{"id":"804e9f39-4e7f-49d0-8f6e-54075b48d7e8","arxiv_id":"2607.04415","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GORIO extends GPU-centered page-cache I/O to NVMe-oF and keeps ANNS miss generation, pending state, and resume decisions on the GPU, yielding 1.31×–4.89× speedups on SIFT1M.","lead":"GORIO keeps graph ANNS search ownership on the GPU while using the CPU only as a thin NVMe-oF transport proxy for remote page service. It reports 1.31× and 4.89× speedups over remote-I/O and direct remote page-cache baselines on SIFT1M, showing a practical path for GPU-centered disaggregated vector search.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"SIFT1M single-run times do not stress the remote-miss regime the paper motivates, so the 1.31×/4.89× substrate claims rest on an unrepresentative working set.","rationale":"The reader correctly isolates the evaluation-scale gap as the weakest assumption behind the strongest claim. The design (GPU-owned miss/pending/resume + thin SPDK proxy) is coherent and the ablations cleanly separate backend vs scheduling, so there is no internal contradiction that would force REJECT. The concern is not “outside consensus” but whether the reported numbers actually exercise the remote path the motivation requires. Expanding to a capacity-exceeding index with miss/hit/outstanding metrics, plus variance, is exactly the check that would settle it; until then CONDITIONAL remains the right verdict and confidence should stay moderate. No stronger load-bearing flaw (e.g., incorrect ownership split or non-reproducible methodology) is visible in the manuscript.","tokens_in":11084,"tokens_out":647,"duration_ms":7620,"concrete_test":"Re-run the exact Table 2 configurations on a DiskANN-style index whose resident set substantially exceeds the 48 GB L40S (e.g., SIFT100M or a synthetic graph with working-set >> GPU memory), reporting per-query remote miss rate, GPU page-cache hit rate, and average outstanding NVMe-oF depth. If G-Gust vs GustANN and G-ABaM vs BaM speedups fall below ~1.1× / ~2× or become statistically indistinguishable under multi-run variance, the substrate claim for the intended disaggregated setting does not hold on the paper’s own evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that GORIO is a concrete GPU-centered remote I/O substrate for graph ANNS over NVMe-oF, evidenced by G-Gust 1.31× over GustANN and G-ABaM 4.89× over BaM on SIFT1M (Table 2, §5.2). That claim is load-bearing on the assumption that SIFT1M (1M vectors, 10k queries, one L40S, one InfiniBand path, one enterprise SSD) produces the fine-grained, data-dependent remote page-miss pressure that §1–2 and Motivation 1–2 say makes CPU-centered NVMe-oF a mismatch. On a small index the GPU page cache can absorb most of the working set after warm-up, so measured gains may largely reflect local scheduling/proxy overhead rather than fabric-crossing miss handling under capacity pressure. The paper itself flags this in §7.2 (“larger indexes and additional datasets are needed… higher graph depth, larger working sets”), yet the abstract, contributions, and Result Takeaway still present the SIFT1M numbers as demonstrating the remote substrate. Without miss-rate, cache-hit, or outstanding-I/O depth numbers, it is unclear whether the evaluation ever left the regime where remote I/O is secondary.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"GORIO is a systems paper that extends GPU-centered local page-cache I/O (in the BaM style) to remote NVMe-oF for graph-based ANNS. The design keeps query evolution, page-miss generation, pending state, and resume decisions on the GPU, while the CPU is restricted to an SPDK/NVMe-oF transport and completion proxy that publishes ready state into GPU-visible memory. Two layers are claimed: a GPU-direct remote path that turns page-cache misses into split-phase remote operations (descriptors, ready table, cache fill), and ANNS-specific persistent GPU scheduling that yields blocked queries and resumes them when pages arrive. On a SIFT1M DiskANN-style workload (10k queries, recall@10) over RDMA NVMe-oF, the paper reports G-Gust 1.31× faster than GustANN and G-ABaM 4.89× faster than a BaM-style remote path, with large gains over a GDS baseline (Table 2, §5.2).","tokens_in":11457,"tokens_out":1531,"duration_ms":14448,"significance":"If the architectural split and speedups hold under realistic remote-miss pressure, the paper would be a useful contribution to GPU-centric storage and disaggregated vector search: it cleanly separates GPU ownership of irregular graph-search state from host-managed NVMe-oF progress, and the ablation structure (G-Gust vs GustANN for the remote backend; G-ABaM vs BaM for ANNS scheduling) is a sensible way to isolate those layers. The work is concrete rather than purely conceptual—it is implemented with CUDA/SPDK over real RDMA NVMe-oF and reports end-to-end search times against named baselines. That said, significance for the motivated large-index, capacity-disaggregated setting is currently limited by evaluation scale; the paper itself notes the need for larger indexes and datasets (§7.2).","major_comments":[{"comment":"Table 2 / §5.2 and abstract/Result Takeaway: the central claim that GORIO demonstrates a GPU-centered remote I/O substrate for graph ANNS rests on SIFT1M (1M vectors, 10k queries, one L40S, one InfiniBand path, one enterprise SSD). §1–2 and Motivations 1–2 motivate the work by large indexes that exceed GPU memory and by fine-grained, data-dependent remote page misses. On a small index the GPU page cache can absorb much of the working set after warm-up, so the 1.31× and 4.89× gains may largely reflect local scheduling/proxy overhead rather than fabric-crossing miss handling under capacity pressure. §7.2 itself states that larger indexes and additional datasets are needed for higher graph depth and larger working sets, yet the abstract, contributions, and takeaway still present SIFT1M as demonstrating the remote substrate. Without miss-rate, cache-hit, outstanding-I/O depth, or working-set","section":null},{"comment":"§5.1–5.2: all end-to-end times appear to be single-run application search times with no variance, repeated trials, or sensitivity to free parameters the design depends on (CPU proxy batching and outstanding depth, GPU page-cache size/slot policy, scheduler yield/claim policy). For a systems result whose headline numbers are 1.31× and 4.89×, single-point timings are weak evidence of a stable substrate advantage. Please report multi-run means/std or ranges, and at least a brief sensitivity study on cache size and outstanding remote I/O depth so readers can see whether the gains are robust or tuned to one configuration.","section":null},{"comment":"§5.1 Comparison and Table 2: the paper asserts “same recall” in the conclusion and uses recall@10 as the evaluation setting, but Table 2 reports only time/QPS/speedup with no Recall@10 column or confirmation that all five paths achieve the same recall on the same index. If any path trades recall for latency (e.g., via different pending/resume or page-service behavior), the throughput comparison is not apples-to-apples. Please report Recall@10 for every system in Table 2 and state the common search parameters (ef/beam, graph degree, page size) used across paths.","section":null}],"minor_comments":[{"comment":"Figure 1 is described in §3.1 but the manuscript text does not include quantitative labels (e.g., where descriptors live, ready-bit protocol). A short caption note on the ready-table vs. request-queue handoff would help readers who only skim the figure.","section":null},{"comment":"§4 Implementation: “GustANN-derived” and “BaM-based” are clear at a high level, but the text does not state which BaM/GustANN components were reused vs. rewritten (page cache, scheduler, SPDK path). A short reuse table would improve reproducibility claims.","section":null},{"comment":"Table 1 lists CUDA 13.0 and SPDK 25.09; if these are intentional, fine, but double-check version strings for the camera-ready so readers can match the stack.","section":null},{"comment":"Related work (§6) is appropriate but could more explicitly contrast GORIO with concurrent GPU-storage ANNS systems that appear in the reference list (e.g., FlashANNS, Neos) on the remote vs. local axis.","section":null},{"comment":"Minor wording: abstract and §1 use both “GORIO” and spaced forms inconsistently in places; unify the product name and the G-Gust / G-ABaM labels early in §5.","section":null}],"recommendation":"major_revision","confidential_remarks":"The architectural story is coherent and the ablation design is better than many systems papers. The main risk is overclaim relative to a SIFT1M-only evaluation that the authors themselves flag as insufficient in §7.2. I would not reject on novelty grounds if the authors either scale the evaluation or tightly qualify the claims; major revision is the right bar. Fit for a systems/DC venue is reasonable if the remote-miss regime is actually demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is a working extension of BaM-style GPU page-cache I/O across real RDMA NVMe-oF, specialized so graph ANNS keeps miss generation, pending state, and resume on the GPU while the CPU is only an SPDK submit/poll/completion proxy. That ownership split is the actual contribution, not a new ANNS algorithm.\n\nWhat is new and done well is the concrete substrate. They turn GPU page-cache misses into split-phase remote ops (descriptor → SPDK NVMe-oF → GPU-visible ready table), pair that with a persistent scheduler that yields blocked queries, and ablate cleanly: G-Gust vs GustANN isolates the remote backend (1.31×), G-ABaM vs BaM isolates the ANNS scheduling layer (4.89×), both against GDS as a weak page-read baseline. Architecture is coherent, citations to BaM, DiskANN, GustANN, and async GPU-storage work are in the right places, and there is no circular math—just measured end-to-end times on a real InfiniBand path. For people building GPU-resident vector search under storage disaggregation, this is useful engineering.\n\nSoft spots are real but proportionate. Evaluation is one SIFT1M DiskANN-style run (10k queries, one L40S, one enterprise SSD). The paper itself says larger indexes and deeper working sets are needed (§7.2), yet the abstract and takeaway still lean on these numbers as demonstrating the remote substrate. Without miss rates, cache-hit stats, or outstanding-I/O depth, it is fair to worry that after warm-up the cache absorbs most of the working set and the gains partly reflect local scheduling/proxy overhead rather than fabric-crossing pressure. Single-run times, no variance, no multi-GPU or multi-tenant data, no artifacts. Those are the limits of a technical report, not a load-bearing contradiction.\n\nThis is for systems people in GPU storage and vector DB infrastructure. It deserves a serious referee; expand scale, report sensitivity, and the claim hardens. I would engage with the design and cite the ownership split if I were working in the same stack.","headline":"Solid systems report with a coherent GPU-owned remote miss path; the 1.31×/4.89× numbers are real on SIFT1M but do not yet prove the large-index disaggregation story the intro sells.","tokens_in":12128,"tokens_out":558,"would_cite":true,"duration_ms":5829,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"GORIO keeps graph ANNS search ownership on the GPU over remote NVMe-oF, with the CPU only as a transport proxy, and reports 1.31× and 4.89× speedups on SIFT1M.","keywords":["ANNS","vector databases","GPU-centric storage","NVMe-oF","GPU-Direct","disaggregated storage","graph traversal","page cache"],"falsifier":"Re-run the same five paths on a multi-billion-vector DiskANN-style index with deeper traversal and a working set far larger than GPU memory; if GORIO’s 1.31× and 4.89× advantages disappear or reverse under that miss load, the substrate claim does not hold for the intended setting.","tokens_in":11911,"feed_emoji":"🔌","tokens_out":772,"duration_ms":6174,"temperature":0.7,"pith_summary":"Large graph indexes for approximate nearest-neighbor search often do not fit in one GPU’s memory, so production systems want storage disaggregated over NVMe-oF. Existing remote paths still let the CPU form I/O, drive the network, and decide when GPU work may resume—exactly the wrong place for control, because the next page miss is discovered inside GPU graph traversal. GORIO extends GPU-centered local page-cache I/O to remote storage: GPU code still owns query evolution, miss generation, pending state, and resume decisions, while the CPU is reduced to submitting NVMe-oF commands and publishing completion bits into GPU-visible memory. A second layer adds ANNS-specific persistent scheduling so blocked queries yield and other ready work continues while remote pages arrive. On a SIFT1M DiskANN-style workload over RDMA NVMe-oF, the system is 1.31× faster than a state-of-the-art remote reference path and 4.89× faster than a direct remote page-cache path that blocks on each miss. The claim is that this split of ownership is both workable and necessary for irregular GPU graph search over fabric storage.","feed_headline":"GPU keeps graph search; CPU only moves remote pages","feed_subtitle":"GORIO is 1.31× and 4.89× faster on SIFT1M by turning page misses into split-phase NVMe-oF work","key_machinery":"Split-phase remote page service: a GPU page-cache miss becomes a lock-free request descriptor, the blocked query yields, independent graph work continues, and a GPU-visible ready-table entry lets the persistent scheduler resume when the CPU proxy has filled the cache slot over NVMe-oF.","core_discovery":"GPU-centered local page-cache I/O can be extended across real NVMe-oF so that graph ANNS keeps query evolution, page-miss generation, pending state, and resume decisions on the GPU, with the CPU restricted to transport and completion proxying; on SIFT1M this substrate is 1.31× faster than the remote-I/O reference path and 4.89× faster than a direct remote page-cache path that waits in place for pages.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["GPU owns graph ANNS; CPU only proxies remote NVMe-oF pages","GORIO keeps ANNS traversal on GPU over split-phase NVMe-oF","Page misses stay GPU-centered across remote NVMe-oF fabric","Graph ANNS evolves on GPU; CPU is NVMe-oF transport proxy","GPU-centered remote I/O beats in-place page waits on SIFT1M"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The paper’s speedups rest on single-run SIFT1M search times over one small index and one RDMA NVMe-oF link, which may not reproduce the miss pressure and cache behavior of the large disaggregated indexes the design is meant for.","fun_headline_variants_meta":{"raw":{"variants":["GPU owns graph ANNS; CPU only proxies remote NVMe-oF pages","GORIO keeps ANNS traversal on GPU over split-phase NVMe-oF","Page misses stay GPU-centered across remote NVMe-oF fabric","Graph ANNS evolves on GPU; CPU is NVMe-oF transport proxy","GPU-centered remote I/O beats in-place page waits on SIFT1M"]},"model":"grok-4.5","effort":"low","cost_usd":0.005996,"raw_usage":{"total_tokens":1658,"prompt_tokens":892,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":59960000,"prompt_tokens_details":{"text_tokens":892,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":675,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":892,"tokens_out":91,"duration_ms":6987,"temperature":1.0,"reasoning_tokens":675,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T19:19:41.653534+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same five paths on a multi-billion-vector DiskANN-style index with deeper traversal and a working set far larger than GPU memory; if GORIO’s 1.31× and 4.89× advantages disappear or reverse under that miss load, the substrate claim does not hold for the intended setting.","supporting_citations":[],"review_version":1}