Receiver Driven Congestion Control and Rate Manipulation

17 concepts12 papers1 proposal

In traditional networks, the sender rules the transmission rate. Senders probe for available bandwidth, react to packet drops, and adjust their congestion windows accordingly. However, this sender-centric model has a fundamental blind spot: the sender cannot directly observe the receiver's local environment. It does not know if a mobile device is juggling multiple applications, if a wireless access point is experiencing transient local interference, or if a datacenter switch buffer is on the verge of overflowing.

To overcome this limitation, a powerful paradigm has emerged across datacenters, wireless networks, and content delivery systems: shifting the intelligence of flow and congestion control from the sender to the receiver or the network edge. By placing the destination in control, networks can achieve precise bandwidth partitioning, near-instantaneous convergence, and strict fairness without requiring complex, coordinated state across distributed senders.

Explicit Pull-Based Transport

The cleanest implementation of this paradigm is a pure "pull" architecture. Instead of a sender pushing data as fast as its congestion window allows, the sender remains silent until the receiver explicitly requests data.

In datacenter networks, where concurrent "push" traffic frequently causes micro-bursts and catastrophic buffer overflows, Cho et al. proposed ExpressPass (Credit-Scheduled Delay-Bounded Congestion Control for Datacenters). ExpressPass shifts control entirely to the receiver via credit pacing: senders are only permitted to transmit a data packet after receiving a lightweight "credit" packet from the receiver. Similarly, Montazeri et al. designed Homa (Homa: A Receiver-Driven Low-Latency Transport Protocol Using Network Priorities), a protocol where receivers issue explicit "grants" to schedule incoming packets from multiple concurrent sources, preventing edge buffer overflows before they can even begin.

Sender                                     Receiver
  |                                           |
  | --------- (Silent / No Data) -----------> |
  |                                           |
  | <-------- Credit / Grant Packet --------- |  (Receiver paces requests)
  |                                           |
  | ========= Scheduled Data Packet ========> |  (Zero buffer overflow)

This pull-based logic is also native to Information-Centric Networking (ICN). As demonstrated by Zhang et al. (Using ORBIT for evaluating wireless content-centric network transport) and Perino et al. (A high speed information-centric network in a mobile backhaul setting), ICN receivers use a pipeline of "interest" or request packets to pull data chunks. By applying additive-increase/multiplicative-decrease (AIMD) logic directly to the request pipeline based on local round-trip times, the receiver manages congestion locally. In heterogeneous multicast scenarios, Chen et al. (SAID: A Control Protocol for Scalable and Adaptive Information Dissemination in ICN) showed that allowing individual receivers to dynamically scale their pending interest windows ensures that the slowest member of a multicast group does not inadvertently throttle the throughput of faster peers.

Indirect Control via Window Manipulation

While explicit pull protocols offer clean semantics, they require rewriting the transport layer. To achieve the same benefits for legacy TCP applications without modifying server-side infrastructure, systems can intercept and manipulate standard TCP acknowledgment (ACK) packets on the client side or at the network edge.

The primary lever for this indirect control is the TCP Advertised Window ($rwnd$). Normally, $rwnd$ is used by the receiver to communicate its remaining hardware buffer space. However, a receiver can deliberately under-report this value to artificially throttle the sender.

Legacy Server                              Mobile Client
    |                                            |
    | <========== TCP ACK (rwnd = 64KB) -------- |  (Normal behavior)
    | ========= Transmits at full rate ========> |
    |                                            |
    |                                            |
    | <========== TCP ACK (rwnd = 16KB) -------- |  (Artificially throttled)
    | ===== Transmits at restricted rate ======> |  (Enforces local priority)

Im et al. leveraged this mechanism in AMUSE (AMUSE: Empowering users for cost-aware offloading with throughput-delay tradeoffs; AMUSE: Empowering Users for Cost-Aware Offloading with Throughput-Delay Tradeoffs) to implement client-side bandwidth throttling and application prioritization on mobile devices. By rewriting the advertisement window in outgoing ACKs, a mobile device can gracefully restrict background downloads to preserve bandwidth for latency-sensitive applications.

This technique is equally effective for enforcing global network policies. At the application layer, Kassa (Hincent: Quick content distribution with priorities and high incentives) demonstrated that manipulating the receive and congestion windows can enforce complex, prioritized max-min fair rates across peer-to-peer content distribution networks without altering the underlying TCP/IP stack. Similarly, Xia et al. (Overhead Control With Reliable Transmission of Popular Packets in Ad-Hoc Social Networks) utilized receiver-side window adjustments to partition bandwidth based on data popularity in ad-hoc networks, proving that a holistic view of competing streams at the destination yields far more accurate link capacity estimation than distributed sender-side protocols.

ACK Pacing, Boosting, and Faking

Beyond modifying window sizes, receivers and intermediate edge nodes can manipulate the timing and generation of ACKs to exploit TCP's self-clocking nature. Because TCP senders rely on the arrival of ACKs to inject new packets into the network, controlling the ACK stream allows the receiver to reshape the sender's behavior.

ACK Pacing

In multi-rate Wi-Fi networks, slow clients operating at low bitrates often monopolize airtime, dragging down the performance of faster clients. To enforce time-based fairness, Tan (Time-based Fairness Improves Performance in Multi-rate WLANs) demonstrated that a wireless access point can strategically delay TCP ACKs destined for specific uplink senders. This trick exploits TCP's self-clocking mechanism to slow down specific flows at the link layer without requiring any modifications to client-side hardware or drivers.

Fake ACKs

Conversely, when legacy TCP interacts with wireless cross-layer schedulers, it can suffer from severe throughput starvation due to transient scheduling delays. Lee et al. (A TCP starvation problem in combining TCP and max-weight scheduling of cross-layer algorithms in WMNs) solved this by introducing a "fake-ACK" algorithm at the mesh access point. When a TCP packet arrives at the access point and moves into the data-link queue, the access point immediately generates a fake ACK back to the sender. This keeps the sender's congestion window open and aligned with the optimal capacity determined by the cross-layer scheduler, even if the physical transmission is temporarily delayed.

ACK Boosting

To accelerate senders and mask network latency over high-delay paths, Xing et al. (NetWarden: Mitigating Network Covert Channels without Performance Loss) utilized "ACK boosting." By proactively generating acknowledgment packets on behalf of the receiver closer to the sender, the system artificially reduces the perceived round-trip time, forcing the sender to ramp up its transmission rate faster than standard network latency would normally permit.

Why It Works: The Power of Destination Context

The unifying insight across these diverse architectures—spanning from datacenter fabrics to application-layer wireless multicast systems like the one proposed by Gupta et al. (Experimental Evaluation of Large Scale WiFi Multicast Rate Control)—is that the receiver is the ultimate aggregation point.

An individual sender only understands its own isolated flow and must infer network state indirectly through drops or delays. In contrast, a receiver (or an edge node sitting immediately in front of it) possesses a complete, real-time view of all competing streams vying for the same local bottleneck link or buffer. Whether by explicitly pacing credits or implicitly guiding legacy TCP via ACK manipulation, controlling traffic from the destination side provides a holistic perspective of network capacity that distributed sender-side algorithms cannot replicate.

Key Foundational Work

A TCP starvation problem in combining TCP and max-weight scheduling of cross-layer algorithms in WMNs (2009) — contributed 6 concept elements to this direction.
AMUSE: Empowering users for cost-aware offloading with throughput-delay tradeoffs (2013) — contributed 2 concept elements to this direction.
Hincent: Quick content distribution with priorities and high incentives (2013) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_holistic_bandwidth_partitioning_bee842a4,
  title = {holistic bandwidth partitioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-bandwidth-partitioning/rdr_bee842a4f5154bbfad9bd86d4c31d0b3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_holistic_bandwidth_partitioning_bee842a4_v1,
  title = {holistic bandwidth partitioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-bandwidth-partitioning/rdr_bee842a4f5154bbfad9bd86d4c31d0b3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake