Tillered

Technology

This page describes the technical problems Tillered is designed to address and the architectural approach used to improve network performance across distributed systems.

The problem: TCP over long distances

TCP is a reliable, congestion-controlled transport protocol designed for correctness and fairness. While this makes it robust on local and moderately sized networks, it also introduces performance limitations when used over long-distance or high-latency links.

Latency and round-trip time (RTT)

TCP relies on acknowledgements to confirm data delivery. As round-trip time increases, the sender must wait longer to receive acknowledgements before transmitting additional data. This directly limits throughput, even when sufficient bandwidth is available.

TCP pacing over long RTT A client sends data to a server and waits for acknowledgements that return after a round-trip time. Over long distance, the RTT delays feedback and reduces how quickly the sender can increase throughput. Client Sender Server Receiver DATA ACK (feedback) Long RTT delays feedback → slower ramp-up and recovery
Over long distances, ACK feedback arrives later (higher RTT), which limits how quickly TCP can increase its sending rate and recover from loss.

Congestion control behaviour

TCP congestion control algorithms increase sending rates gradually and reduce them sharply in response to packet loss. On long-distance paths, a single loss event can cause significant throughput reduction, followed by slow recovery due to increased RTT.

Bandwidth-delay product

High-latency links require large congestion windows to fully utilise available bandwidth. In practice, default TCP configurations often underutilise long fat networks, resulting in lower-than-expected throughput.

Effective Throughput vs Latency (RTT) Illustrative comparison of a single end-to-end TCP flow versus Tillered optimised transport using segmented paths. 0 2 4 6 8 10 0 50 100 150 200 250 300 Latency (RTT ms) Effective Throughput (Gbps) Standard TCP Tillered (segmented paths)
As round-trip time increases, standard TCP throughput drops sharply due to congestion window constraints. Tillered's segmented paths maintain higher effective throughput by keeping per-segment RTT low.

These effects compound across multiple hops and regions, making distributed systems slower and less predictable as physical distance increases.

Tillered's approach

Tillered addresses these limitations by optimising how traffic is carried between nodes rather than relying on end-to-end TCP behaviour alone.

The system establishes optimised paths between nodes and manages traffic flows based on observed network characteristics. This allows performance to be improved without requiring application-level changes.

Reducing effective transport distance A before-and-after view: a user connecting directly to a server over a long-haul path versus a segmented path through intermediate Tillered nodes, reducing the RTT and loss impact per segment. Before: single long-haul flow After: segmented path through Tillered nodes User source Server destination high RTT, end-to-end loss impact User source Node A Tillered node Node B Tillered node Server destination shortRTT shortRTT shortRTT
A single long-haul flow must retransmit end-to-end when loss or corruption occurs. Segmenting the path through Tillered nodes means each segment can detect and correct errors locally with shorter RTT, recovering faster without impacting the full path.

Segmented paths and locality

By reducing the effective distance of individual transport segments, Tillered limits the impact of RTT and packet loss on any single connection. When packet loss or corruption occurs, the affected segment can detect and retransmit locally rather than requiring a round trip across the entire path. Localised congestion control reacts more quickly and recovers faster than a single long-haul TCP flow.

Continuous measurement

Nodes observe latency, loss, and throughput characteristics over time. This information is used to inform routing and transport decisions, improving predictability under changing network conditions.

Control plane and data plane separation

Tillered maintains a strict separation between control-plane operations and data-plane traffic. Control-plane components coordinate configuration and topology, while data-plane traffic flows directly between nodes.

Control plane versus data plane A diagram showing a control plane coordinating nodes, while data plane traffic flows directly between nodes. The hosted control plane is not in the data path. CONTROL PLANE DATA PLANE Control Plane configuration • coordination • policy Node A site / region Node B site / region control control data-plane traffic (direct)
Control-plane coordination and configuration are separate from data-plane traffic. Data-plane traffic flows directly between nodes and is not proxied through external services.

This separation allows the same core technology to operate in both centrally managed and fully decentralised deployment models.

What Tillered does not do

Tillered is intentionally focused in scope. The system does not:

  • Inspect or modify application payloads
  • Proxy traffic through external services
  • Require application-level integration
  • Depend on continuous internet connectivity

These constraints simplify deployment, reduce trust assumptions, and make behaviour easier to reason about in regulated environments.

Network layer operation

Tillered works at the network infrastructure layer. It accelerates TCP traffic transparently, meaning existing applications and transfer tools continue to work without modification. If you already use specialised transfer software, Tillered provides additional acceleration underneath, complementing application-level optimisations.

Zero-risk integration

Deployment uses policy-based routing configured on your firewall or network hardware to steer selected traffic through Tillered nodes. If a node becomes unavailable, the firewall falls back to the direct path and can load-balance across links -- no traffic is dropped. Tillered nodes do not need to sit on your LAN. They can be deployed in a DMZ or isolated VLAN, with routing handled entirely by the firewall.

Relationship to deployment models

The technology described here is shared across all Tillered deployment models. Differences between Tillered Cloud and Tillered Self-Hosted relate to control-plane placement and operational responsibility, not core functionality.

Read more about deployment models →