At ByteDance, cloud gateway clusters orchestrate petabit-scale aggregate traffic. Traditional ASIC-only gateways fail to meet these escalating demands due to severe on-chip resource constraints and limited programmable flexibility, while pure software solutions or alternatives like disaggregated SmartNICs struggle to match terabit-scale line-rate throughput. To bridge this gap, we present Gryphon, a hyperscale cloud gateway built on a hybrid architecture that integrates DPUs directly into the switching ASIC's forwarding path. This design resolves the fundamental tension between capacity and speed, expanding table scale by up to 1000× and augmenting programmability, while sustaining 1.6 Tbps line-rate throughput at a cost of only ~8 μs in additional average latency. To manage this hardware heterogeneity, we introduce Hierarchical Co-Offloading (HLCO) in the data plane, achieving >99.9% fast path hit rate, while retaining software fallback for complex operations. In the control plane, we develop an abstraction layer (P4Bridge) that decouples hardware specifics from policy configuration. Gryphon has been operating at production scale for over a year, deployed on hundreds of nodes across multiple Availability Zones. We also share production measurements and operational experiences that serve as the first hyperscale-proven guidelines for next-generation DPU-augmented cloud gateways.

Gateways based on the programmable P4 language are becoming a key component in data center traffic management, offering cost-effective solutions for high-throughput environments. However, traditional monitoring techniques like sFlow and NetFlow lack the needed precision to meet the demands of large-scale data centers. In this paper, we presentFlowLog, the first sketch-based and end-to-end flow monitoring system capable of accurate flow size estimation in 400 Gbps production environments. FlowLog integrates the novelByteSketchalgorithm, a transmission subsystem, and a high-speed analysis subsystem, achieving high accuracy even in demanding data center scenarios. Deployed for over six months in ByteDance’s data center with peak bandwidths exceeding 400 Gbps, FlowLog outperforms existing solutions such as Bytehunter sFlow and state-of-the-art sketches in both accuracy and efficiency. Additionally, through real-world deployment, we gained valuable insights that guided improvements in system compatibility, integration ease, and traffic detection. These lessons resulted in a more adaptable system, better handling complex traffic patterns and ensuring minimal overhead during monitoring.

In-band Network Telemetry (INT) plays a vital role in modern computer networks since it can frequently collect device-internal states hop-by-hop to enable high-resolution network monitoring. To perform network-wide monitoring, high-level orchestration is required to establish multiple monitoring paths across the entire network. Existing research minimizes telemetry overhead by maintaining minimal path overlap, allowing the system to capture the network state with low overhead while assuming that there is no network anomalies. However, in production deployments, the measurement system is not always executed in an environment without network anomalies. When INT probes are dropped or delayed caused by anomalies, this can result in measurement failures at additional locations than just the anomaly sites, leading to (i.) incorrect decisions by network applications, (ii.) hindering the analysis, and (iii.) localization of anomalies. To overcome these issues, we introduce ATINT, a novel path planning algorithm that guarantees uninterrupted measurements in the locations without anomalies when the number of network anomaly locations is below a predefined threshold. We defined the problem as an Integer Linear Programming model and proposed a heuristic algorithm to find possible optimal solutions. ATINT orchestrates redundant measurements in a staggered manner, continuously providing visibility of the network state when anomalies take place at a low measurement cost. Our evaluation within a network composed of 245 switches shows that our solution reduces the interruption probability to 19.4% and decreases the additional unmonitored links to 0.07% with respect to state-of-the-art INT probe solution.

The emergency backup problem of optical inter-datacenter (inter-DC) networks is widely studied to avoid data loss caused by disasters. However, when unpredictable disasters such as earthquakes and large-scale power outages attack the optical inter-DC network, it is impossible to predict the network damage under a discrete time domain through the early warning system (EWS). Designing a reliable emergency backup scheme is becoming a key challenge in the optical inter-DC network in unpredictable disasters. Previous work only optimized emergency backup in the optical inter-DC networks in predictable disasters, which would lead to the unreliability of emergency backup of the optical inter-DC networks in unpredictable disasters. Therefore, this work proposes a reliable emergency backup method for the optical inter-DC networks in unpredictable disasters. Firstly, the disaster probability propagation model of the optical inter-DC network in unpredictable disasters is established based on the Markov process to quantify the network damage under a discrete time domain and install the emergency backup reliability problem. Next, constructing a time-sensitive variable time extension network (TS-VTEN) transforms the dynamic emergency backup reliability problem into a static flow problem. Finally, the problem is expressed as integer linear programming (ILP). The proposed ILP method outperforms state-of-the-art emergency backup methods by simultaneously achieving high reliability and time efficiency.

Network traffic measurement is central to successful network operations, especially for today’s large-scale networks. Although existing researches have made great contributions, they fail to achieve the following three criteria simultaneously: 1) low total overhead, which refers to the ability to acquire information in low bandwidth cost; 2) constant overhead of on-path measurement; and 3) easy to organize, meaning that the system is able to satisfy most of the user’s measurement requests. We propose SPT to meet these three criteria simultaneously. To realize flexible control, SPT supports various measurements in INTs passport mode. To save the bandwidth, SPT adopts sketch-guided elephant flow selection, extension filed reuse, and compression of control instructions. To achieve flexed-length extension fields, SPT decomposes multiple measurement tasks carried by one packet of a flow into separate measurement tasks carried by multiple sequential packets of the flow. We also need to consider the deployment of the framework so we optimize the pipeline to save the SRAM resources. We implemented a SPT prototype on a testbed with P4 software switches in a linear topology. We conduct comparative experiments with a classical measurement framework SINT and a state-of-the-art measurement framework PINT. Experimental results show that SPT can obtain per-hop flow-level information with the highest accuacy when the overhead is the same.