Latency as a Physical Property of Distributed Systems

Abstract

Network latency is not a software inefficiency; it is a fundamental physical constraint. It represents the immutable time required for data packets to traverse the spatial distance between distributed nodes. While data theoretically travels at the speed of light, propagation delay creates a non-negotiable lower bound proportional to geographical separation. This paper analyzes how unavoidable delays are compounded by network topology, hardware interfaces, and routing fragmentation. We argue that protocols must account for the physical persistence of signals exemplified by the Maximum Segment Lifetime rather than attempting to eliminate the physically impossible.

1. Definitions and Metric Space

To formalize the analysis of delay, we define the following core metrics:

• Latency: The total duration required for a data packet to travel from a source to a destination.

• Round Trip Time (RTT): The elapsed time for a request to reach a destination and for the acknowledgement to return.

• Bandwidth vs. Throughput: Bandwidth is the theoretical maximum capacity of the medium, whereas throughput is the actual successful delivery rate over time.

2. The Physics of Propagation

Latency is bounded primarily by physics.

• Distance: Signals traveling through a medium (fiber optic glass, copper) move slower than the speed of light in a vacuum (approximately 2/3c in fiber). A transcontinental request (e.g., New York to London) incurs tens of milliseconds of delay solely due to distance, regardless of network speed.

• Medium Constraints: Wired connections offer consistency, while wireless transmissions introduce variance due to signal attenuation and environmental interference.

• Immutability: No amount of code optimization can reduce the delay imposed by the physical distance between two points.

3. Network Topology and Path Effects

Packets rarely travel in a straight line. They traverse a complex graph of independent operators.

• The Cost of Hops: At every Internet Exchange Point (IXP), routers must inspect headers, apply Border Gateway Protocol (BGP) policies, and decrement the Time-to-Live (TTL) field.

• Policy over Geometry: Routing is driven by business agreements (peering), not Euclidean geometry. A packet may travel a longer physical path to satisfy a peering contract, increasing RTT.

  
  

• Fragmentation: When packets exceed the Maximum Transmission Unit (MTU), they are fragmented. This increases overhead, as the receiver must wait for all fragments to arrive before reassembly.

4. System-Level Synchronization

Beyond the wire, delay is introduced at the host level to ensure reliability.

• The Handshake Tax: Protocols like TCP require a multi-step handshake (SYN, SYN-ACK, ACK) to establish shared state before a single byte of application data is sent. In an asynchronous system without a global clock, this round-trip synchronization is mandatory.

• Computational Overhead: Context switching, buffer management, and checksum verification are not "inefficiencies" but necessary safety mechanisms for data integrity.

• Maximum Segment Lifetime (MSL): Systems enforce quiet periods to allow "ghost" packets to drain from the network, preventing old data from corrupting new connections.

5. Implications for Observation

The most critical finding for distributed systems design is the "Visibility Lag." Because information takes non-zero time to travel, every observer in a distributed system sees a version of the world that is already in the past.

• Divergent Views: A sender’s view of "sent" data differs from a receiver’s view of "received" data.

• Partial Ordering: Without a global real-time view, systems must rely on sequence numbers and logical clocks to reconstruct causal history.

Core Finding

Latency dictates the topology of value.

While bandwidth can be scaled with capital, latency remains bounded by physics. This constraint forces a structural stratification of financial infrastructure:

• Global Consensus must accept high latency to maintain a unified state.

• High-Frequency Execution must accept localization to minimize physical distance.

There is no single architecture that optimizes for both. The future of distributed systems is not convergence toward speed, but specialized layering based on proximity to the speed of light.