing HiDP's communication, scheduling, and execution stack. The reported 38% latency reduction, 46% energy reduction, and 56% throughput improvement may therefore reflect the gap between HiDP and its own stripped-down variants, not a gap against the actual published methods. The absence of released code and the lack of error bars further prevent independent verification. While the hierarchical idea is plausible and the hardware testbed is real, the experimental evidence as described does not establish the headline comparative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HiDP, a two-tier hierarchical DNN partitioning strategy for distributed inference on heterogeneous edge clusters. At the global level, HiDP partitions the DNN model (model partitioning) or input data (data partitioning) across edge nodes; at the local level, it further partitions the assigned workload across each node's heterogeneous CPU/GPU cores. The framework includes a runtime scheduler, a dynamic programming-based design-space exploration, and a Python middleware. The evaluation is performed on a real cluster of five commercial edge platforms (Jetson Orin NX, Jetson Nano, Jetson TX2, Raspberry Pi 4B, Raspberry Pi 5) using four DNNs (ResNet-152, EfficientNet-B0, VGG-19, InceptionNet-V3). The paper reports lower latency, lower energy, and higher throughput compared with three prior strategies: MoDNN, OmniBoost, and DisNet.","tokens_in":9425,"tokens_out":4094,"duration_ms":37921,"significance":"If the comparative claims hold, HiDP would be a meaningful contribution to edge inference: the idea of jointly optimizing global and local partitioning with core-level awareness is plausible, and the use of a real heterogeneous testbed with five boards and four DNNs is a genuine strength. The framework is lightweight (reported 400-line middleware and ~15 ms DP overhead), and the measured trends across varying cluster sizes and workload mixes are informative. However, the headline comparative claim is undermined by the way the baselines are implemented, and by the absence of statistical validation. The central contribution is defensible but the evidence as presented does not currently establish the reported advantage over the actual published systems.","major_comments":[{"comment":"The three baselines are not faithful re-implementations of the original published methods. The text states: 'We implemented MoDNN using the data partitioning module of HiDP framework', 'We used the data and model partitioning algorithm of HiDP to implement DisNet', and for OmniBoost only 'the throughput estimator' was implemented, omitting the Monte-Carlo tree search and scheduling components. This means the comparison is effectively between HiDP and ablated variants of HiDP that share the same communication, scheduling, and execution stack. The reported 38% lower latency, 46% lower energy, and 56% higher throughput may therefore reflect the gap between HiDP and its own restricted versions, not a gap against MoDNN, OmniBoost, and DisNet as actually published. To support the central comparative claim, the authors must implement the actual algorithms of these baselines (including MoDNN's local scheduling, DisNet's heuristic, and OmniBoost's MCTS-based search), or use original code if available, and ideally validate the re-implementations against published results on a common benchmark.","section":"Section IV.A, 'Comparison w.r.t. state-of-the-art approaches'"},{"comment":"All quantitative results are reported as single-point measurements with no error bars, confidence intervals, or repeated-trial statistics. Edge devices are subject to DVFS, thermal throttling, and wireless variability; without repeated runs, the reported latency/energy/throughput differences may be within run-to-run noise. The paper should report mean and standard deviation over multiple runs (at least 5-10), and state the number of trials, for every figure in Section IV. This is load-bearing because the headline improvements are comparative and would be more convincing if shown to be statistically significant.","section":"Section IV.B, 'Experimental Results'"},{"comment":"The latency model used for partitioning decisions is not validated. Eqs. (5)-(6) define total computation time Θ as γ · ω (or γ · σ), where γ is the computation-to-communication ratio vector and ω/σ is the block width or submodel count. The units of γ and the justification for this multiplicative form are unclear, and no empirical calibration is provided to show that minimizing Θ indeed minimizes measured latency on the testbed. Since the DP algorithm in Algorithm 1 selects partitions based on this model, the paper should demonstrate that the chosen partitions are optimal or near-optimal on the actual hardware, for example by comparing predicted versus measured latency across partitioning configurations (as in Figure 1, but including the model's predictions). Without this, the claimed optimality of HiDP's decisions rests on an unvalidated heuristic.","section":"Section III, Eqs. (5)-(6)"},{"comment":"The headline number '38% lower latency' in the abstract is not consistent with the numbers reported in the paper. Section IV.B reports 'On average HiDP has 37%, 44%, and 56% lower latency than DisNet, OmniBoost, and MoDNN', which average to about 46%, not 38%; Figure 8 reports 30%, 46%, and 38% for different experiments; the conclusion states 'latency and energy improvements of 38%, and 46%'. The origin of the abstract's 38% figure is unclear. This is a load-bearing issue because the abstract is the primary statement of the contribution. The authors should harmonize the abstract, results section, and conclusion, and report how each aggregate number was computed.","section":"Abstract and Section IV.B/Conclusion"}],"minor_comments":[{"comment":"The caption says 'different workload partitioning configurations (P1-P9)', but the text does not define what P1 through P9 mean concretely (number of data partitions and CPU-GPU split). Adding a table or a sentence enumerating the configurations would improve reproducibility.","section":"Figure 1"},{"comment":"The text says 'Each device hosts Linux 18.04 OS' - this is likely Ubuntu 18.04, not Linux 18.04. Also, '80 MBps wireless control' is ambiguous: it should be clarified whether this is megabytes per second or megabits per second, and whether it is the Wi-Fi link speed or measured throughput.","section":"Section IV.A, 'Middleware'"},{"comment":"The column 'heterogeneous block size' is not explained; it is unclear what makes a block size 'heterogeneous' as opposed to the other partitioning-related columns. A definition or footnote would help.","section":"Table I"},{"comment":"The notation for communication rates µ_k and β_φ is introduced as 'transmission overhead between two processors/nodes for a given time duration t', which is not a conventional rate. The equations would benefit from explicit units or a sentence explaining what a 'scalar' communication rate represents and how it is measured.","section":"Section III, Eqs. (1)-(6)"},{"comment":"The y-axis label 'Gigaflops/s' is non-standard; consider using 'GFLOP/s' or 'GFLOPS'. Also, the figure caption says 'Performance (Gigaflops/s)', which is fine, but the axis text should be consistent.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is already accepted at DATE 2025, so this report should be read as input for a revision or a published version rather than a gatekeeping decision. The most serious concern is the baseline re-implementation issue; if the authors can obtain or faithfully recreate the actual baselines, or at least clearly delimit that the comparison is against 'HiDP with restricted partitioning modes', the contribution may be salvageable. The lack of error bars is also important but is a more standard fix. I would not reject outright because the hierarchical idea is plausible and the hardware demonstration is real, but the central comparative claim currently overstates what the evidence shows."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the hierarchical idea, not for the headline percentages. The two-level split — global model/data partitioning across the cluster, then core-level CPU/GPU partitioning inside each node — is a real extension of DisNet and the single-node heterogeneity schedulers. The paper shows convincingly that default TensorFlow schedules on the GPU and leaves CPU cycles idle, and that the optimal CPU/GPU split varies by model. The hardware testbed is real: four Jetson boards and two Raspberry Pis, four DNNs, dynamic workload mixes. That is a lot of honest measurement work.\n\nThe weak spot is the baseline comparison, and it is not minor. The authors state plainly that MoDNN was implemented using HiDP's data partitioning module, DisNet using HiDP's partitioning algorithm, and OmniBoost only through a re-trained throughput estimator with its Monte-Carlo tree search omitted. Those are not the original systems; they are ablated versions of HiDP. So the 38% latency, 46% energy, and 56% throughput gaps likely measure the distance between HiDP and its own stripped-down variants, not between HiDP and the published methods. The absence of error bars and the lack of released code make the numbers even harder to verify. The latency model is informal — Eqs (5) and (6) are basically product forms with a ratio and a width — and it may not transfer to workloads where memory contention or DVFS dominate, but the authors do not lean on the model for the headline; they lean on measurement.\n\nStill, the central claim is plausible: jointly optimizing global and local heterogeneity should help, and the experiments support that direction, even if the magnitude is uncertain. The paper is clearly written, the related work is appropriate, and the authors disclose exactly what they did, which is more than many systems papers do.\n\nWho gets value: anyone working on edge inference orchestration, especially on Jetson-class hardware. It deserves a serious referee, but the referee should push for faithful baseline implementations (or at least a discussion of what is lost in re-implementation), error bars, and code. I would not cite the headline numbers in my own work yet, but I would keep the paper on the radar for the hierarchical partitioning idea.","headline":"A plausible hierarchical partitioning idea with a real hardware testbed, but the baseline comparison is partially a self-comparison, so treat the headline gains as unverified.","tokens_in":9919,"tokens_out":2699,"would_cite":false,"duration_ms":22686,"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":"This paper proposes HiDP, a two-level DNN partitioning strategy that splits a model across heterogeneous edge nodes and then splits each node's share across its CPU and GPU cores, reporting 38% lower latency, 46% lower energy, and 56%…","keywords":["edge inference","DNN partitioning","heterogeneous edge platforms","hierarchical partitioning","model partitioning","data partitioning","CPU-GPU scheduling","distributed inference"],"falsifier":"Run the same four DNN models on a cluster where memory bandwidth or dynamic voltage/frequency scaling materially changes per-core compute rates, and compare the HiDP-chosen partition against a grid search over block widths and submodel counts; if the grid search finds a configuration with meaningfully lower latency than the model's optimum, the cost model's predictive power is refuted.","tokens_in":8936,"feed_emoji":"⚡","tokens_out":3686,"duration_ms":31025,"temperature":0.7,"pith_summary":"HiDP is a two-level strategy for running DNN inference across a cluster of heterogeneous edge devices. At the global level it partitions the model (or its input data) among nodes; at the local level it further splits each node's share across CPU and GPU cores, overriding the default framework behavior that usually runs on the GPU alone. The paper argues that this hierarchical view, which accounts for core-level heterogeneity both when creating blocks and when scheduling them, is what prior distributed inference methods miss. On four DNN models across five commercial boards (Jetson and Raspberry Pi), HiDP reports 38% lower latency, 46% lower energy, and 56% higher throughput on average compared with three state-of-the-art approaches. If correct, the claim is that heterogeneity-aware partitioning inside each node is as important as partitioning across nodes for edge inference.","feed_headline":"Two-tier DNN splitting cuts edge inference latency 38%","feed_subtitle":"A hierarchical partitioner also claims 46% lower energy and 56% higher throughput across Jetson and Raspberry Pi clusters.","key_machinery":"The cost model in Eqs. (5)–(6) estimates total computation time $\\Theta$ as the product of a computation-to-communication ratio $\\gamma$ (global $\\Psi$ or local $\\psi$) and a decision variable: block width $\\omega$ for model partitioning or number of parallel submodels $\\sigma$ for data partitioning. A standard subset-sum dynamic program searches this model to pick the partition with minimal $\\Theta$ at both tiers. The same DP routine is reused for global and local decisions because the function arguments are identical in form.","core_discovery":"The central discovery is that the optimal workload partition for a distributed DNN inference request cannot be found globally alone: the same model partition performs differently on different devices because each device has a CPU-GPU mix with different compute rates, and the default deep-learning runtime does not exploit that mix. HiDP therefore treats partitioning as a two-stage decision, guided by a dynamic-programming search over block widths (model partitioning) and sub-model counts (data partitioning), using measured computation-to-communication ratios at both the node level and the processor level. As a result, the configuration that wins on latency, energy, and throughput is one that earlier global-only schemes never consider.","pith_inferences":["The same hierarchical idea could extend to NPUs and other accelerators if the cost model gains per-core bandwidth and contention terms.","Because the paper reports accuracy identical to global-only methods, data partitioning appears to preserve correctness for these models; a natural test is whether that holds for segmentation or detection heads that share intermediate features.","The cost model's reliance on linear products of ratios suggests its predictions may degrade under heavy memory pressure; adding a memory-bandwidth term would be a testable extension."],"forward_implications":["Edge clusters can service more inference requests per unit time because shorter per-inference latency frees nodes for queued requests.","Energy per inference drops alongside latency, which matters for battery-operated and passively cooled edge devices.","Dynamic workload mixes benefit from choosing data versus model partitioning per request based on DNN characteristics.","The advantage of local core-level partitioning grows as the number of worker nodes shrinks, since the local node becomes the bottleneck."],"supporting_citations":[{"why":"MoDNN is the data-partitioning baseline that HiDP must beat in latency, energy, and throughput.","marker":"[4]"},{"why":"DisNet is the hybrid partitioning baseline whose data/model partitioning algorithm is reused to implement one of the comparison strategies.","marker":"[5]"},{"why":"OmniBoost is the model-partitioning baseline using Monte-Carlo search that HiDP compares against.","marker":"[7]"},{"why":"TensorFlow is the default runtime whose GPU-only scheduling motivates the local partitioning tier.","marker":"[18]"},{"why":"Jetson TX2 is the platform used in the motivational experiments showing default partitioning latency.","marker":"[23]"},{"why":"CoEdge supplies the definition of computation rate as frequency over compute intensity, used in the cost model.","marker":"[24]"}],"fun_headline_variants":["Two-tier DNN partition cuts edge latency 38%","Hierarchical split beats flat edge inference by 38%","Edge AI: core-aware DNN split slashes latency 38%","Global + local DNN partitioning lifts edge throughput 56%","HiDP: two-level partition tames heterogeneous edge nodes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the latency model in Eqs. (5)–(6), which multiplies a measured computation-to-communication ratio by a block width or submodel count, predicts real inference latency accurately enough that the partition it picks is truly the fastest.","fun_headline_variants_meta":{"raw":{"variants":["Two-tier DNN partition cuts edge latency 38%","Hierarchical split beats flat edge inference by 38%","Edge AI: core-aware DNN split slashes latency 38%","Global + local DNN partitioning lifts edge throughput 56%","HiDP: two-level partition tames heterogeneous edge nodes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000813,"raw_usage":{"total_tokens":3498,"prompt_tokens":809,"completion_tokens":2689,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":2604}},"tokens_in":425,"tokens_out":2689,"duration_ms":17576,"temperature":1.0,"reasoning_tokens":2604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:33:22.524488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four DNN models on a cluster where memory bandwidth or dynamic voltage/frequency scaling materially changes per-core compute rates, and compare the HiDP-chosen partition against a grid search over block widths and submodel counts; if the grid search finds a configuration with meaningfully lower latency than the model's optimum, the cost model's predictive power is refuted.","supporting_citations":[{"cited_title":"MoDNN: Local distributed mobile computing system for Deep Neural Network,","cited_arxiv_id":null,"evidence_quote":"MoDNN is the data-partitioning baseline that HiDP must beat in latency, energy, and throughput."},{"cited_title":"Disnet: Distributed micro-split deep learning in heterogeneous dynamic iot,","cited_arxiv_id":null,"evidence_quote":"DisNet is the hybrid partitioning baseline whose data/model partitioning algorithm is reused to implement one of the comparison strategies."},{"cited_title":"Omniboost: Boosting throughput of heterogeneous embedded devices under multi-dnn workload,","cited_arxiv_id":null,"evidence_quote":"OmniBoost is the model-partitioning baseline using Monte-Carlo search that HiDP compares against."},{"cited_title":"Tensorflow,","cited_arxiv_id":null,"evidence_quote":"TensorFlow is the default runtime whose GPU-only scheduling motivates the local partitioning tier."},{"cited_title":"Jetson tx2 module,","cited_arxiv_id":null,"evidence_quote":"Jetson TX2 is the platform used in the motivational experiments showing default partitioning latency."},{"cited_title":"Coedge: Cooperative dnn inference with adaptive workload partitioning over heterogeneous edge devices,","cited_arxiv_id":null,"evidence_quote":"CoEdge supplies the definition of computation rate as frequency over compute intensity, used in the cost model."}],"review_version":1}