{"id":"e18dbecd-3f3f-48d4-a27d-7619251a88c4","arxiv_id":"1908.10739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Measured Age of Information on real Internet and local Wi-Fi IoT testbeds and derived how synchronization bias shifts average age under linear, exponential, and logarithmic penalties.","lead":"This paper measures how old the newest information is at the receiver, called Age of Information, over real TCP and UDP links on desktop PCs and low-power IoT devices. It also derives how clock differences between sender and receiver distort AoI measurements, and recommends TCP over UDP for the tested IoT hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central empirical claims rely on an unverified constant clock-offset assumption; clock drift would invalidate the 'absolute AoI up to a constant offset' statement and could distort the reported age curves.","rationale":"The reader's weakest_assumption correctly identifies the constant clock-offset assumption of Eq. (10) as the critical point. The entire measurement methodology treats B as fixed, and the statement that plotted values are absolute AoI up to a constant offset (Section III) is the bridge from raw timestamps to the empirical claims in Figures 6 and 7. If B drifts or RTT varies, the average age at different throughput settings is corrupted by different amounts, which could produce artificial U-shapes or plateaus. The theoretical derivation in Section III remains valid as a conditional statement, but its application to the testbed results is unverified. A concrete drift-measurement experiment would settle this. Other issues, such as the missing data and error bars and the apparent unit inconsistency in Figure 8, are real but secondary: they affect reproducibility and presentation, whereas the sync assumption directly threatens the validity of the quantitative age comparisons that the central claim rests on. Therefore the reader's CONDITIONAL verdict remains appropriate: authors should verify the sync assumption and release the data before the empirical claims are accepted.","tokens_in":8041,"tokens_out":7574,"duration_ms":82184,"concrete_test":"Re-run the Internet and IoT experiments while measuring the TX-RX clock offset at the start and end of each throughput step, or continuously with a GPS-disciplined reference. Then compute average age with and without drift compensation for each rate. If the compensated and uncompensated average-age values differ by more than a few percent of the observed range, or if the U-shape or plateau features shift, the constant-B assumption fails and the headline empirical claims would need to be re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III derives the synchronization bias under the explicit assumption that the difference between transmitter and receiver clocks is a constant B, Eq. (10), and the measurement section then states that the values plotted represent absolute AoI up to a constant offset, with the variation among values being correct. This is the hinge for the paper's central empirical results: the U-shaped TCP age curve and the flat UDP 'Panicked' region are comparisons of average age across increasing packet rates. If B is not constant over the measurement sweep, the offset differs between measurement points, so the plotted shape is not purely AoI variation. The paper estimates the offset only from RTT measurements taken before the experiments; no clock-drift or repeated RTT characterization is reported. On Internet paths with roughly 80 ms RTT (Section IV-D), even modest drift or congestion-dependent RTT changes can produce biases comparable to the age differences that define the curve shapes. Without drift data, the claim that the observed shapes are caused by TCP retransmissions versus UDP loss, rather than by a time-varying synchronization error, is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Age of Information (AoI) on real-life networks using two testbeds: an Internet-based testbed with regular PCs and a local Wi-Fi IoT testbed with ESP32 devices. It derives the effect of a constant clock offset between transmitter and receiver on average age for linear, exponential, and logarithmic penalty functions, showing that the bias is a constant shift only for linear age and becomes non-constant for nonlinear penalties. Empirically, it reports that TCP retransmissions yield a U-shaped average-age-versus-rate curve over the Internet, while UDP losses dominate and can keep average age flat in a 'Panicked' region; on the IoT testbed, device CPU/memory limitations prevent a clear U-shape. The paper claims to be the first reported investigation of AoI on real IoT testbeds and argues for age-aware transport protocol design.","tokens_in":8210,"tokens_out":4868,"duration_ms":51650,"significance":"If the measurements are trustworthy, the paper would provide one of the first empirical mappings of transport-protocol behavior to AoI curves on real Internet and IoT paths, and a useful warning that synchronization errors distort nonlinear age penalties nonlinearly. The synchronization-bias derivation (Section III) is a parameter-free consequence of the AoI definition and is a solid contribution: it correctly shows that a constant clock bias shifts linear average age by alpha B and produces non-constant bias for exponential and logarithmic penalties. The paper also makes concrete, falsifiable predictions about when U-shaped age curves appear. However, the empirical core currently rests on methodology that is not sufficiently reported, and the central synchronization assumption is unverified over the measurement duration.","major_comments":[{"comment":"The claim that 'the values plotted represent absolute AoI up to a constant offset, the variation of the values within themselves being correct' relies on the constant clock bias B in Eq. (10) remaining constant over the entire experiment. The paper estimates B from RTT measurements taken only before the experiments and provides no clock-drift monitoring or repeated RTT characterization. On the Ankara-Istanbul path with approximately 80 ms RTT, even a small drift would introduce a time-varying offset that shifts the average age between different measurement points, so the U-shaped TCP curve and the flat UDP 'Panicked' region could partly reflect synchronization artifacts rather than pure AoI variation. The authors should either measure and correct for clock drift (e.g., using GPS/PTP synchronization) or report drift bounds that are small compared to the age differences visible in Figs. 6 and 7.","section":"Section IV-E, after Eq. (15)"},{"comment":"The experimental sections do not report confidence intervals, number of repetitions, raw data, packet sizes, or the exact rate-increase schedule. All central conclusions—the U-shaped TCP curve, the flat UDP Panicked region, the 300 ms jitter claim, and the protocol comparison—are drawn from single-trace plots without error bars, so the statistical significance of the observed shapes is not established. At minimum, the authors should provide multiple trials with variability information (e.g., box plots or confidence bands) and make the measurement data available for independent verification.","section":"Sections IV-D, IV-E, IV-F"},{"comment":"The statement that 'without loss of generality, the queuing delay is negligible' for the UDP Internet experiments is an unverified, load-bearing assumption for the claim that packet loss is the dominant age-inflating factor. The same section later reports that packet-wise delays jump to a higher value in the Panicked region, suggesting that delays are not irrelevant. To support the loss-dominated interpretation, the authors should measure or bound queuing delay (e.g., through active probing or knowledge of intermediate router buffers) rather than asserting negligibility from the experimental results themselves.","section":"Section IV-E"}],"minor_comments":[{"comment":"The y-axis label 'average age (usec)' appears to conflict with the plotted values, which range around 88-100 and resemble the 'Success Percentage' values; the figure should use a dual-axis layout with clearly labeled axes or correct the unit/scale.","section":"Figure 8"},{"comment":"Please specify the packet size, the exact range and step of the packet generation rate, and whether the rate was measured at the application layer or the socket layer; this is necessary for reproducibility and for interpreting the 'throughput' axis in Figs. 6 and 7.","section":"Section IV-D"},{"comment":"The abstract states that 'excessive number of AoI measurements are provided,' but no raw measurement data or dataset link is included in the manuscript; either provide the data or rephrase this statement.","section":"Abstract and Section IV"},{"comment":"The symbols beta and theta are defined after the equations; define them just before Eq. (16) to improve readability, and consider presenting the exponential/logarithmic bias expressions in a more structured form to reduce the risk of transcription errors.","section":"Section III, Eqs. (16)-(17)"},{"comment":"The author list of reference [14] contains an empty field ('H. B. Beytur, , S. Baghaee'); please correct the citation.","section":"Reference [14]"},{"comment":"The claim of being the 'first reported investigation of AoI on real IoT testbeds' should be qualified relative to reference [7], which already performed AoI measurements over WiFi/LTE links; specify that the novelty concerns low-power IoT devices (e.g., ESP32) and the TCP/UDP comparison, and define the reported '300 ms jitter' metric explicitly.","section":"Introduction and Section IV-F"}],"recommendation":"major_revision","confidential_remarks":"The synchronization-bias derivation is sound and a useful citable result, but the empirical claims require substantial additional support: clock-drift characterization and statistically meaningful measurement reporting. The novelty claim about being 'first' on real IoT testbeds is somewhat overstated given [7], and the authors should position their contribution more carefully. If the authors can add drift measurements, confidence intervals, and a fuller experimental description, the paper could become acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is a solid incremental extension of the authors' earlier AoI measurement work, and the sync-bias derivation is the cleanest part. Equations (13)–(17) are new to me: they show that a constant clock offset shifts linear average age by a constant αB but distorts nonlinear penalties nonlinearly, which is a genuinely useful caution for anyone measuring AoI on real systems. The ESP32 testbed data is also rare; there is almost no public AoI measurement from actual IoT devices, so the LWIP UDP jitter observation and the TCP/UDP throughput asymmetry are worth knowing.\n\nThe experiments themselves are the weak spot, mostly in presentation. No raw data, no confidence intervals, no exact packet sizes or generation-rate steps, and Figure 8 has a unit inconsistency on the y-axis. More importantly, the paper leans on the constant-offset assumption (Eq. 10) for the claim that plotted values represent absolute AoI up to a constant offset. If the clock offset drifts over the sweep, the \"variation within themselves being correct\" statement no longer holds, and the shapes of the TCP U-curve and the flat UDP Panicked region could partly be artifacts. The stress-test note is right that the RTT estimate made before the experiment does not bound drift during the experiment. I would not call this fatal — the observed delay behavior and packet-loss patterns are consistent with the protocol mechanisms described — but the authors should either provide clock-drift measurements or soften the claim to \"relative AoI with potential bias variation.\" Their own Eq. (15) only works for constant B.\n\nThe citation pattern is fine; they build on their own prior work, which is expected in a measurement program, and they cite the FCFS queueing theory they compare against. The paper does not oversell the theoretical contribution — it is explicit that the formulas are for a constant bias.\n\nWho should read it? Anyone building AoI measurement tools or experimenting with TCP/UDP age on real networks. It deserves a serious referee, but only with the expectation that the authors supply data, fix the figure, and address drift.","headline":"Useful sync-bias analysis and rare IoT AoI measurements, but the clock-drift assumption and missing data keep the empirical claims conditional.","tokens_in":8710,"tokens_out":2072,"would_cite":true,"duration_ms":22199,"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 paper presents the first reported Age of Information measurements on real IoT testbeds and shows that transport protocol choice and device capability, not queueing theory alone, set the age-versus-rate curve.","keywords":["Age of Information","AoI measurement","IoT testbed","TCP","UDP","clock synchronization","status updates"],"falsifier":"A concrete test: on the IoT testbed, deliberately introduce a time-varying clock skew instead of a constant offset and re-measure average age; if the measured age changes by more than the RTT-bounded constant offset predicted by the paper's bias formula, the synchronization model fails. Separately, replay the UDP Internet experiment on a path whose bottleneck is a large buffer rather than loss, and check whether average age keeps decreasing with rate; if it flattens anyway, the 'Panicked' flat-age mechanism is not specific to loss-dominated paths.","tokens_in":7831,"feed_emoji":"📡","tokens_out":8769,"duration_ms":82185,"temperature":0.7,"pith_summary":"The paper sets out to measure Age of Information (AoI), the time since the newest received status update was generated, on real networks rather than in queueing models alone. It builds two testbeds, one using regular PCs over the Internet and one using low-power IoT nodes on a local Wi-Fi network, and records average age for TCP and UDP traffic. The central finding is that the transport protocol and the hardware at both ends materially shape the age curve: TCP's retransmissions produce a U-shaped age-versus-throughput plot similar to FCFS queues, while UDP's packet losses dominate at high load and can hold the average age flat. The paper also shows that a constant clock offset between transmitter and receiver shifts linear age by a constant but distorts non-linear age penalties nonlinearly, which can make an unsynchronized measurement select the wrong operating point. If these results hold, they ground AoI theory in real hardware and caution that protocol selection and synchronization must be part of any practical freshness metric.","feed_headline":"Real IoT tests: TCP retransmissions U-shape age, UDP flattens","feed_subtitle":"First fresh-data measurements on real IoT hardware show why TCP, UDP, and clock sync decide how old your updates look.","key_machinery":"The load-bearing machinery is the sawtooth age process $\\Delta(t)=t-U(t)$, together with its area-normalized time average; the constant clock-bias model $t_{RX}=t_{TX}+B$; and the mapping of measured age-throughput curves onto FCFS queue predictions. This machinery turns synchronization error into an explicit age-bias expression, and the paper derives from it that linear penalties receive a constant shift $\\alpha B$ while exponential and logarithmic penalties receive a rate-dependent distortion, which is why the measurements restrict attention to linear age with an offset bounded by round-trip time.","core_discovery":"The paper claims that average Age of Information on real networks is not determined by queueing alone: the transport protocol and the transmit/receive hardware reshape the age-versus-rate curve. In Internet-scale TCP flows, retransmissions make the average age follow a U-shape reminiscent of FCFS queues, because packet delays and queueing grow with load. In UDP flows there are no retransmissions, so at high load the curve enters a 'Panicked' region where heavy packet losses offset the higher update rate and average age stays flat. On constrained IoT nodes, CPU and buffer limits dominate, so the U-shape may not appear at all and the practical choice between TCP and UDP can be reversed. The paper also derives that when the age penalty is nonlinear, a constant clock offset between sender and receiver does not merely shift the measured average by a constant; it distorts it nonlinearly and can point to the wrong operating point, so the paper restricts its measurements to linear age with a synchronization error bounded by round-trip time.","pith_inferences":["One testable consequence not drawn in the paper: an adaptive transport that switches between TCP and UDP based on measured loss and delay could keep average age below either protocol alone, because the two protocols fail in complementary regimes.","The constant-bias analysis suggests a practical diagnostic for AoI measurement campaigns: re-run the same load with a deliberately injected clock offset and check whether the average age shifts by the predicted amount; agreement validates the synchronization model, and disagreement reveals clock drift.","The device-bottleneck result generalizes to other constrained radios, so on a very low-bandwidth link the U-shape should reappear because the channel, not the CPU, becomes the bottleneck, which is directly testable.","If the reported UDP jitter is indeed a buffer-management bug in the lightweight IP stack, fixing it could make UDP the better choice for IoT age, so the paper's TCP recommendation is contingent on stack quality rather than on protocol fundamentals."],"forward_implications":["Average AoI versus offered load is not protocol-agnostic: TCP's retransmissions produce a U-shaped curve, while UDP's losses can keep average age flat at high load, so AoI-optimal rate control must account for the transport layer.","On low-power IoT nodes, the bottleneck shifts from the network queue to the device CPU and buffers, so the U-shape disappears and throughput alone is not a reliable proxy for freshness.","Selecting UDP over TCP on constrained IoT devices is risky given the measured multi-hundred-millisecond jitter from the lightweight IP stack, so TCP can give fresher updates until UDP buffer handling improves.","For non-linear age penalties, clock synchronization is not cosmetic: a constant offset changes the measured average penalty nonlinearly, so unsynchronized deployments may pick the wrong sampling rate.","With linear age and an RTT-bounded synchronization offset, the measured values are absolute up to a constant, so comparisons of age variation across rates remain valid."],"supporting_citations":[{"why":"defines Age of Information and poses the update-rate question that motivates measuring age in real systems.","marker":"[2]"},{"why":"gives the FCFS/LCFS queue analysis whose U-shaped age-versus-rate prediction is used to interpret the TCP measurements.","marker":"[5]"},{"why":"supplies the general single-server queue age formula used as the theoretical baseline for the measured curves.","marker":"[6]"},{"why":"reports earlier real TCP/IP age measurements over WiFi, Ethernet, and LTE, the direct precursor this work extends.","marker":"[7]"},{"why":"establishes the Gamma-arrival FCFS result behind the expected U-shaped average-age characteristic.","marker":"[8]"},{"why":"is the authors' earlier conference report of real-connection age measurements that this paper deepens with IoT testbeds.","marker":"[14]"},{"why":"provides the non-linear age penalty framework used to derive the synchronization-bias expressions.","marker":"[15]"},{"why":"documents the TCP-versus-UDP throughput and CPU overhead difference on wireless links used to explain device bottlenecks.","marker":"[16]"}],"fun_headline_variants":["Real IoT: TCP U-shapes AoI, UDP flattens it","Transport protocol decides AoI curve shape on real IoT","IoT hardware can flip the best protocol for fresh data","Age-of-Information curves depend on TCP vs UDP on real systems","On constrained IoT, TCP vs UDP tradeoff reverses for AoI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole measurement interpretation rests on the assumption, stated in the synchronization section, that the transmitter and receiver clocks differ by a fixed amount for the whole experiment; the paper itself warns its RTT-based synchronization is 'not the best way,' so if the offset drifts or the round-trip time varies, the reported ages stop being correct up to one constant shift.","fun_headline_variants_meta":{"raw":{"variants":["Real IoT: TCP U-shapes AoI, UDP flattens it","Transport protocol decides AoI curve shape on real IoT","IoT hardware can flip the best protocol for fresh data","Age-of-Information curves depend on TCP vs UDP on real systems","On constrained IoT, TCP vs UDP tradeoff reverses for AoI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3356,"prompt_tokens":892,"completion_tokens":2464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":2377}},"tokens_in":508,"tokens_out":2464,"duration_ms":18059,"temperature":1.0,"reasoning_tokens":2377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:35:31.088577+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: on the IoT testbed, deliberately introduce a time-varying clock skew instead of a constant offset and re-measure average age; if the measured age changes by more than the RTT-bounded constant offset predicted by the paper's bias formula, the synchronization model fails. Separately, replay the UDP Internet experiment on a path whose bottleneck is a large buffer rather than loss, and check whether average age keeps decreasing with rate; if it flattens anyway, the 'Panicked' flat-age mechanism is not specific to loss-dominated paths.","supporting_citations":[{"cited_title":"Real-time status: How often should one update?","cited_arxiv_id":null,"evidence_quote":"defines Age of Information and poses the update-rate question that motivates measuring age in real systems."},{"cited_title":"Status updates through queues,","cited_arxiv_id":null,"evidence_quote":"gives the FCFS/LCFS queue analysis whose U-shaped age-versus-rate prediction is used to interpret the TCP measurements."},{"cited_title":"Age-of- information in practice: Status age measured over tcp/ip connections through wifi, ethernet and lte,","cited_arxiv_id":null,"evidence_quote":"reports earlier real TCP/IP age measurements over WiFi, Ethernet, and LTE, the direct precursor this work extends."},{"cited_title":"Age of Information: The Gamma Awakening","cited_arxiv_id":"1604.01286","evidence_quote":"establishes the Gamma-arrival FCFS result behind the expected U-shaped average-age characteristic."},{"cited_title":"Measuring age of information on real-life connections,","cited_arxiv_id":null,"evidence_quote":"is the authors' earlier conference report of real-connection age measurements that this paper deepens with IoT testbeds."},{"cited_title":"Age and value of information: Non-linear age case,","cited_arxiv_id":null,"evidence_quote":"provides the non-linear age penalty framework used to derive the synchronization-bias expressions."},{"cited_title":"Tcp and udp performance over a wireless lan,","cited_arxiv_id":null,"evidence_quote":"documents the TCP-versus-UDP throughput and CPU overhead difference on wireless links used to explain device bottlenecks."}],"review_version":1}