{"id":"126ec2d9-4901-4997-8436-d37f0f716994","arxiv_id":"2501.15504","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A literature survey that classifies geo-distributed task scheduling research into four environments and three objectives, without providing new algorithms or results.","lead":"This survey organizes more than 130 research papers on scheduling computing tasks across data centers, edge devices, and supercomputers spread around the world. It groups approaches by computing environment and by goal (performance, fairness, fault tolerance) to help researchers find the main techniques and open problems.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HPC pillar of the four-environment taxonomy mixes generic HPC/grid schedulers into the 'geo-distributed supercomputer' category, so the survey's distinctness and comprehensiveness claims rely on an unverified scope filter.","rationale":"The reader's weakest assumption, unverified corpus comprehensiveness, is real and remains valid: no search queries, databases, or inclusion criteria are given, so completeness cannot be audited. My stress-test found a more specific, internally checkable version of the same risk: the HPC section, the survey's differentiator, appears to contain multiple references that do not satisfy the paper's own geo-distributed definition. This is not an external-consensus disagreement; it is an internal scope and classification inconsistency that can be settled by reading the cited papers. If the proposed check shows most Section 5 refs are in fact geo-distributed, the concern disappears and the main remaining issue is the missing methodology. The survey's organization and breadth are genuine strengths, and the classification of the cloud and edge sections is mostly clear, so a rejection is not warranted. Conditional acceptance with a required scope-definition and methodology revision is the appropriate verdict; this matches the reader's conditional verdict, so no change to the verdict is needed.","tokens_in":31818,"tokens_out":6491,"duration_ms":58611,"concrete_test":"For each of the 13 references in Section 5 (refs [123] to [135]), read the full text and determine whether the proposed scheduler's model or evaluation includes more than one geographically separated site (e.g., multiple data centers, regions, or HPC centers connected by WAN links) with explicit inter-site network cost or latency. Count the refs that fail this test. If 3 or more of the 13 fail, the GDSC section needs an explicit geo-distribution scope filter, reclassification of out-of-scope refs, and a softened 'comprehensive' claim; if all 13 pass, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central claim is a comprehensive map of task scheduling across four distinct geo-distributed systems, and Section 5 (Geo-Distributed Supercomputer Computing, HPC) is the pillar the authors use to distinguish this survey from prior work (Section 1). Yet Section 5 opens by defining GDSC as coordinating tasks across globally distributed HPC nodes, without giving inclusion criteria that separate GDSC from ordinary HPC or grid scheduling. Several references in Section 5 appear, from their titles and venues, to be conventional single-cluster or non-geo-distributed HPC schedulers: [125] intra-and-inter-server scheduling in HPC data centers, [126] MRSch multi-resource scheduling for HPC, [128] hierarchical task allocation for energy-aware HPC clouds, [130] HPC grid meta-heuristics, [131] HPC task scheduling on IaaS cloud infrastructures, [133] malleable job scheduling for supercomputers, and [134]/[135] rollback-recovery in HPC systems. If these works do not model multiple geographically separated sites or WAN costs, they are misclassified under the paper's own definition, and the 'four distinct systems' taxonomy plus the 'comprehensive' claim for the HPC environment is substantially overstated. The absence of a documented search and inclusion protocol, noted by the reader, compounds this: without stated scope criteria, readers cannot tell whether Section 5 is a review of geo-distributed HPC or a convenience sample of HPC scheduling papers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper surveys task scheduling in geo-distributed computing, organizing 137 references into four environments (geo-distributed cloud, edge, cloud-edge, and geo-distributed supercomputer/HPC) and classifying each work by scheduling objective (performance, fairness, fault-tolerance) and by method family (heuristic, AI-based, mathematical, hybrid). The authors argue in Section 1 that earlier surveys are limited to one environment or to cloud/grid combinations and that this is the first comprehensive review to cover all four environments, with the HPC pillar being a key differentiator. They present a taxonomy for each environment, summarize representative algorithms, and close with challenges and open issues in Section 6 and a conclusion in Section 7.","tokens_in":32096,"tokens_out":4328,"duration_ms":38452,"significance":"If the taxonomy and coverage are accurate, the survey would be a genuinely useful reference: it applies a consistent objective-based organizing principle across four paradigms, it spans recent works up to 2024 from reputable venues, and it explicitly addresses fairness and fault-tolerance in every environment, which many prior surveys treat only in passing. The paper also identifies meaningful open issues such as emerging AI/LLM workloads and next-generation hardware. However, the two central claims—comprehensiveness of the corpus and distinctness of the four-environment taxonomy—are currently not verifiable because the survey omits its literature selection methodology, and the HPC pillar contains entries that appear to be conventional HPC/grid schedulers without explicit geo-distribution. Since these claims define the survey's value, they need to be substantiated or carefully scoped before the paper can be recommended for acceptance.","major_comments":[{"comment":"The survey claims in Section 7 to provide 'a comprehensive review of state-of-the-art task scheduling techniques across four distinct geo-distributed computing systems,' but Section 1 gives no search queries, databases, date range, or inclusion/exclusion criteria. As a result, the 137-reference corpus cannot be verified as comprehensive or representative; it is a hand-selected sample. Please add a methodology subsection documenting the literature search and screening process, or revise the 'comprehensive' claims to be explicitly scoped to the works surveyed.","section":"Sections 1 and 7"},{"comment":"Section 5 defines geo-distributed supercomputer computing (GDSC) as coordinating tasks across globally distributed HPC nodes, but it does not specify how cited works qualify as geo-distributed. Several entries appear to target conventional HPC or grid environments without modeling multiple geographically separated sites or WAN costs: reference [125] on intra- and inter-server scheduling in HPC data centers, [126] MRSch for multi-resource HPC scheduling, [128] hierarchical task allocation for energy-aware HPC clouds, [130] parallel job scheduling in HPC grids with nature-inspired meta-heuristics, [131] HPC task scheduling on IaaS cloud infrastructures, [133] malleable job scheduling for supercomputers, and [134]/[135] rollback-recovery in HPC systems. If these works are not geo-distributed under the paper's own definition, they are misclassified, and the claimed distinctness of the HPC pillar (Section 1) is substantially overstated. Please add per-reference justification for geo-distribution, or narrow the section's title and claims accordingly.","section":"Section 5"},{"comment":"There is an internal contradiction in the taxonomy: Figure 2 classifies references [38] and [39] under 'MILP with Branch and Cut,' but Section 2.1.1 (under the heading 'MILP with Branch and Bound') states that OPRS [39] employs a branch-and-bound algorithm. Please correct the figure or the text and verify the solver classification for both references, because this inconsistency undermines confidence in the accuracy of the taxonomy.","section":"Figure 2 and Section 2.1.1"}],"minor_comments":[{"comment":"The heading 'Mixed Linear Integer Programming' should read 'Mixed Integer Linear Programming' to match standard terminology and the MILP abbreviation.","section":"Section 2.1.1"},{"comment":"In the Data Placement Optimization paragraph, 'Data placement problem ca also be solved using reinforcement learning' contains a typo: 'ca' should be 'can.'","section":"Section 2.1.2"},{"comment":"In the Divide-and-Conquer paragraph, 'breath-first search' should be 'breadth-first search.'","section":"Section 3.1.1"},{"comment":"The text attributes algorithms named 'Okita' and 'Okita*' to reference [98], whose listed authors are Pang et al.; please confirm that these algorithm names are correct and appear in the cited paper, and clarify the naming for readers.","section":"Section 3.1.2"},{"comment":"The entry 'Extreme Low' latency for GDSC is counterintuitive for geographically distributed supercomputers, where WAN communication typically introduces high latency; please clarify whether this refers to user-perceived response latency for certain workloads or to inter-node communication latency, and consider rewording.","section":"Table 1"},{"comment":"References [136] and [137] are YouTube videos; consider replacing them with archival citations (papers, technical reports, or official documentation) for stability and verifiability.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper has the makings of a useful survey, but the missing literature-selection protocol and the questionable inclusion of non-geo-distributed HPC papers in Section 5 are load-bearing for the 'comprehensive' and 'four distinct systems' claims. The internal inconsistency around reference [39] is a concrete sign that the taxonomy needs a careful fact-check. I recommend major revision rather than rejection, because these issues are fixable by adding a methodology section, tightening the HPC scope, and correcting the taxonomy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a useful but uneven survey. The four-environment taxonomy (cloud, edge, cloud-edge, HPC) with a per-environment split by objective is genuinely new in structure – earlier surveys stuck to one environment or lumped cloud and grid. The paper earns credit for that organization and for the generally accurate, readable one-paragraph summaries. If I were a grad student starting in geo-distributed scheduling, I'd want this on my desk.\n\nBut there are soft spots, and one is load-bearing. Section 5, the HPC pillar that the authors use to distinguish their survey from prior work, includes several references that are not geo-distributed by the paper's own definition. From their titles and venues, [125], [126], [128], [130], [131], [133], [134], and [135] look like ordinary single-cluster or cloud schedulers, not work coordinating tasks across globally distributed HPC nodes. The section opens with a definition of GDSC that requires multiple geographically separated sites and WAN considerations, but then the corpus doesn't consistently meet that bar. That means the 'four distinct systems' taxonomy – the main contribution – is overstated for at least one of the four.\n\nThe other big gap is methodology. The paper claims to be comprehensive but never states which databases, search queries, date ranges, or inclusion criteria were used. For a survey whose value rests on coverage, that's not a minor omission; it leaves the reader unable to tell whether the HPC section is a genuine review or a convenience sample. The stress-test note hits this exactly.\n\nThere are also smaller inconsistencies: reference [39] appears as 'MILP with Branch and Cut' in Figure 2 but 'branch-and-bound' in the text, and the challenges section leans on a couple of YouTube videos as evidence of emerging workloads – not a deal-breaker, but not scholarly.\n\nIs the central argument sound? The taxonomy itself is plausible as a way to organize the space, and the summaries are honest to the sources. The problem is execution, not conception. I'd send this to peer review, but with major revisions expected: the authors need to add a selection methodology, re-classify or justify the HPC references, and fix the internal inconsistencies. After that, it would be a serviceable survey. For now, I'd be cautious about citing it as authoritative.","headline":"Plausible taxonomy and broad coverage, but the HPC pillar inflates the corpus with non-geo-distributed schedulers and the survey lacks a documented selection method, so 'comprehensive' is doing too much work.","tokens_in":32549,"tokens_out":3798,"would_cite":false,"duration_ms":34271,"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 survey contends that task scheduling across geo-distributed cloud, edge, cloud-edge, and supercomputer systems can be usefully classified by a two-axis taxonomy of environment and objective, and that this map exposes where research…","keywords":["Geo-Distributed","Task scheduling","Workflow scheduling","Optimization","Edge computing","Cloud-edge computing","HPC scheduling","Scheduling objectives"],"falsifier":"A reader could test the comprehensiveness claim by running a defined literature search with explicit queries and a date range on the topic and checking how many of the retrieved papers are absent from the survey's reference list and cannot be assigned to one of its four environments; a substantial number of such omissions would falsify the claim.","tokens_in":31660,"feed_emoji":"🗺️","tokens_out":7460,"duration_ms":64697,"temperature":0.7,"pith_summary":"This paper is a survey that aims to give researchers a complete map of task scheduling in geo-distributed computing, covering four settings: cloud data centers spread across regions, edge networks, cloud-edge collaborations, and geographically distributed supercomputers. The authors' central claim is that every existing scheduling method in these settings can be classified by environment and by the objective it targets—performance, fairness, or fault tolerance—and that doing so reveals where the field is concentrated and where it is thin. If the map is right, a researcher entering the area can quickly locate the relevant technique family for their problem and see which objectives and environments remain underserved. The survey also argues that previous reviews missed the HPC and supercomputer case, and that emerging workloads and new hardware are the key open challenges.","feed_headline":"Survey maps task scheduling across four geo-distributed systems","feed_subtitle":"One taxonomy now covers cloud, edge, cloud-edge, and HPC scheduling methods, exposing the gaps.","key_machinery":"The load-bearing mechanism is the two-axis taxonomy: environment (geo-distributed cloud, edge, cloud-edge, geo-distributed supercomputer) crossed with scheduling objective (performance, fairness, fault tolerance). Under performance, the taxonomy splits into computing-resource utilization and data management, and each cell is further filled by method type—heuristic, AI-based, mathematical, or hybrid. This grid is what allows the survey to place each surveyed article in a unique cell and to read the distribution of work as evidence about where the field is concentrated and where it is empty.","core_discovery":"The authors' central claim is that the literature on geo-distributed task scheduling, which earlier reviews treated environment by environment, can be read as one coherent design space. They organize that space along two axes: four computing infrastructures—geo-distributed cloud, edge, cloud-edge, and geo-distributed supercomputer (HPC)—and three scheduling objectives—performance, fairness, and fault tolerance. Within performance, they further separate work that optimizes the use of computing resources from work that optimizes data management, and they classify every surveyed method as heuristic, AI-based, mathematical, or hybrid. On this basis they conclude that no previous survey covered all four environments, that supercomputer scheduling in particular had been left out, and that the resulting map exposes the field's gaps, with fairness and fault tolerance far less developed than performance, along with the open problems posed by emerging AI and multimodal workloads, next-generation hardware, and cross-border security and privacy.","pith_inferences":["A natural next step the authors do not take is to use their own grid as a quantitative map, counting papers per cell and then weighing whether the sparse cells reflect genuine research gaps or simply differences in how problems are named across communities.","The taxonomy could be stress-tested on boundary systems such as serverless geo-distributed computing, satellite edge computing, and federated learning clusters, which may not fit cleanly into any single one of the four environment boxes.","Because the survey is qualitative, it does not say which method wins on a given objective; an editorially inferred consequence is that the field now needs a common benchmark with shared workloads, failure models, and cost metrics before the taxonomy can guide technique selection."],"forward_implications":["A researcher facing a scheduling problem can use the taxonomy to identify which environment and objective their task belongs to and immediately see the method families—heuristic, AI-based, mathematical, or hybrid—that have been tried there.","The survey's explicit comparison across all four environments supports the conclusion that performance-driven scheduling is far more developed than fairness and fault tolerance, marking the latter as the most promising space for new algorithms.","Naming geo-distributed supercomputing as a fourth environment widens the scheduling research agenda beyond latency-sensitive web and IoT workloads to tightly coupled, bandwidth-intensive scientific computing.","The open-issue analysis in Section 6 gives a concrete agenda: scheduling for LLM inference and multimodal AI workloads, for next-generation quantum and nano-computing hardware, and for security and privacy constraints that span jurisdictions."],"supporting_citations":[{"why":"A comprehensive survey of cloud scheduling techniques that the authors position as environment-specific, establishing the gap they aim to fill.","marker":"[1]"},{"why":"A state-of-the-art review of edge computing task scheduling that likewise covers only one environment, used to show the lack of cross-environment coverage.","marker":"[5]"},{"why":"A survey of resource scheduling in edge computing that spans one distributed environment, cited as evidence that prior reviews are not comprehensive across geo-distributed systems.","marker":"[7]"},{"why":"A review of energy-optimization task scheduling algorithms in cloud environments, another single-environment baseline the taxonomy extends.","marker":"[8]"},{"why":"A review of cost-optimization methods for scientific workflow scheduling in cloud and grid environments; the authors cite it as the closest prior cross-environment survey that still omits supercomputing.","marker":"[9]"}],"fun_headline_variants":["Four geo-distributed systems, one scheduling taxonomy","Survey maps scheduling across cloud, edge, cloud-edge, HPC","Survey exposes gaps in geo-distributed scheduling","Geo-distributed task scheduling: one map for four realms","New taxonomy unifies geo-distributed scheduling from cloud to HPC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey assumes its selected set of papers is a comprehensive and representative sample of geo-distributed task scheduling research, because it does not report a reproducible search protocol or inclusion criteria.","fun_headline_variants_meta":{"raw":{"variants":["Four geo-distributed systems, one scheduling taxonomy","Survey maps scheduling across cloud, edge, cloud-edge, HPC","Survey exposes gaps in geo-distributed scheduling","Geo-distributed task scheduling: one map for four realms","New taxonomy unifies geo-distributed scheduling from cloud to HPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001079,"raw_usage":{"total_tokens":4489,"prompt_tokens":898,"completion_tokens":3591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":3511}},"tokens_in":514,"tokens_out":3591,"duration_ms":24134,"temperature":1.0,"reasoning_tokens":3511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:12:25.855103+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could test the comprehensiveness claim by running a defined literature search with explicit queries and a date range on the topic and checking how many of the retrieved papers are absent from the survey's reference list and cannot be assigned to one of its four environments; a substantial number of such omissions would falsify the claim.","supporting_citations":[],"review_version":1}