{"id":"01c4b53d-7f35-4c54-a9ce-da1ef9d0dc71","arxiv_id":"2411.17118","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GASKAP-HI imaging workflow was successfully ported to DUG Technology's commercial supercomputer, producing four pilot-survey image cubes at an estimated A$5,250 per field in 11 elapsed days.","lead":"Radio astronomers ported a large-scale imaging pipeline for the GASKAP-HI survey to a commercial supercomputer, producing four science-ready image cubes and documenting a four-step porting process. The trial gives the community concrete cost, time, and optimization data for using commercial HPC for future SKA-era processing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4's per-field cost estimate is internally inconsistent: the imaging step lists 528 jobs while the text states 2,112 native-resolution channels are imaged separately, so the A$5,250 figure may be understated by roughly a factor of three.","rationale":"I read the paper in good faith. The central claim has two parts: that commercial HPC can process GASKAP-HI pilot data, and that the full survey can be processed at about A$5,250 per field in about 11 days. The first part is well supported by direct evidence: four image cubes were produced, one field was compared with prior processing, and the workflow and scripts are publicly available. The second part is load-bearing for sub-objective 2 and for the abstract's headline resource estimate, and it is not independently supported. The weakest link is the internal accounting in Table 4, especially the unresolved tension between 2,112 native channels and 528 imaging jobs. This is a correctness risk internal to the paper, not a disagreement with external consensus. The reader's weakest assumption identifies the same issue, and I agree with the CONDITIONAL verdict. The concrete test is a source-level accounting check that would settle the concern: if the true imaging job count is 2,112, the cost estimate changes by roughly a factor of three; if it is 528, the authors need to document the grouping and otherwise reconcile the table. No change to the reader's verdict is needed, but the paper should be revised to make Table 4 self-consistent.","tokens_in":17268,"tokens_out":10382,"duration_ms":89089,"concrete_test":"Reconstruct the SBID 38466 accounting from the DUG RJS/SLURM job logs and storage records: count the number of WSClean imaging invocations and sum actual node-hours by node type, then compare with the 528 jobs and 2,112 channels stated in the paper. If the true count is 2,112, recompute Step 08 at 5,525.6 node-hr and the total per-field cost as roughly A$15,800 including storage; if the true count is 528, identify the channel-grouping or binning mechanism and add it to the methods. Also verify the elapsed-time and node-hour columns for Steps 01 and 06 using job start and end timestamps, so that the concurrency implied by each row is made explicit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 4 is the sole quantitative basis for the A$4,720 processing cost and the A$5,250 per-field extrapolation, but its job counts and node-hour totals do not follow from the text. Section 6.1 states that the production runs used the full native resolution of 2,112 channels of width 1.157 kHz, and Section 5 says each spectral channel is imaged separately. Row 08 of Table 4 lists only 528 imaging jobs at 9,419 s each, which gives 1,381.4 node-hr. If the true job count is 2,112, the imaging node-hours become 5,525.6 and the imaging cost alone rises from about A$3,522 to about A$14,090, raising the per-field estimate from roughly A$5,250 to roughly A$15,800 with all other rows unchanged. If the 528 jobs each process four channels, the paper needs to state that grouping explicitly; it does not. The cost rule is also applied unevenly: Steps 04 and 05 are halved for dual-socket nodes, while Step 08 is not (though socket pinning may justify this). More concerning, Step 01 reports 108 jobs x 2,443 s = 73.3 node-hr against an elapsed time of 6.2 hr, implying large concurrency, while Step 06 reports 36 jobs x 256,283 s = 2,562.8 node-hr against an elapsed time of 38.9 hr, which is only consistent if 36 jobs each lasted 71 hr. None of this undermines the feasibility demonstration, which rests on four actually produced cubes, but the resource estimate is not reproducible from the paper as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a trial of commercial high-performance computing (HPC) for radio-astronomy imaging, processing ASKAP GASKAP-HI pilot survey data on DUG Technology's commercial supercomputer. The authors ported an existing WSClean-based joint-deconvolution workflow, optimized it for the commercial platform through machine- and software-level changes, and produced four image cubes for the science team. They report a resource estimate of approximately A$5,250 per field and an elapsed time of 11 days per field for future full-survey processing, based on timing data from SBID 38466. The paper also documents lessons learned for researchers considering commercial HPC.","tokens_in":84,"tokens_out":4208,"duration_ms":99186,"significance":"If the resource estimate is reliable, the paper demonstrates that commercial HPC is a practical and affordable option for large-scale radio-astronomy processing, complementing public facilities and informing plans for SKA regional centres. The feasibility claim is well supported by four successfully produced image cubes, and the paper provides openly available code and an imaging guide. The main weakness is that the quantitative resource estimate, a stated sub-objective, rests on Table 4, whose job counts and node-hour accounting are internally inconsistent and not reproducible from the text; correcting this is essential before the cost estimate can be used by the community.","major_comments":[{"comment":"The paper states in §6.1 that production runs used the full native resolution of 2112 channels of width 1.157 kHz, and §5 says each spectral channel is imaged separately. However, Table 4 row 08 lists only 528 imaging jobs at 9,419 s each, yielding 1,381.4 node-hr. If 2112 channels are imaged separately, the job count should be 2112, which would raise the imaging node-hours to about 5,526 and increase the per-field cost from roughly A$5,250 to roughly A$15,800, assuming all other rows are unchanged. If instead each of the 528 jobs processes four channels, that grouping must be stated explicitly and reconciled with the text's claim of separate per-channel imaging. This inconsistency directly affects the headline resource estimate and must be resolved.","section":"§6.1 and Table 4"},{"comment":"The node-hour arithmetic in Table 4 is not reproducible without additional information about concurrency and nodes per job. The text says node-hours are the product of time per job and number of jobs, divided by two for dual-socket nodes without socket pinning, but Step 06 lists 36 jobs at 256,283 s each (2,562.8 node-hr) against an elapsed time of 38.9 hr, which implies each job ran about 71 hr unless jobs used multiple nodes or concurrent node allocations; neither case is specified. Similarly, Step 08 lists 528 jobs at 9,419 s each (1,381.4 node-hr) with 97.6 hr elapsed, and Step 01 lists 108 jobs at 2,443 s (73.3 node-hr) with 6.2 hr elapsed, implying very different concurrency levels that are not explained. Provide, for each step, the number of nodes per job, the wall-clock time per job, and the concurrency, so that the node-hour totals and the A$4,720 cost can be verified and the discrepancy between 2112 and 528 imaging jobs can be resolved.","section":"Table 4 (node-hour accounting)"},{"comment":"The per-field estimate of A$5,250 includes A$530 of storage cost, which is derived by assuming that 'multiple workflows can be managed at 3 per month' against a peak storage of ~25 TB per cube. This assumption is stated but not justified; if fields are processed one at a time, the storage cost per field would be approximately A$1,600, raising the total to about A$6,320. The sensitivity of the estimate to the field-processing cadence should be quantified or the assumption should be relaxed, as it is a nontrivial component of the headline figure.","section":"§7 (storage-cost assumption)"}],"minor_comments":[{"comment":"The phrase 'processing and storages costs' contains a grammatical error ('storages' should be 'storage').","section":"Abstract"},{"comment":"There is a typo in 'because of the the additional processing capacity available' — delete the duplicated 'the'.","section":"§6.1"},{"comment":"The word 'esentially' should be 'essentially'.","section":"§3"},{"comment":"The lesson numbering skips items 16 and 18; renumber for consistency.","section":"Table 5"},{"comment":"The sentence 'the sidelobe structure becomes more apparent as the weighting transitions from uniform (−2.0) to natural (2.0)' is clear, but the sign convention for Briggs robustness (negative = uniform, positive = natural) is not defined; add a brief definition for readers unfamiliar with WSClean.","section":"§6.2"},{"comment":"The table lists lessons 1–20 but the body text discusses only some of them; consider cross-referencing each lesson number or adding a sentence noting that unlisted lessons were considered lower priority.","section":"§8 and Table 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a practical experience report rather than a theoretical contribution, and the feasibility demonstration is credible. However, the resource estimate, one of the four stated sub-objectives, is not currently reproducible from the presented data. The inconsistency between the 2112 native-resolution channels and the 528 imaging jobs in Table 4 is particularly serious because it could change the headline cost by a factor of three. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I also note that the paper's self-citation to its own imaging guide and code is appropriate given the nature of the work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper deserves a serious referee. It does exactly what it claims: it ports the GASKAP-HI imaging pipeline to a commercial supercomputer, produces four science-ready cubes, and reports measured cost and elapsed time. The four-step porting process and the documented optimizations—especially the factor of 4.5 reduction in imaging time—are genuinely useful for anyone planning commercial HPC for SKA regional centres or other large surveys.\n\nWhat's new here is concrete: real node-hour totals, dollar costs, a before/after optimization table, and a frank lessons-learned list. The paper also shows that the joint deconvolution outputs from WSClean at DUG match earlier processing for at least one field, which gives confidence in the feasibility claim. That claim does not rest on a simulation or an extrapolation; it rests on four actually produced cubes.\n\nThe soft spot is the cost model in Table 4. The text says the full 2112 native-resolution channels are imaged separately, but the table lists only 528 imaging jobs at 9,419 s each. Multiply those together and you get 1381 node-hr, which is the number in the table. So the likely explanation is that each imaging job covers four channels—consistent with the measured 39 minutes per channel. But the paper never says that, and a reader who assumes one job per channel will think the imaging cost is understated by a factor of four. That ambiguity needs to be fixed. Similarly, the 'divide by two' rule for dual-socket nodes is applied to some rows and not others; the text explains that socket pinning changes the accounting, but Table 4 alone is confusing. Neither issue invalidates the feasibility demonstration, but the resource estimate, which is a stated sub-objective, is not reproducible from the paper as written.\n\nOne more minor point: the storage-cost extrapolation assumes three workflows can run per month. That's a reasonable planning assumption, but it should be flagged as an assumption rather than a measured constraint.\n\nWho is this for? Astronomers thinking about using commercial HPC, and the SKA regional centre community. It's an experience report with measurements, not a theoretical advance, and it's honest about the limitations. I'd accept it for peer review. The revision should focus on cleaning up Table 4's job-count documentation and adding one sentence about the four-channels-per-job grouping.","headline":"A solid, practical trial of commercial HPC for radio imaging; the feasibility claim holds, but the cost model needs a clarifying edit in Table 4.","tokens_in":18234,"tokens_out":3628,"would_cite":true,"duration_ms":29235,"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":"Commercial supercomputer processes radio survey for A$5,250 per field","keywords":["radio astronomy","high performance computing","commercial supercomputing","GASKAP-HI survey","neutral hydrogen imaging","joint deconvolution","workflow porting","cost estimation"],"falsifier":"Count the actual number of WSClean imaging jobs in the optimized production run: if it is 2112 rather than 528, the imaging node-hours and the per-field cost would be roughly four times larger, invalidating the A$5,250 estimate.","tokens_in":17092,"feed_emoji":"📡","tokens_out":8255,"duration_ms":68061,"temperature":0.7,"pith_summary":"The paper claims that a commercial supercomputer can run a full-scale radio astronomy imaging workload—processing GASKAP-HI pilot survey data with 108 beams and 2112 spectral channels—and that the optimized workflow is fast and cheap enough for the full survey: about A$5,250 and 11 days per field. The demonstration is actual production: four science-ready image cubes were produced and delivered to the survey team, not just benchmark timings. This matters because it gives astronomers an on-demand alternative to waiting in queues at public supercomputing facilities, with a documented recipe for porting an existing astronomical pipeline to a rented platform. The key to the result is optimization: machine-level tuning (socket pinning, RAM-disk scratch space, concurrent imaging instances) and software-level tuning (clean thresholds, Briggs weighting, disabling model updates) together cut the per-channel imaging time by a factor of 4.5.","feed_headline":"Commercial supercomputer processes radio survey for A$5,250 per field","feed_subtitle":"Porting an astronomy imaging workflow to rented nodes cut imaging time 4.5x; full survey fits on-demand compute.","key_machinery":"The load-bearing mechanism is the joint deconvolution imaging mode of the WSClean software, which stitches the 108 individual beams into one image before applying CLEAN, preserving the large-scale diffuse emission that would be lost if each beam were deconvolved separately. The argument hinges on the measured performance of this software on the commercial provider's heterogeneous node pool, combined with a set of optimizations: assigning memory-hungry tasks to high-memory nodes, pinning processes to sockets, using a RAM disk for temporary files, running four imaging instances per socket, and suppressing file-system writes by disabling model-column updates. These choices, together with a drip-fed data transfer from the archive, produce the reported cost and elapsed-time figures.","core_discovery":"The central claim is that renting commercial high-performance computing is a viable way to produce science-ready image cubes for a large H I survey, at a predictable dollar cost and elapsed time. The paper substantiates this by porting a previously developed joint-deconvolution imaging workflow to the commercial platform, optimizing it, and measuring the production cost on one pilot field: roughly A$4,720 in compute plus A$530 in storage, over 11 days, for a cube spanning 2112 channels at 7-arcsecond resolution across 108 beams. The paper further claims that the optimization gains are substantial and repeatable—imaging time per channel fell from 2.9 hours to 39 minutes—and that the resulting cubes serve both as reference images for an upgrade to the telescope's science data processor and as science products for the survey team.","pith_inferences":["If the imaging job count discrepancy is real, the per-field cost may be understated by up to a factor of four; the full-survey budget should not be set until the 528-versus-2112 question is resolved.","The optimization techniques (socket pinning, RAM-disk scratch, concurrency tuning) are generic enough to transfer to other radio imaging workloads on shared-memory clusters, whether commercial or academic.","The cost model is specific to the GASKAP-HI geometry; scaling to other surveys with different beam counts, channel counts, or weighting schemes would require re-measuring per-channel imaging time rather than reusing A$5,250 directly."],"forward_implications":["If the estimate holds, the full GASKAP-HI survey can be processed on demand at a known cost and an 11-day cadence per field, allowing budget planning without queue-time uncertainty.","The documented porting and optimization steps give other radio surveys a template for moving existing pipelines to commercial HPC, including the future regional centres for the SKA.","The WSClean joint-deconvolution cubes provide independent reference products for validating and refining the ASKAPSoft joint-deconvolution upgrade.","The observed factor-of-4.5 speedup suggests similar tuning could reduce turnaround times on public supercomputers even when compute is not billed."],"supporting_citations":[{"why":"Provides the WSClean software whose joint deconvolution mode is the core imaging engine for the workflow.","marker":"Offringa et al., 2014"},{"why":"Defines the original GASKAP-HI imaging workflow that this paper ports to the commercial platform.","marker":"Pingel et al., 2022"},{"why":"Supplies the CASA package used for data manipulation, concatenation, and feathering steps.","marker":"CASA Team et al., 2022"},{"why":"Demonstrates prior processing of simulated SKA data on a top supercomputer, motivating the commercial-HPC trial.","marker":"Wang et al., 2020"},{"why":"Defines the GASKAP-HI survey parameters that set the scale for the full-survey cost estimate.","marker":"Dickey et al., 2013"}],"fun_headline_variants":["Rented HPC images GASKAP-HI field in 11 days for A$5,250","Commercial supercomputer cuts imaging time 4.5x for radio survey","Astronomy data ported to cloud: GASKAP-HI imaged at low cost","On-demand supercomputing delivers science-ready radio cubes","GASKAP-HI pilot survey processed on rented supercomputer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The per-field cost and elapsed-time estimate assumes that the single benchmark field (SBID 38466) is representative of every full-survey field, and that the node-hour counts in the cost table are internally consistent—including the unexplained difference between the stated 2112 native-resolution channels and the 528 imaging jobs listed.","fun_headline_variants_meta":{"raw":{"variants":["Rented HPC images GASKAP-HI field in 11 days for A$5,250","Commercial supercomputer cuts imaging time 4.5x for radio survey","Astronomy data ported to cloud: GASKAP-HI imaged at low cost","On-demand supercomputing delivers science-ready radio cubes","GASKAP-HI pilot survey processed on rented supercomputer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000494,"raw_usage":{"total_tokens":2464,"prompt_tokens":1021,"completion_tokens":1443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":1342}},"tokens_in":637,"tokens_out":1443,"duration_ms":12198,"temperature":1.0,"reasoning_tokens":1342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:29:58.851760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the actual number of WSClean imaging jobs in the optimized production run: if it is 2112 rather than 528, the imaging node-hours and the per-field cost would be roughly four times larger, invalidating the A$5,250 estimate.","supporting_citations":[],"review_version":1}