QUIC Loss Detection vs TCP Fast Retransmit

The reliability of transport protocols over inherently unreliable IP networks hinges on their ability to detect and recover from packet loss efficiently. Both TCP and QUIC implement mechanisms for loss detection and retransmission, but their designs diverge significantly due to their respective protocol architectures and underlying philosophies. TCP, a legacy protocol operating over IP and implemented in the kernel, relies on implicit signaling through acknowledgments and timers to trigger retransmissions. QUIC, a modern transport protocol built on top of UDP and implemented in user space, integrates richer signaling capabilities and more sophisticated algorithms for loss detection, taking advantage of its tightly-coupled encryption, multiplexing, and pacing mechanisms. Comparing QUIC’s loss detection with TCP’s fast retransmit mechanism reveals how the evolution of transport protocols has led to more responsive and accurate loss recovery strategies.

In TCP, fast retransmit is a cornerstone of its loss detection mechanism. When a TCP sender transmits a series of segments, it expects acknowledgments (ACKs) in return as the receiver processes them. If a segment is lost in transit, the receiver cannot acknowledge it directly. Instead, when subsequent segments arrive, the receiver continues to send duplicate acknowledgments for the last in-order segment received. Upon receiving three duplicate ACKs for the same segment, the sender infers that the next segment in sequence was lost and triggers a fast retransmit, resending the missing segment without waiting for a timeout. This threshold of three duplicate ACKs helps distinguish between actual packet loss and minor reordering, which is common in IP networks.

However, TCP’s reliance on duplicate ACKs as an implicit signal for loss has limitations. The approach is reactive and conservative, often resulting in delayed retransmissions in high-latency or lossy environments. TCP is also vulnerable to reordering, where out-of-order packet delivery can trigger spurious retransmissions, unnecessarily consuming bandwidth and reducing throughput. Additionally, TCP uses retransmission timeouts (RTOs) as a fallback for loss detection when duplicate ACKs are not observed, such as in cases of sparse or one-way data flow. RTOs are based on round-trip time (RTT) estimations and backoff algorithms, which can introduce significant delays, particularly in asymmetric or high-delay links.

QUIC addresses many of these challenges through its more refined and flexible loss detection framework, defined in RFC 9002. Unlike TCP, which relies primarily on ACKs and timeouts, QUIC uses acknowledgment-based loss detection augmented by per-packet metadata, packet number spaces, and precise timers. Each QUIC packet is assigned a unique, strictly increasing packet number, and acknowledgments include both the largest packet number seen and optionally a set of packet number ranges indicating which packets were received. This rich acknowledgment format allows the sender to precisely identify which packets have been lost without waiting for a threshold of duplicate acknowledgments.

In QUIC, loss detection is based on both packet number gaps and packet timers. When the sender observes that a packet with a lower number than an acknowledged packet has not been acknowledged within a certain time frame, it marks the packet as lost. The loss delay is typically set to 1.125 times the estimated RTT, providing a rapid response to detected loss while maintaining a safety margin to accommodate minor reordering. QUIC also maintains separate packet number spaces for different encryption levels—Initial, Handshake, and Application Data—allowing for independent loss detection and retransmission across stages of the connection lifecycle. This separation is particularly useful during connection establishment, where handshake messages may be lost or delayed.

Another notable improvement in QUIC is the use of Probe Timeout (PTO) instead of RTO. While similar in purpose to TCP’s retransmission timeout, the PTO in QUIC is calculated more aggressively and is used to trigger the sending of probe packets when no acknowledgment is received within a computed interval. These probe packets serve to elicit an acknowledgment from the receiver and can be either retransmissions of unacknowledged data or new application data if available. The PTO strategy enables QUIC to recover more quickly from idle or uncertain states, especially in the face of asymmetric loss or tail loss, where the final packets of a burst go unacknowledged.

QUIC also integrates packet pacing into its loss recovery mechanisms, spacing out transmissions to prevent bursts that might overwhelm the network. Combined with its congestion control, QUIC’s loss detection dynamically adjusts the congestion window and pacing rate in response to loss signals, which are tightly coupled to accurate RTT measurements and acknowledgment delays. The acknowledgment delay field in each ACK frame allows the sender to account for receiver-side delays in processing and transmitting ACKs, reducing the likelihood of spurious loss detection.

Additionally, QUIC’s user-space implementation enables rapid innovation and customization of loss detection algorithms. Because it is not bound to kernel constraints like TCP, QUIC stacks can experiment with adaptive heuristics, tailored congestion responses, and context-aware retransmission strategies without requiring operating system-level updates. This agility is particularly valuable for applications such as video streaming, online gaming, or real-time communication, where low-latency loss recovery can significantly enhance user experience.

In conclusion, QUIC’s approach to loss detection represents a significant advancement over TCP’s fast retransmit mechanism. By leveraging explicit acknowledgment metadata, multiple packet number spaces, refined timers, and user-space flexibility, QUIC provides faster and more accurate loss detection that adapts well to modern network conditions. While TCP fast retransmit remains effective and is deeply entrenched in today’s Internet, QUIC’s innovations demonstrate how transport protocols can evolve to meet the growing demands for performance, reliability, and responsiveness in an increasingly diverse and dynamic network landscape.

The reliability of transport protocols over inherently unreliable IP networks hinges on their ability to detect and recover from packet loss efficiently. Both TCP and QUIC implement mechanisms for loss detection and retransmission, but their designs diverge significantly due to their respective protocol architectures and underlying philosophies. TCP, a legacy protocol operating over IP and…