(lec 2) CS 536 Park Network Performance - Flashcards

Architecture and design philosophy

• Networks are designed with performance in mind; speed is a premium and slow networks are often not used for practical tasks
  • Cryptographic protocols may be turned off at routers due to overhead
• End-to-end paradigm / lightweight core
  • Push heavyweight processing toward the edge (hosts/servers) and keep the core lightweight
  • This has guided Internet design and evolution historically; many alternative approaches have been tried and failed
• Practical implication: system design aims to minimize cost and overhead while maximizing usable performance at the edge of the network

Performance yardsticks

• Bandwidth (bps)
  • Defined as the bandwidth of physical media; from raw physical layer capacity to bits per second
  • Represents link bandwidth, independent of contention, software overhead, or protocol overhead
• Throughput (bps)
  • Actual data delivered per second end-to-end
  • Includes overhead from software layers: firmware in NICs, device drivers in OS, user-space overhead in applications
  • In practice, app/user-space overhead leads to further slow-down beyond raw link bandwidth
• Latency and delay (ms)
  • Latency: propagation delay plus processing and buffering delay (queueing)
  • Propagation delay depends on distance and signal speed
  • Processing delay includes router/switch computation; buffering delay arises from queues
• Jitter
  • Variation in delay across packets
  • Even if average delay is small, large maximum delays can degrade multimedia performance

Meaning of “high-speed” networks

• Propagation speed bound (SOL) limits how fast a single bit can move; the speed-of-light sets a fundamental bound
  • Approximate signal propagation velocity: v \approx 1.86\times 10^5\ \text{miles/s} \approx 3\times 10^5\ \text{km/s}
  • This bound is independent of the media (optical fiber, copper) and is slower than the ideal SOL
• Lower bound latency example (Purdue to West Coast)
  • Distance: about d = 2000\ \text{miles}
  • One-way propagation lower bound: t_{\text{prop}} \ge \frac{d}{v} = \frac{2000}{186{,}000} \approx 0.0108\ \text{s} \approx 10\ \text{ms}
• Geostationary satellite latency example
  • Distance: approx d \approx 22.2{,}000\ \text{miles}
  • One-way latency: t_{\text{prop}} \approx \frac{d}{v} \approx \frac{22{,}200}{186{,}000} \approx 0.119\ \text{s} \approx 120\ \text{ms}
  • End-to-end (one-way) latency ≈ 120 ms; Round-trip (two-way) ≈ 240 ms; roughly half a second in some scenarios
• Fundamental consequence for applications
  • Some latency is unavoidable due to physics; this constrains interactive and real-time applications

Meaning of high-speed (interpretation)

• A single bit cannot move faster; speed increases by increasing bandwidth (bits per second)
  • Analogy: widening a highway increases throughput by adding more lanes
  • Also called broadband
• Interpretation of high-speed ≈ many lanes; effect on completion time
  • For large files, higher speed shortens completion time
  • For small files, the marginal benefit is smaller
• Internet workload distribution
  • Most files are small, a minority are very large
  • The minority of large files consumes a bulk of network bandwidth

Purdue backbone snapshot (illustrative network picture)

• Purdue’s backbone network features a large set of nodes and interconnections (example device names and links shown in the transcript)
• Link speeds observed in the snapshot include:
  • 1.54 Mb
  • 10 Mb
  • 155 Mb
  • 1 Gbps (1 GigE)
  • 2 Gbps (2 GigE)
  • 10 Gbps (10 GigE)
• Indicates a heterogeneous backbone with regional and campus links, ISP interfaces, and core aggregation
• Document is labeled as “Data Network Version 1.3” and is a NOC copy; illustrates the diversity and scale of a university backbone

Level 3 backbone (Tier-1 ISP) context

• Level 3 (www.level3.com) is a Tier-1 backbone provider; now part of CenturyLink (as of historical context)
• The backbone speed shown in an outdated diagram is 10 Gbps, equal to Purdue’s shown backbone
• Reality (as of the time of the material): faster backbone speeds exist now, including 40 Gbps, 100 Gbps, and 400 Gbps

What is traveling on the wires?

• Traffic mix on networks includes:
  • Bulk data (data, image, video, audio files)
  • Voice
  • Streaming video/audio
  • Real-time interactive data (e.g., games, social media interaction)
  • AI-related traffic
• Most Internet traffic historically has been TCP-based bulk file transfers (data traffic)
• Multimedia streaming has grown rapidly (YouTube, Netflix)
  • Real-time interactive services (VoIP, video conferencing, online gaming) are also significant
• This mix is a driver for traffic management practices, including traffic shaping and “unlimited” data plan policies

Burstiness of traffic (example)

• Bursty traffic example: MPEG-compressed real-time video shows bursty traffic patterns
• Visual intuition: video compression exploits inter-frame dependencies, producing bursts of activity
• Consequence: burstiness is challenging for networks because bursts can exceed average capacity and cause buffering or packet loss

90/10 property: mice and elephants

• Observation: traffic often follows a 90/10 (or 80/20) split
  • 90% of flows are mice (small flows)
  • 10% are elephants (large, bursty flows)
• Why it matters
  • The elephants cause spikes that dominate congestion and delay
  • Active traffic control (e.g., TCP congestion control) may have limited efficacy due to bursts and the disproportionate impact of elephants

How to make sense of all this?

• The study of network performance is approached through three interconnected dimensions:
  • Architecture: system design and how real networks realize those designs in practice
  • Algorithms: how components operate and coordinate (e.g., congestion control, scheduling, routing)
  • Implementation: how algorithms and architectures are realized in hardware and software, including practical constraints and complexities
• Central concern: performance; slow performance means a design is unlikely to be used in practice
• Interdependencies
  • Performance characteristics influence architectural decisions (core vs edge), the algorithms selected (congestion control, QoS), and the implementation details (drivers, firmware, OS stacks)
• Real-world implications
  • Design choices affect latency, throughput, and quality of experience for users
  • Edge-centric designs raise considerations about security, privacy, and reliability at the host/application boundary
  • Traffic shaping and policy decisions impact user experience and service fairness

Equations and numerical references (summary)

• Propagation speed bound (sol):
  • Propagation velocity: v \approx 1.86\times 10^5\ \text{miles/s} \approx 3\times 10^5\ \text{km/s}
• Propagation latency (lower bound):
  • For distance d: t_{\text{prop}} \ge \frac{d}{v}
• Purdue-to-West-Coast example (approx 2000 miles):
  • t_{\text{prop}} \approx \frac{2000}{186{,}000} \approx 0.0108\ \text{s} \approx 10\ \text{ms}
• Geostationary satellite example (distance ≈ 22{,}200 miles):
  • One-way latency: t_{\text{prop}} \approx \frac{22{,}200}{186{,}000} \approx 0.119\ \text{s} \approx 120\ \text{ms}
  • Round-trip latency: ≈ 240\ \text{ms}
• Throughput and bandwidth relationships (conceptual):
  • Throughput is limited by bandwidth and overhead: \text{Throughput} \approx \text{Bandwidth} \times \text{efficiency}
  • File transfer time for size S with end-to-end throughput R: T = \frac{S}{R}

Summary of practical implications

• Physics imposes a hard limit on latency due to propagation speed; higher speeds can only reduce the impact by reducing distance or improving routing efficiency, but the speed of light is a fundamental constraint
• Increasing bandwidth (more lanes) is the primary lever to reduce completion times for large transfers, with diminishing returns for small transfers
• Real networks must manage burstiness and heterogeneous link speeds; a small fraction of large flows (elephants) can dominate congestion behavior
• Effective performance requires integrating architecture, algorithms, and implementation with an understanding of traffic patterns, real-world constraints, and policy considerations
• Edge-centric designs, while beneficial for performance, raise considerations for security, privacy, and reliability that must be addressed in practical systems