{"id":"8e1b6ea3-e8e0-485c-8601-835da325f6dd","arxiv_id":"2412.00266","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A schedule-aware routing algorithm with in-switch buffering achieves loss-free operation in fast-switched optical data center networks at microsecond-scale time slices, shown on a Tofino2 prototype and in large simulations.","lead":"URO is a routing framework for optical data center networks whose circuits reconfigure quickly, letting packets wait at intermediate switches instead of being lost when a circuit changes. It is demonstrated on commodity programmable switches at 2 microsecond time slices, and large-scale simulations show large latency gains over prior systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The offline path-planning model ignores per-hop propagation/transmission delays; for time slices shorter than per-hop OWD, computed 'optimal' paths are physically infeasible, undermining the sub-OWD central claim.","rationale":"The reader's weakest assumption concerned empty queues and packet arrival at the beginning of a time slice. My concern is related but more fundamental: the model abstracts away per-hop propagation and transmission entirely, which is precisely the effect that defines the sub-OWD regime. In §III-A the authors wave at these delays but do not include them in Eqn. 1 or in the feasibility conditions of Algorithm 1. For any time slice shorter than the per-hop OWD, a packet that departs a ToR in slice t will physically arrive at the next ToR in a later slice, so the offline-computed path can schedule a connection the packet cannot catch. This invalidates Property 1's optimality in the target regime; the system would rely on Property 3 rerouting at every hop, turning the offline computation into a warm start rather than a guarantee. The paper's own prototype at 2 us with 500 ns propagation does not enter this regime per hop, and the 0.12 us theoretical comparisons in §VI-C use the same delay-free model, so the sub-OWD claim is not empirically or theoretically established. This is a correctness risk in the central modeling contribution, not merely a missing artifact or a deferred derivation. It is addressable by adding per-hop delay to the graph model and to the proofs, or by restricting the claimed regime to slices longer than per-hop OWD. Because the reader already assigned CONDITIONAL for other reasons and this concern strengthens the need for revision without necessarily invalidating the 2-50 us prototype results, I leave the verdict unchanged. The proposed concrete test would settle whether the concern lands by making the physical timing explicit in simulation and checking whether planned paths remain feasible.","tokens_in":24149,"tokens_out":10422,"duration_ms":99385,"concrete_test":"Augment the htsim simulator with explicit per-hop propagation (500 ns) and transmission (120 ns) delays, set u=0.12 us (as in §VI-C), and replay URO's offline-computed paths for all ToR pairs at t0. Count the fraction of packets that arrive at each intermediate ToR after the planned departure slice and thus miss their reserved circuit; compare end-to-end FCTs against the current model that ignores these delays. If the miss fraction is non-negligible or FCTs deviate by more than 10%, the sub-OWD claims are not supported. A minimal analytical test: for u=100 ns and 500 ns propagation, the path S->R (slice 0) -> D (slice 1) is predicted to deliver in 200 ns but actually delivers no earlier than the next R->D circuit after the packet reaches R at slice 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that URO supports time slices shorter than a packet's OWD rests on the time-sliced graph model of §III-A, where edge availability times are the only constraint. The latency definition lat(p) = (tend - tstart + 1) × u (Eqn. 1) and the SUBPATH feasibility check t(src,r) <= t (Alg. 1, lines 17-18) assume that a packet which departs a ToR in slice t is present at the next ToR in time to use a circuit whose departure slice is >= t. This is only valid if per-hop propagation plus transmission is less than the slice duration u. The paper asserts these delays are 'substantially smaller than lat(p)' (§III-A), but for the sub-OWD regime the paper targets (e.g., PULSE 20 ns, Sirius 40 ns in Table I) and even for the 0.12 us theoretical comparisons in §VI-C, the 500 ns per-hop propagation used in §VI-A is several times u. A two-hop path S->R at slice 0 and R->D at slice 1 is declared feasible with latency 2u, but the packet physically arrives at R at time 500 ns (slice 5 with u=100 ns), missing the R->D circuit; it must wait until the next cycle. Thus Property 1's 'optimal path' is optimal only in a discrete model that ignores the very delay that makes slices sub-OWD. Runtime rerouting (Property 3) can mask misses, but then the actually-taken path is not the offline-computed one, the optimality guarantee is vacuous in the target regime, and loss-freedom depends on unbounded rerouting capacity rather than the planned schedule. The prototype's 2 us slices exceed per-hop OWD in the testbed, so it does not exercise the sub-OWD regime the paper motivates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Unified Routing for Optical networks (URO), a routing framework intended to work across fast-switched optical data center networks whose time slices may be shorter than a packet's one-way delay. The core idea is to exploit the fact that the cyclic optical schedule is known a priori: a backtracking algorithm (Algorithm 1) computes, for each source-destination pair and arrival slice, a path that may wait (buffer) at intermediate ToRs, minimizing a slice-counting latency metric (Eqn. 1). Three properties are proved: the chosen path has minimal latency (Property 1), per-hop lookup tables preserve the optimal path (Property 2), and rerouting after a missed slice yields the next optimal path (Property 3). A prototype on Intel Tofino2 implements calendar queues with queue pause/resume and reports a minimum 2 us slice; large-scale htsim simulations on Microsoft production traces compare URO with Opera and Sirius/VBS, reporting up to 12.8x lower FCTs at 2 us slices and resource usage within commodity limits.","tokens_in":24390,"tokens_out":21066,"duration_ms":189841,"significance":"If the claims hold, URO is a valuable step toward decoupling routing from the specifics of optical hardware, an area where most systems are tightly co-designed (Opera, Sirius, VBS). The paper's strengths are its explicitly stated modeling assumptions, a real prototype (2 us slices on Tofino2 is an impressive data point for commodity switches), the breadth of the evaluation (production traces, multiple baselines, failure scenarios, resource accounting), and the open-framework positioning (Lighthouse). The three properties are hand-checkable and appear correct within the stated discrete model. The main risk is that the model omits per-hop propagation and transmission delays, and the paper's own evaluation enters a regime (0.12 us slices with 500 ns per-hop propagation) where that omission is most consequential; the quantitative sub-microsecond comparisons should therefore be treated with caution until the model is refined. The needed corrections are within the manuscript's scope, so the contribution remains substantial.","major_comments":[{"comment":"The model omits per-hop propagation and transmission delays, and this omission is load-bearing exactly in the regime the paper targets. The feasibility check at Algorithm 1 lines 17-18, t(src,r) <= t, together with the latency definition in Eqn. (1), treats a packet as arriving at the next ToR at the start of its departure slice. The footnote in §III-A justifies this by saying these delays are 'substantially smaller than lat(p)', but this is contradicted by the paper's own settings: §VI-A sets per-hop propagation to 500 ns, while §VI-C evaluates 0.12 us slices and Table I lists 20-40 ns slice architectures. For u = 120 ns, a packet sent over S->R in slice 0 arrives at R at 500 ns, i.e., during slice 4, so a planned R->D circuit in slice 1 is missed and the path declared feasible by the algorithm is physically infeasible. Consequently, Properties 1-3 prove optimality only with respect to the discrete slice model, and in the sub-OWD regime the path actually taken is the product of runtime rerouting, whose 'optimality' is relative to the same idealized model. The fix is within reach: introduce a per-hop constant delay delta (propagation + transmission + guardband), replace the check with t(src,r) + ceil(delta/u) <= t, and either re-derive the properties under this refined model or explicitly restrict the optimality claims to slices with u >= delta. In particular, the 0.12 us lower bound in Fig. 11c should be recomputed, since with delta = 500 ns and u = 120 ns a single hop already costs several slices.","section":"§III-A, §III-C, Algorithm 1"},{"comment":"The abstract's claim that the prototype 'ensur[es] end-to-end, loss-free application performance' is stronger than what the system actually provides. In §IV-C, the one-shot lookup drops a packet when no departure queue is feasible, and the paper reports only that no drops were observed in the evaluation. Because the offline plan is built under the empty-queue assumption (§III-A) and the number of alternative paths stored per lookup entry is fixed, there is no argument ruling out drops under heavy bursts or adversarial arrivals; loss-freedom is an empirical observation, not a guarantee. Please state the load conditions under which loss-freedom is expected to hold, or soften the abstract and §I claims accordingly.","section":"§IV-C; abstract"},{"comment":"The complexity analysis of the offline algorithm is not rigorous as written. The claimed reduction to O(N^M / 2^(M^2+M)) rests on the unexplained assertion that 'half of the nodes are filtered out at each level', which is schedule-dependent and not established; moreover, the product in the text (1 × N/2 × N/4 × ... × N/2^M) evaluates to N^M / 2^(M(M+1)/2), not N^M / 2^(M^2+M), so the stated exponent needs correction. The 'polynomial time' wording is also imprecise: for a fixed hop limit M, the worst-case O(N^M) is polynomial in N, but the heuristic reduction is what makes the 1024-ToR projections in Table II plausible, and the pre-computation time at that scale is not reported. Please present the complexity bound formally and report the achieved pre-computation time for the larger configurations.","section":"§III-B"}],"minor_comments":[{"comment":"There is a typo in §IV-B: 'if the departure time slice is later than the arrival time slice of' should read '...arrival time slice of the packet'.","section":"§IV-B"},{"comment":"The walk-through of Fig. 4 is hard to follow because the path numbering in Fig. 4a and the table entries in Fig. 4b are not visually aligned in monochrome; a small table listing (arrival slice, destination, egress port, departure slice) per path would clarify the example.","section":"Fig. 4"},{"comment":"Both the 2 us minimum-slice derivation and the 50 ns queuing-delay estimation accuracy are deferred to the companion Lighthouse paper [33]; since these numbers support headline claims, please include the key derivation steps here or confirm the stable public availability of [33].","section":"§V"},{"comment":"The caption says architectures supported by URO are 'highlighted in blue', but the table is not color-coded in this version; please mark the supported rows with a symbol or footnote.","section":"Table I"},{"comment":"Several axis labels and legends (e.g., Fig. 2, Fig. 6) show the unit as 's' where the micro sign is missing; the units should render as 'us'.","section":"Fig. 2 and Fig. 6"},{"comment":"The footnote that h=2 outperforms h=3 for VBS in this scaled network is worth a sentence in the main text, since Fig. 11b otherwise suggests monotone improvement with h.","section":"§VI-C"}],"recommendation":"major_revision","confidential_remarks":"The paper's two headline quantitative anchors—the 2 us minimum slice and the 50 ns queue-delay estimation accuracy—are both delegated to the companion Lighthouse manuscript [33]; I recommend the editor verify that this companion paper is available and has undergone review, since the present manuscript's claims cannot be independently checked otherwise. The relationship to the earlier HOHO workshop paper [51] is acknowledged, but the delta (full system, prototype, extensive evaluation, failure handling) should be clear to readers. Scope-wise, the manuscript fits ToN well as a systems and evaluation paper; the theoretical part is best treated as an idealized model that needs the timing refinement described in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"URO is a serious systems paper: the Tofino2 prototype with calendar queues and one-shot rerouting is real engineering, the simulation study against Opera is careful, and the 1.4-12.8x FCT gains at 2us slices are credible. Giving credit where it is due, the paper is also honest about its HOHO ancestry and does not oversell the novelty of waiting-based routing. The formal latency model and the three optimality properties are plausible under the stated empty-queue and fixed-arrival assumptions. If you work on optical DCN routing, this is worth your time.\n\nThe soft spots are real, though. The biggest one is the propagation/transmission assumption in Section III-A. The feasibility check t(src,r) <= t is only valid if a packet that leaves a ToR in slice t arrives at the next ToR before the onward circuit departs. That requires per-hop propagation plus transmission to be smaller than the slice duration. The paper asserts this, but in the sub-OWD regime it explicitly targets (PULSE 20ns, Sirius 40ns) it is false: the 500ns propagation used in the evaluation is several times the slice. The theoretical 0.12us comparisons in Section VI-C violate this too. So Property 1's \"optimal path\" is optimal only in a model that ignores the very delay that makes slices sub-OWD. The 2us prototype does not exercise sub-OWD, since 2us is larger than the per-hop delay in the testbed. This does not kill the contribution for microsecond-scale slices, but it means the central \"sub-OWD\" claim is not supported as stated.\n\nThe time-complexity analysis is also garbled: the product formula \"1 x N/2 x N/4 x ...\" and the resulting O(N^M / 2^{M^2+M}) do not parse, and the \"half the nodes are filtered\" pruning factor is asserted without proof. Minor, but it should be fixed. The 2us limit and the 50ns queue-delay accuracy are deferred to the companion Lighthouse paper, so the feasibility derivation is not in this manuscript. No artifacts are released, which makes independent replication harder but is not unusual.\n\nFor a reader in this area, the paper is valuable and deserves a serious referee, but it needs revision: restrict the claims to slices larger than per-hop delay, or extend the model to account for propagation/transmission in the feasibility check. I would send it to peer review, with the expectation of heavy revision.","headline":"A real prototype and thorough simulations make URO worth reading, but the sub-OWD claim rests on an assumption that breaks in the very regime the paper motivates.","tokens_in":25089,"tokens_out":2158,"would_cite":true,"duration_ms":21864,"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":"URO claims that a single offline routing framework can make fast-switched optical data center networks work even when circuit time slices are shorter than a packet's own one-way delay, and it demonstrates this on commodity programmable…","keywords":["optical data center networks","fast-switched optical circuits","time slices","routing","programmable switches","calendar queues","Valiant load balancing","flow completion time"],"falsifier":"Load a URO testbed with a burst that makes a scheduled departure queue genuinely non-empty, then measure whether the rerouted packet's actual end-to-end latency ever exceeds the latency of the path that would have been chosen with accurate queue information. If such a case exists, the optimality claim under real queueing is refuted; if none exists, the zero-queue simplification holds under the tested regime.","tokens_in":23814,"feed_emoji":"⚡","tokens_out":5097,"duration_ms":43134,"temperature":0.7,"pith_summary":"URO (Unified Routing for Optical networks) is a routing framework for fast-switched optical data center networks, where optical circuits are reconfigured at microsecond or even nanosecond scale and a time slice can be shorter than a packet's one-way delay. Because the optical schedule is known in advance, URO computes paths offline so that packets know where to wait, and programmable switches buffer packets in calendar queues that are paused and resumed on a per-slice beat. The authors argue that this single design covers a wide range of optical hardware, from very short sub-one-way-delay slices to long slices where it reduces to k-shortest-path routing. They demonstrate a prototype on Intel Tofino2 switches with a minimum circuit duration of 2 microseconds and loss-free end-to-end application traffic, and simulations against Opera, Sirius, and VBS show lower latency and shorter paths at microsecond-scale slices. The payoff is that network architects no longer need to co-design routing with each new optical hardware point; one routing layer can serve many.","feed_headline":"URO routes fast-switched optical data center networks down to 2-microsecond circuits","feed_subtitle":"Offline path planning plus per-switch buffering lets packets survive circuit changes shorter than a packet's own one-way delay.","key_machinery":"The carrying mechanism is an offline backtracking algorithm over a time-varying graph. The graph has top-of-rack switches as vertices and optical circuits as edges labeled with the time slices when they exist; path latency is defined as the number of elapsed time slices times the slice duration. ROUTING finds the earliest last-hop circuit to the destination, then SUBPATH recursively searches backward for the shortest feasible chain of intermediate ToRs whose circuit times are nondecreasing, so each hop's departure is no later than the next hop's circuit time. Per-hop lookup tables let each ToR act on only the next hop while preserving optimality, and calendar queues, physical queues per egress port cyclically assigned to time slices, implement the waiting; an on-chip packet generator triggers queue pause and resume each slice. When a packet misses its planned slice, a one-shot lookup consults a queue-state bit array to reroute onto the next optimal path without recirculation.","core_discovery":"The paper claims that a single offline routing algorithm can make fast-switched optical data center networks work even when a circuit's time slice is shorter than a packet's one-way delay. Because the optical schedule is known ahead of time, URO computes the latency-minimizing path for each source-destination pair per time slice, allowing packets to stop and wait at intermediate top-of-rack switches until the next needed optical circuit appears. On Intel Tofino2 switches the authors implement this waiting as time-synchronized queue pausing and report a minimum circuit duration of 2 microseconds with loss-free application performance; in simulations with production DCN traffic, URO shortens paths and reduces flow completion times compared to the Opera architecture and to VLB-based designs at microsecond-scale slices.","pith_inferences":["The paper's zero-queue assumption suggests a testable design point: URO's advantages should shrink under sustained high load where queues are non-empty, and one could extend the offline path computation to use predicted queue occupancy per slice rather than zero.","Because URO treats any known cyclic schedule the same way, it opens the door to schedule design as a control knob, where future work could jointly optimize schedules with URO's latency metric.","The one-shot rerouting with multiple precomputed backup paths turns link failures into a missed-slice event, suggesting URO's robustness could be quantified analytically as a function of schedule redundancy rather than only by simulation.","Since URO separates routing from architecture, it could act as a compatibility layer that lets two different optical fabrics be bridged by a common ToR software stack, an interoperability benefit the authors invoke but do not demonstrate."],"forward_implications":["Fast-switched architectures with sub-one-way-delay slices become deployable on commodity programmable switches rather than requiring custom silicon.","One routing layer can replace architecture-specific routing co-designs, so new optical hardware can be adopted without building a new network stack.","Mice flows get lower latency and shorter paths than continuous-path designs like Opera, while elephant flows can still be offloaded to Valiant load balancing for throughput.","At slice durations above one-way delay the algorithm collapses to ordinary shortest-path routing, so the same framework covers both fast-switched and slow-switched regimes.","If packet-granularity nanosecond slices ever become practical in ToR hardware, URO's schedule-based offline approach reaches lower-bound flow completion times comparable to dedicated designs like Sirius and VBS."],"supporting_citations":[{"why":"Opera, the baseline continuous-path routing design URO must beat, and the source of the optical schedule and super-OWD comparison.","marker":"[12]"},{"why":"Sirius, the packet-granularity VLB architecture URO compares against at microsecond-scale slices.","marker":"[13]"},{"why":"VBS, the theoretical oblivious reconfigurable-network scheme whose lower bounds URO compares to.","marker":"[21]"},{"why":"Supplies the queue pause/resume capability on programmable switches that implements time-sliced buffering.","marker":"[25]"},{"why":"Companion work that derives the 2 microsecond Tofino2 timing limit and the 50 nanosecond queue-delay estimation accuracy used by URO.","marker":"[33]"},{"why":"Programmable calendar queues, the mechanism URO uses to map packets to future time slices.","marker":"[37]"},{"why":"Sub-round-trip-time transport protocol selected as URO's default and used for fair comparison with Opera.","marker":"[42]"}],"fun_headline_variants":["URO slashes optical circuit switching to 2 microseconds","Loss-free optical DCN routing at 2-microsecond circuits","Unified routing tames fast-switched optical networks","2-microsecond optical circuits, zero loss with URO","URO enables loss-free routing on 2us optical DCNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole path plan assumes packets arrive exactly at the start of a time slice and that top-of-rack switch queues are empty; if real queueing builds up, the precomputed fastest route may no longer be fastest, and the system's rerouting only re-optimizes under that same idealized model.","fun_headline_variants_meta":{"raw":{"variants":["URO slashes optical circuit switching to 2 microseconds","Loss-free optical DCN routing at 2-microsecond circuits","Unified routing tames fast-switched optical networks","2-microsecond optical circuits, zero loss with URO","URO enables loss-free routing on 2us optical DCNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000517,"raw_usage":{"total_tokens":2479,"prompt_tokens":889,"completion_tokens":1590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":1520}},"tokens_in":505,"tokens_out":1590,"duration_ms":11806,"temperature":1.0,"reasoning_tokens":1520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:33:03.637728+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Load a URO testbed with a burst that makes a scheduled departure queue genuinely non-empty, then measure whether the rerouted packet's actual end-to-end latency ever exceeds the latency of the path that would have been chosen with accurate queue information. If such a case exists, the optimality claim under real queueing is refuted; if none exists, the zero-queue simplification holds under the tested regime.","supporting_citations":[{"cited_title":"Expanding across time to deliver bandwidth efficiency and low latency,","cited_arxiv_id":null,"evidence_quote":"Opera, the baseline continuous-path routing design URO must beat, and the source of the optical schedule and super-OWD comparison."},{"cited_title":"Sirius: A flat datacenter network with nanosecond optical switching,","cited_arxiv_id":null,"evidence_quote":"Sirius, the packet-granularity VLB architecture URO compares against at microsecond-scale slices."},{"cited_title":"Optimal oblivious reconfigurable networks,","cited_arxiv_id":null,"evidence_quote":"VBS, the theoretical oblivious reconfigurable-network scheme whose lower bounds URO compares to."},{"cited_title":"Advanced congestion & flow control with programmable switches,","cited_arxiv_id":null,"evidence_quote":"Supplies the queue pause/resume capability on programmable switches that implements time-sliced buffering."},{"cited_title":"Programmable calendar queues for high-speed packet scheduling,","cited_arxiv_id":null,"evidence_quote":"Programmable calendar queues, the mechanism URO uses to map packets to future time slices."},{"cited_title":"Bolt: Sub-rtt congestion control for ultra-low latency,","cited_arxiv_id":null,"evidence_quote":"Sub-round-trip-time transport protocol selected as URO's default and used for fair comparison with Opera."}],"review_version":1}