Flow-CoDel Cuts Queuing Delay by 31 Percent in Network Bufferbloat Simulations

In network simulation trials evaluating Active Queue Management strategies, an enhanced version of the Controlled Delay algorithm designated as Flow-CoDel demonstrated a 31 percent reduction in average queuing delay under near-congestion conditions compared to standard CoDel, while drastically curbing packet drops and distributing remaining losses equitably. The study, conducted by researchers Morteza Seyyed Ashrafi and Ali Ghiasian and published in Discover Networks, investigates solutions to bufferbloat—excessive queuing latency that degrades communication performance despite ample transmission bandwidth. Evaluating queue behavior across stable and overloaded traffic regimes over simulation windows up to 25,000 milliseconds, the authors showed that coordinating packet discards across active flows resolves congestion bottlenecks more effectively than traditional algorithms that discard packets indiscriminately. Their findings suggest that integrating flow awareness into single-queue router mechanisms can reconcile the competing operational demands of low latency, throughput stability, and fairness.

The fundamental motivation behind Flow-CoDel stems from the operational limitations of conventional end-to-end congestion control protocols operating without proactive queue management at intermediate routing devices. The Transmission Control Protocol, which governs internet traffic, relies on transport-layer mechanisms like TCP Reno, TCP Cubic, and Bottleneck Bandwidth and Round-trip propagation time to regulate data transmission between endpoints based on observed network signals such as packet loss and round-trip time variations. However, because end-to-end transport protocols lack direct visibility into internal router queues, they cannot independently prevent localized buffer accumulation before congestion impairs overall throughput. Active Queue Management mechanisms address this operational gap by actively monitoring queue metrics inside network routers and signaling congestion back to senders before buffers become completely saturated. When implemented effectively, queue management disciplines reduce queue waiting times, maintain high throughput, and ensure balanced resource allocation among multiple simultaneous connections competing for shared bottleneck link capacity.

Historically, router queue management has evolved from early heuristic drop techniques toward sophisticated delay-centric and flow-differentiating architectures. Early schemes like Random Early Detection sought to avert bufferbloat and global TCP synchronization by probabilistically discarding packets based on running average queue lengths, but their operational efficacy depended on manual parameter tuning that frequently produced unstable queue dynamics under volatile traffic loads. Weighted Random Early Detection expanded upon this foundation by incorporating Quality of Service differentiation based on IP precedence or Differentiated Services Code Point markings, prioritizing latency-sensitive applications like Voice over IP by selectively dropping lower-priority traffic. However, Weighted Random Early Detection required complex threshold configurations and precise packet classification to function reliably. Modern queue management research subsequently diverged into two complementary directions: delay-based algorithms like PIE and Controlled Delay that focus primarily on holding down queue latency, and single-queue mechanisms like CHOKe and Stochastic Fair Blue that employ statistical methods to penalize aggressive flows.

Among delay-centric mechanisms, the standard Controlled Delay algorithm gained prominence due to its parameter-free operational model and its reliance on packet sojourn time—the actual time a packet spends traversing a queue—rather than aggregate buffer size. Controlled Delay tracks minimum queuing delay over a 100-millisecond observation interval, entering an active dropping state whenever packet sojourn time exceeds a target threshold of 5 milliseconds and scheduling drops at intervals inversely proportional to the square root of the accumulated drop count. However, standard Controlled Delay enforces a blind packet-dropping policy that treats all enqueued packets identically, ignoring flow-specific attributes such as round-trip times, congestion window sizes, or individual throughput demands. When Controlled Delay discards a packet belonging to a larger TCP stream, the sender retransmits subsequent packets, generating redundant transmissions. Furthermore, because standard Controlled Delay fails to signal congestion universally across all competing streams, uninformed senders continue transmitting at elevated rates, triggering cascading drop cycles.

Prior efforts to rectify these fairness anomalies produced multi-queue solutions such as Flow Queue CoDel, which segregates separate data streams into distinct individual queues scheduled via Deficit Round Robin while applying delay-based dropping independently within each queue. Although Flow Queue CoDel successfully isolates aggressive flows and protects short-lived data bursts like DNS queries and TCP handshake exchanges, it requires substantial computational overhead to manage per-flow state tables and presents complex implementation hurdles in high-speed hardware environments. Similarly, DualQ architectures like L4S require fundamental changes across network protocol stacks. In contrast, Flow-CoDel preserves a single First-In-First-Out queue architecture while incorporating selective packet inspection during congestion events. By avoiding multi-queue scheduling structures, the algorithm approximates flow fairness within a standard single-queue framework. The proposed mechanism operates by inspecting buffer contents when delay thresholds are breached, pinpointing the specific traffic streams driving queue growth and applying targeted drop decisions without maintaining separate queues.

To achieve practical hardware and software execution within a single queue, the authors designed Flow-CoDel around an efficient flow tracking and constant-time deletion architecture. The router maintains a compact flow table indexed by flow identifiers, storing the arrival timestamp of the most recent packet along with an activity counter. A flow is formally designated as active if it has at least one packet currently enqueued within the buffer during a recent measurement window, ensuring that the algorithm targets only those connections actively contributing to buffer buildup. Rather than scanning the entire queue sequentially during congestion episodes, Flow-CoDel maintains per-flow registers containing pointers directly referencing the last enqueued packet of each active flow within a doubly-linked list structure. When a drop must occur, the controller accesses the identified flow’s last packet via its pointer and unlinks it in O(1) constant time without disturbing the sequential order of surrounding packets, bounding processing overhead for software queue disciplines and programmable gate arrays.

The execution logic of Flow-CoDel formally monitors packet arrivals and departures through a structured control loop. When a packet is dequeued from the First-In-First-Out buffer, the controller calculates its sojourn time by subtracting its recorded arrival timestamp from the current system clock. If the sojourn time remains below the designated target delay of 5 milliseconds, the algorithm resets its tracking variable to zero. When queuing delay exceeds the target threshold continuously across a 100-millisecond monitoring interval, the algorithm transitions into dropping mode, sets its drop count to one, and schedules the next drop event using the interval scaled by the inverse square root of the drop count. Crucially, Flow-CoDel does not discard consecutive packets from arbitrary flows; instead, it systematically selects each active flow and discards exactly one packet from its queue backlog during that drop cycle, prompting competing TCP senders to initiate coordinated rate backoffs that rapidly alleviate buffer pressure.

To rigorously measure the performance of Flow-CoDel against baseline queue management strategies, the authors developed a specialized four-part simulation framework benchmarking the new algorithm against unmanaged First-In-First-Out tail-drop queuing and standard Controlled Delay. The simulation architecture consists of four distinct sub-programs: a TCP window management module that tracks sequence numbers and modulates congestion windows; a packet prioritization stage that assigns time-based priorities; a queue execution engine that processes packet extractions and applies the Flow-CoDel control loop; and an acknowledgment handler that returns standard ACKs and issues triple duplicate ACKs upon detecting lost segments. The simulated network modeled a 24 Mbps transmission link utilizing fixed 1000-byte packets, yielding an effective channel capacity of 3000 packets per second, or exactly 3 packets per millisecond time slot. Experimental trials evaluated performance across two observation windows—2,500 milliseconds and 25,000 milliseconds—across traffic loads ranging from near-congestion at rho = 0.9 to severe overload conditions spanning rho values from 1.1 to 1.5.

The quantitative simulation findings documented significant queuing delay improvements alongside an exceptional reduction in total packet loss when evaluating Flow-CoDel under steady-state conditions. At a near-congestion load of rho = 0.9, Flow-CoDel achieved a 31 percent reduction in average queuing delay compared to standard Controlled Delay, while providing an overall 68 percent reduction in average delay compared to unmanaged First-In-First-Out queues. As traffic loads escalated into sustained saturation, the delay margin over standard Controlled Delay shifted to 20 percent at rho = 1.1, 21 percent at rho = 1.2, 14 percent at rho = 1.3, and approximately 4 percent at both rho = 1.4 and rho = 1.5, yielding an average delay improvement of 16 percent across all tested load conditions. Concurrently, packet drop measurements revealed dramatic disparities: at rho = 0.9, standard Controlled Delay discarded 1,684 packets, whereas Flow-CoDel discarded only 89 packets, producing a remarkably balanced, uniform distribution of dropped packets across all participating user connections.

Simulations incorporating heterogeneous workloads demonstrated that Flow-CoDel maintains its drop reduction advantages under diverse traffic conditions, while revealing an inherent operational trade-off regarding link utilization efficiency. When evaluating mixed traffic environments where twenty-five percent of connections operated as short-lived flows active only between 1.0 and 1.5 seconds, Flow-CoDel dropped just 53 packets compared to 39,541 packets discarded by standard Controlled Delay, preventing short-lived flows from being starved of bandwidth. However, this conservative packet preservation strategy introduces a trade-off in instantaneous transmission efficiency: while standard Controlled Delay sustains high link utilization by aggressively purging packets to drain buffers, Flow-CoDel experiences intermittent intervals of reduced link utilization in exchange for enhanced flow fairness and packet conservation. The authors note that while these simulation trials confirm the viability of Flow-CoDel within controlled settings, further research remains necessary to evaluate the algorithm across broader network parameters, including varying physical link capacities, larger flow populations, non-responsive UDP traffic, and wide-ranging round-trip time distributions.

Subject of Research: Active Queue Management and bufferbloat mitigation through per-flow packet dropping in single FIFO queues

Article Title: Flow-CoDel: per-flow active queue management for balanced delay and drops in bufferbloat mitigation

Article References: Ashrafi, M. S., & Ghiasian, A. (2025). Flow-CoDel: per-flow active queue management for balanced delay and drops in bufferbloat mitigation. Discover Networks, 1(1), Article 12. https://doi.org/10.1007/s44354-025-00012-z

Image Credits: AI Generated

DOI: 10.1007/s44354-025-00012-z

Keywords: Active Queue Management, bufferbloat, Controlled Delay, Flow-CoDel, FQ-CoDel, queuing delay, packet loss, TCP congestion control, single-queue AQM, FIFO queue, Deficit Round Robin, Random Early Detection, sojourn time, flow fairness, link utilization, round-trip time, drop-next, FPGA