{"id":"91e3495b-4ef8-4865-863e-d8fd302c81e4","arxiv_id":"1908.06115","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Energy consumption of NWP model configurations on the Intel E5-2697v4 is U-shaped in wall-clock time, with an energy-optimal core count that differs from the runtime-optimal one; optical processors are more energy-efficient but not competitive in time or precision.","lead":"This report measures how much energy weather-forecasting codes use on an Intel-based Cray cluster, and finds a U-shaped relationship between energy use and runtime as the number of cores grows. It also compares a Fourier-transform code on the CPU with an optical processor, which uses far less energy but is slower and less precise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Acraneb multi-node energy curve is constructed from a linear assumption, not measured; the U-shape for that configuration is not empirical.","rationale":"I read the main quantitative claim as the U-shaped energy-versus-walltime relation, supported by dedicated-node PM-counter measurements. The most load-bearing weak point is not the optical comparison (already flagged by the reader and explicitly caveated in §5.3.2.1 as not demonstrated), but the Acraneb branch of the U-shape: it is the only configuration whose multi-node behavior is not measured at all. Because the right-hand side is constructed by assuming perfect linear energy scaling, the existence and position of the Acraneb minimum cannot be verified from Figure 3. The paper is transparent about this, and the U-shape remains plausible for BiFFT and ALARO, so the reader's CONDITIONAL verdict is not overturned; it is reinforced. The concrete test would settle whether the pseudo-parallel construction is conservative or misleading.","tokens_in":12899,"tokens_out":6910,"duration_ms":74013,"concrete_test":"Run a genuinely parallel Acraneb (or the Acraneb component inside ALADIN) on 2, 4, and 8 full cca nodes and measure node energy with the same PM-counter script used in §4.1. Compare each E_N to n×E_1 from the pseudo-parallel construction. If E_N/(n E_1) deviates from 1 by more than the run-to-run scatter in Figure 3, the Acraneb U-shape in that figure is an artifact of the linear assumption and must be re-derived from measured points.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 builds the Acraneb multi-node data by assuming 'the energy consumption of n full nodes equals n times that of 1 full node while the wall-clock time remains the same.' Each 'multi-node' point is therefore one node running an adapted workload, with energy multiplied by n and wall-clock time taken from that single node. The right-hand, rising branch of the U-shape for Acraneb in Figure 3 is a consequence of this linear extrapolation, not an observed phenomenon. The text immediately qualifies the wall-clock side by noting the pseudo-parallel runs scale worse than the D4.3 profiling data, so the curve would be 'stretched out to lower wall-clock times' in reality. The same linear additivity is used in §4.4 to build dwarf energy pie charts from standalone dwarf average powers. If real multi-node execution introduces non-additive energy (extra memory traffic, MPI/communication, uncore power), both the Acraneb U-shape and the ALARO dwarf contribution estimates change. This matters because the paper's headline U-shape claim is presented as an empirical result across configurations, and Acraneb is one of the four configurations supporting it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This deliverable reports node-level energy measurements with PAPI Cray PM counters on an Intel Broadwell cluster for four NWP-related workloads: the BiFFT dwarf, the Acraneb2 radiation dwarf, the ALARO 2.5 km reference configuration, and the COSMO-EULAG 2.2 km configuration. The central empirical claim is that energy consumption versus wall-clock time is U-shaped as the number of cores varies, because average node power is roughly constant below one full node and then scales approximately linearly with core count. The paper also estimates relative dwarf energy contributions in ALARO, compares BiFFT energy on CPU with projected Optalysys optical-processor performance, and discusses the energy roofline as a candidate metric. The resulting message is that energy-optimal core counts differ from wall-clock-optimal core counts for these workloads.","tokens_in":13155,"tokens_out":5403,"duration_ms":51720,"significance":"If the U-shape is confirmed by direct measurements, the observation is practically relevant for scheduling and procurement in NWP: users can trade a small increase in wall-clock time for substantial energy savings, and the energy-optimal parallelism is not the fastest configuration. The paper's measurement methodology, including dedicated nodes, idle-power baselines, repeated runs with error bars, and explicit caveats about unmeasured infrastructure energy, is a useful template for future energy studies in operational weather prediction. The core observation is not produced by a fitted model, and the average-power explanation is supported by the measured power-versus-core-count behavior. The Acraneb multi-node curve and the Optalysys projections are, however, extrapolations rather than measurements, so the breadth of the empirical claim exceeds the data; the energy roofline discussion is a reasonable forward-looking proposal but is not applied to the measurements.","major_comments":[{"comment":"The multi-node Acraneb points are not measured but constructed by the assumption that 'the energy consumption of n full nodes equals n times that of 1 full node while the wall-clock time remains the same.' This makes the rising branch of the Acraneb U-shape an artifact of the extrapolation rather than an observed phenomenon. Since the Executive Summary and Section 6 state the U-shape as a result across the measured configurations, the claim should be narrowed to BiFFT and ALARO, or the Acraneb curve should be explicitly labeled as a model-based estimate. The same section's caveat that the pseudo-parallel runs scale worse than the D4.3 profiling data reinforces this concern.","section":"Section 4.3, Figure 3"},{"comment":"The ALARO dwarf energy pie charts are built by multiplying wall-clock time fractions by average powers measured in separate dwarf runs. This assumes the energy contributions of BiFFT, Acraneb, and the rest of the model add linearly and that dwarf power remains unchanged when embedded in ALARO. Communication energy is explicitly excluded, so the claim in the Executive Summary that the energy pie charts are 'more representative of the relative computational workloads' is not supported unless this additivity assumption is stated and justified.","section":"Section 4.4, Figure 4"},{"comment":"Only three COSMO-EULAG data points are available, with the 1152-core run lost and runs below 72 cores impossible due to memory limits. Three points, on differing grids and timesteps, cannot establish the U-shape for this configuration. The paper should not imply that a U-shape was measured for COSMO-EULAG; at most the available points are consistent with the pattern observed elsewhere.","section":"Section 4.5, Figure 6"},{"comment":"The Optalysys energy comparison is based on projected operation times that, as the text states, 'have not been demonstrated,' and on an assumed packing of 8 BiFFT transformations into a 1920x1080 frame. The Executive Summary and Conclusion state that the optical processors are 'much less energy costly' without carrying this conditionality. Please present the comparison as a conditional projection or add the qualification to every summary statement of the finding.","section":"Sections 5.3.2.1 and 5.5"}],"minor_comments":[{"comment":"The processor is called the Intel Xeon E5-2695v4 here while the title, abstract, and introduction say E5-2697v4; please reconcile the inconsistency.","section":"Section 4.1"},{"comment":"The text says the optimal configuration appears to be 'in the range between 30-150 cores or so,' but the figure axes show energy versus wall-clock time; please clarify how the reader should map that statement to the displayed data points.","section":"Section 4.2, Figure 2"},{"comment":"The log-log power-versus-core-count plot is central to the explanation of the U-shape; adding a numerical fit or at least the measured slopes for the sub-node and multi-node regimes would make the 'slope close to 1' statement easier to verify.","section":"Section 4.4, Figure 5"},{"comment":"The energy roofline is presented as a promising metric but is not applied to any of the measured workloads in this deliverable; please clarify that it is a proposal for future work rather than a result of the current measurements.","section":"Section 3.2"},{"comment":"The 0.1 ms, 6.6 mJ entry for a 2048x1536 complex-to-complex 2D FT is inferred from the convolution scenario rather than measured directly; the footnote helps, but the caption could state this more prominently.","section":"Section 5.4, Table 2"},{"comment":"Several citations refer to ESCAPE deliverables (D4.3, D3.3) and internal profiling data; for a public audience, please indicate how those documents can be accessed or summarize the relevant profiling results needed to reproduce the comparisons.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a project deliverable rather than a conventional research paper, and its main value is the measured energy data and the practical U-shape observation. The Acraneb multi-node curve and the Optalysys figures are explicitly extrapolated or projected, so the current abstract and conclusions overstate the empirical coverage. The issues are fixable within the manuscript's scope by recalibrating the claims and labeling the model-based estimates; I do not see a fundamental methodological error that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper before reading. First, it contains real, new energy measurements for several NWP configurations on a specific Intel node, and those measurements are worth having. Second, one of the four U-shaped curves is not fully measured: the Acraneb multi-node points are built by assuming energy scales linearly with node count and wall-clock time stays constant. The text is honest about this, but it means the paper's central generalization is a bit weaker than the abstract suggests.\n\nWhat the paper does well: the energy measurement methodology is sound and described in enough detail to be credible. They use dedicated nodes with an idle-power baseline, and the error bars on the BiFFT and ALARO figures give the central U-shape real empirical support. The explanation for the U-shape in terms of average node power being roughly flat below one node and then scaling linearly with cores is plausible and matches the data for the measured configurations. The paper also explicitly flags its own limitations, including the pseudo-parallel Acraneb runs and the fact that communication energy is not accounted for. That is refreshingly candid for a project deliverable.\n\nThe soft spots are real but localized. The Acraneb curve in Figure 3 is an extrapolation, not an observation, and the rising right-hand branch follows directly from the linear assumption. The optical comparison in Section 5 is even more conditional: the Optalysys throughput figures are explicitly described as not yet demonstrated, so placing the black dot on the same plot as measured CPU data gives it more evidentiary weight than it currently has. There is also a minor inconsistency in the CPU model name (E5-2697v4 in the abstract, E5-2695v4 in the method section), and no raw data or scripts are provided.\n\nWho should read this: HPC operators and NWP developers who want practical guidance on choosing core counts to minimize energy. The measured U-shape for BiFFT and ALARO is credible, and the power-scaling explanation is a useful mental model. If the optical comparison and the Acraneb extrapolation were clearly separated from the empirical results, this would be a more solid contribution. I would bring it to a reading group if the conversation is about energy measurement methodology, and I would cite it for the measured U-shape data.\n\nMy recommendation: this deserves serious referee time, but not as-is. I would send it back with requests to (1) clearly mark which points are measured versus extrapolated, (2) move the optical comparison to a clearly labeled projected estimate, and (3) add a data availability statement. If those changes were made, the central result would stand on much firmer ground.","headline":"Genuinely useful energy measurements for NWP cores, but the headline U-shape rests on one extrapolated curve and an optical comparison built on projected throughput.","tokens_in":13647,"tokens_out":1721,"would_cite":true,"duration_ms":18517,"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":"Four weather-prediction workloads show energy consumption that first falls and then rises as core count increases, with a local minimum at an intermediate configuration.","keywords":["energy efficiency","numerical weather prediction","high-performance computing","energy roofline","power measurement","BiFFT","ALARO","COSMO-EULAG"],"falsifier":"Measure total system energy including network switches, network interface cards, and cooling for the ALARO reference configuration at the reported core counts; if communication-plus-cooling energy grows faster than compute-node energy as N increases, the rising arm of the U-curve steepens and the energy-minimizing N is smaller than the one reported.","tokens_in":12699,"feed_emoji":"⚡","tokens_out":8252,"duration_ms":74356,"temperature":0.7,"pith_summary":"This report measures energy consumption of four numerical-weather-prediction workloads on the Intel E5-2697v4 processor and finds a U-shaped relation between consumed energy and wall-clock time as the number of cores is varied. Each configuration has an intermediate core count where energy use is locally minimal; using more cores shortens runtime but eventually costs more total energy. The U-shape is explained by node-level average power: below one full node, power stays roughly constant as cores are added, while once entire nodes are used average power grows about linearly with core count. The authors propose the energy roofline, an energy analogue of the time-based roofline, as a metric for deciding whether code is compute-bound or memory-bound in energy terms. The results imply that energy-optimal and wall-clock-optimal configurations are not the same, so operational choices involve an explicit tradeoff.","feed_headline":"Energy use vs core count is U-shaped in weather runs","feed_subtitle":"Above a sweet spot, extra cores shorten runtime but spend more total energy; below it, energy stays high.","key_machinery":"The load-bearing object is the energy-versus-wall-clock curve, measured at compute-node level, and its explanation is the average-power function P(N). For a partially filled node, P is nearly independent of the number of active cores; once a full 36-core node is used, P grows approximately linearly with N. Combined with the observed leveling off of wall-clock time at high core counts, this makes E(N)=P(N)t(N) U-shaped. To judge whether a code is compute-bound or memory-bound in energy terms, the report also introduces the energy roofline, an 'arch line' analogue of the time-based roofline whose balance point separates energy-compute-bound from energy-memory-bound regions; the gap between time-balance and energy-balance points measures how much harder energy optimization may be than time optimization.","core_discovery":"The paper's central claim is that for the BiFFT and Acraneb dwarfs and for the ALARO 2.5 km and COSMO-EULAG reference configurations, the energy consumed by a run, plotted against its wall-clock time as the core count is varied, follows a U-shaped curve with a local minimum at some core count N. The mechanism is the average power of a compute node: it rises slowly for partial-node runs, then becomes nearly linear in the number of cores once full nodes are used; because real applications are not perfectly scalable, wall-clock time eventually stops improving, so total energy E = P × t starts increasing linearly with N. The U-shape does not define a universal optimum, since some users need minimum wall-clock time rather than minimum energy. The paper also reports that relative energy contributions of the BiFFT dwarf to ALARO differ from its wall-clock contributions because communication energy is not measured, and that optical processors are much less energy-hungry for BiFFT but slower and numerically inadequate.","pith_inferences":["If communication energy were included in the measurements, the right side of the U-curve would likely rise more steeply, shifting the energy-minimizing core count to a lower value than reported.","The U-shape analysis suggests a practical diagnostic: for any strongly scaled code, plotting energy against wall-clock time separates the flat-power regime from the linear-power regime; deviations from the expected shape could expose communication or load-balance overheads.","The energy roofline's balance gap implies that a code with arithmetic intensity between the time-balance and energy-balance points is simultaneously compute-bound in time and memory-bound in energy; raising its arithmetic intensity would help energy efficiency more than time efficiency.","The optical-processor comparison hints at a crossover resolution where the constant wall-time scaling of optics would beat the n-log-n scaling of CPUs, but the paper's own numerical-precision caveat suggests this crossover is moot until the precision problem is solved."],"forward_implications":["Each measured configuration has a core count that minimizes energy for the fixed workload; beyond it, additional cores buy little runtime and cost extra energy.","Energy-optimal and runtime-optimal configurations differ, so an operational scheduling decision is required: pay more energy for faster forecasts or accept slower forecasts to save energy.","Because communication and cooling energy are not accounted for, the reported energy shares of individual dwarfs are closer to computational-workload shares than the wall-clock shares, which include communication overhead.","The energy roofline, if instrumented at high sampling frequency, could identify whether a loop or subroutine is easier to optimize for time or for energy, via the time-energy balance gap.","Optical processors can beat the CPU on energy alone for the BiFFT workload, but their wall-clock time and numerical precision keep them non-competitive for NWP at current resolutions."],"supporting_citations":[{"why":"Supplies the energy roofline model and the time-energy balance-gap concept used to propose an energy metric.","marker":"Choi et al. 2013"},{"why":"Defines the ALARO dynamical kernel that underlies the ALARO 2.5 km reference configuration measured here.","marker":"Bénard et al. 2010"},{"why":"Provides the ARPEGE/ALADIN framework equations on which the ALARO reference configuration is built.","marker":"Bubnova et al. 1995"},{"why":"Describes the COSMO-EULAG dynamical core and the weather-dwarf framework used for the COSMO-EULAG reference configuration.","marker":"Piotrowski et al. 2018"},{"why":"Reports the wall-clock time scaling and dwarf contribution data that the energy measurements are compared against.","marker":"D4.3"},{"why":"Documents the numerical accuracy of the optical processors, used to judge them non-competitive despite lower energy.","marker":"D3.3"}],"fun_headline_variants":["U-shaped energy curve found for weather model runs","Energy vs cores: U-shaped curve for weather runs","Weather runs: energy minimum at optimal core count","More cores don't always cut energy in weather models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central result assumes that the energy consumed by inter-node communication and by cooling can be neglected, so compute-node energy readings capture the full story; if communication energy grows with core count, the U-shaped curve's rising side is understated and the energy-minimizing core count would be lower.","fun_headline_variants_meta":{"raw":{"variants":["U-shaped energy curve found for weather model runs","Energy vs cores: U-shaped curve for weather runs","Weather runs: energy minimum at optimal core count","More cores don't always cut energy in weather models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000966,"raw_usage":{"total_tokens":4184,"prompt_tokens":1091,"completion_tokens":3093,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":3032}},"tokens_in":707,"tokens_out":3093,"duration_ms":19238,"temperature":1.0,"reasoning_tokens":3032,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:55:26.873716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure total system energy including network switches, network interface cards, and cooling for the ALARO reference configuration at the reported core counts; if communication-plus-cooling energy grows faster than compute-node energy as N increases, the rising arm of the U-curve steepens and the energy-minimizing N is smaller than the one reported.","supporting_citations":[],"review_version":1}