• The Congested Curve
  • Background
  • Scenario
• The Hidden Layer
  • Background
  • Scenario
• The Constrained Choice
  • Background
  • Scenario
• The Tradeoff Tunnel
  • Background
  • Scenario
• The Mistaken Sprint
  • Background
  • Scenario
• The Sleepy Socket
  • Background
  • Scenario
• The Queued Crossroad
  • Background
  • Scenario
• The Overcommitted Buffer
  • Background
  • Scenario
• The Ticking Timeout
  • Background
  • Scenario
• The Buffered Bottleneck
  • Background
  • Scenario
• The Idle Interface
  • Background
  • Scenario
• The Overstuffed Switch
  • Background
  • Scenario

50.005 Computer System Engineering
Information Systems Technology and Design
Singapore University of Technology and Design
Natalie Agus (Summer 2025)

Network Performance

The Congested Curve

Background

Recall that network delay is made up of four components:

• Processing delay: time to examine and route the packet at each hop
• Queuing delay: time spent waiting in buffers before transmission
• Transmission delay: time to push the packet onto the link
• Propagation delay: time for bits to travel through the medium

Each hop in a network path contributes to total delay differently depending on hardware, load, and distance. Understanding which delay dominates at each point helps identify bottlenecks and optimize performance.

Scenario

A packet travels across four routers to reach a destination. The network includes a mix of low-latency fiber and high-load shared links. The student is shown a timeline diagram as shown below where each hop is broken into segments of processing, queuing, transmission, and propagation delay. The queuing segment grows significantly at one hop, causing total delay to spike:

Answer the following questions:

• How can you tell which delay is dominating at a hop?
• What conditions cause queuing delay to grow?
• Why does processing delay usually remain constant?
• What network behaviors could flatten or steepen the delay curve?

Hints:

• Queuing depends on arrival rate vs. service rate
• Processing time depends on router speed and configuration
• Transmission is based on packet size and link speed
• Propagation depends only on physical distance
• Bursts or congestion affect queuing the most

The Hidden Layer

Background

Network performance depends not only on bandwidth and propagation, but also on which layers of the system introduce delay. The Internet is modeled as a 5-layer architecture:

1. Application – protocols like HTTP, FTP
2. Transport – TCP, UDP
3. Network – IP
4. Link – Ethernet, Wi-Fi
5. Physical – cables, signals

The operating system is responsible for parts of layers 2–4. The transport layer, including socket buffers and TCP congestion control, is implemented inside the OS kernel. The network layer is handled by routing logic and packet forwarding. The link layer involves device drivers and network interface queues. Delays at these OS-managed layers — such as socket buffering or driver queuing — contribute to total response time but are often overlooked by application developers.

Scenario

An HTTP request from a user appears slow. A network diagnostic shows low propagation delay and high bandwidth. However, further investigation reveals the OS socket buffer is full, and the network interface queue is also growing.

Answer the following questions:

• Which network layers are managed by the operating system?
• What types of delay can occur inside the OS kernel?
• Why does a full socket buffer introduce delay even when the network is fast?
• What tools or measurements help isolate OS-induced delay?
• How can understanding layer boundaries improve performance debugging?

Hints:

• Layers 2–4 interact with kernel code
• Socket buffers and interface queues reside in the OS
• Application-layer delays may result from kernel-side queuing
• Tools like netstat, ss, or tcpdump can observe kernel behavior
• Not all delay is caused by the external network

The Constrained Choice

Background

In multi-hop networks, the bottleneck link (the one with the lowest bandwidth) determines the maximum throughput along a path. To choose the best path, we must compute:

• Transmission delay per link
• Total transmission delay for a packet
• End-to-end throughput, considering bottlenecks

Assume 1000-byte packets are sent from Host A to Host B. Ignore propagation and queuing delay.

Here is the diagram of a multi-path network from A to B, passing through multiple routers with different bandwidths. Note that there are two main paths from A to B.

Scenario

Using the graph above:

• Compute the transmission delay for each link (in ms)
• For each complete path from A to B, compute:
  • Total transmission delay
  • Bottleneck bandwidth
• If you had to send 1000 packets, which path would complete transmission faster?

Answer the following questions:

• What are the possible paths from A to B?
• What is the transmission delay for each link?
• What is the bottleneck link for each path?
• What is the total time to send 1000 packets over each path (pipelined)?
• Which path offers higher throughput and lower latency?

Hints:

• Transmission delay = (packet size × 8) ÷ link bandwidth
• Total transmission delay = sum of all link delays
• Bottleneck = min(link bandwidths in path)
• Pipelined time = 1st packet delay + (999 × bottleneck link delay)
• Prefer paths with higher bottleneck and lower total delay

The Tradeoff Tunnel

Background

In pipelined transmission, packets are sent back-to-back while earlier ones are still in transit. This enables high link utilization but requires understanding both transmission delay and propagation delay. These two combine to form the total time a packet takes to reach the destination.

The bandwidth-delay product (BDP) tells us how many bits can be “in flight” on the link at once. A space-time diagram helps visualize how pipelining works.

In the diagram below:

• Each diagonal line represents a 1000-byte packet
• The x-axis shows direction: from sender to receiver
• The y-axis shows time in milliseconds, growing downward
• Each blue diagonal line shows a packet being transmitted and propagating from sender to receiver.

This format helps visualize packet timing in a way that aligns with packet capture tools or stack traces, where newer events appear lower.

Scenario

A host sends 5 packets (each 1000 bytes = 8000 bits) over a 1 Mbps link with 6 ms one-way propagation delay. The packets are sent back-to-back with no gaps.

Refer to the space-time diagram above for packet timing.

Answer the following questions:

• What is the transmission delay per packet?
• When does the first packet arrive at the receiver?
• When does the last packet arrive?
• What is the total time to complete the transmission of all 5 packets?
• How many bits are in flight just after the third packet finishes sending?

Hints:

• Transmission delay = size ÷ bandwidth
• First packet arrives after transmission + propagation
• Each packet starts right after the previous one
• BDP = bandwidth × propagation delay

The Mistaken Sprint

Background

Transmission delay is the time it takes to push all bits of a packet onto the wire. It depends on the packet size and link bandwidth. Propagation delay is the time it takes for bits to physically travel through the medium (like fiber or copper) from sender to receiver. It depends on distance and the speed of signal propagation — not on packet size.

A common misconception is to confuse these two. In reality, a tiny packet sent across the globe has nearly zero transmission delay but significant propagation delay. Conversely, a large packet on a slow link may experience long transmission delay even over a very short distance.

Scenario

A 1000-byte packet is sent over two different links:

• Link A: 100 meters long, 1 Mbps bandwidth
• Link B: 3000 kilometers long, 1 Gbps bandwidth

Students assume Link A is faster because it is “shorter.”

Answer the following questions:

• What is the transmission delay for each link?
• What is the propagation delay for each link?
• Which link delivers the packet sooner?
• Why does distance dominate in one case, and bandwidth in the other?
• How does this distinction help when analyzing latency?

Hints:

• Transmission = size ÷ bandwidth
• Propagation = distance ÷ signal speed
• Signal speed ≈ 2 × 10⁸ m/s
• Tiny delay differences can become significant at large scale
• Always compute both components to understand total delay

The Sleepy Socket

Background

Network communication often involves blocking system calls like read() or recv(), which wait for data to arrive. If the data has not yet reached the local socket buffer (due to propagation, queuing, or slow sender) the process is put to sleep by the OS. This frees the CPU to run other processes while the kernel waits for incoming data. When the data finally arrives, an interrupt notifies the OS, which wakes the blocked process and schedules it to resume. This mechanism ensures CPU efficiency but introduces performance sensitivity to network delay.

Scenario

A process issues read(sockfd, buf, 100) to receive 100 bytes from a remote server. The network introduces 120 ms of delay before the data reaches the socket buffer. During this time, no CPU is used, and the process is marked as sleeping.

Answer the following questions:

• Why does the OS block the process instead of spinning?
• How is this blocking behavior related to network performance?
• What role does the interrupt mechanism play in this scenario?
• When does the process resume, and how is scheduling involved?
• How can this type of delay be measured or profiled?

Hints:

• Blocking avoids wasting CPU cycles
• Sleep queues allow efficient multitasking
• Arrival of data triggers a hardware interrupt
• Wakeup is followed by re-scheduling
• Tools like strace, perf, or timestamps can expose this delay

The Queued Crossroad

Background

In real networks, intermediate routers often have asymmetric bandwidth between input and output links. When packets arrive faster than they can be forwarded, queuing delay occurs. This happens even if the arrival rate is brief or bursts: the bottleneck at the router determines throughput and latency.

In this setup:

• Sender transmits packets at 10 Mbps
• Router outputs to Receiver at 1 Mbps
• Each packet is 1000 bytes (8000 bits)
• Propagation delay is 1 ms for both Sender→Router and Router→Receiver

The space-time diagram below shows 4 packets being transmitted. You can observe queuing delay as packets pile up at the router before being sent onward.

Notice how packets queue up at the router and experience increasing delays before reaching the receiver.

• Blue lines: Packets traveling from Sender to Router (fast link, low delay).
• Green lines: Packets leaving Router to Receiver (slow link, queuing occurs).

Scenario

The Sender sends 4 packets back-to-back at 10 Mbps. The Router receives them and forwards them one by one at 1 Mbps to the Receiver.

Answer the following questions:

• What is the transmission delay for each hop (Sender→Router, Router→Receiver)?
• When does each packet arrive at the router?
• When does each packet leave the router?
• What is the queuing delay at the router for each packet?
• When does each packet arrive at the receiver?
• What is the total time from first packet sent to last packet received?

Hints:

• Transmission delay = size ÷ bandwidth
• Propagation delay is fixed (1 ms per hop)
• Router sends packets sequentially: no overlap
• Queuing delay = wait time at router before transmission starts

The Overcommitted Buffer

Background

When a network interface card (NIC) receives packets, it transfers them into a fixed-size area in memory called a ring buffer. This buffer is managed by the operating system kernel and the NIC driver. The ring buffer works like a circular queue: each new packet is placed in the next available slot, wrapping around when it reaches the end. The OS must process and remove packets from this buffer before it fills. If the OS is delayed (for example, due to high CPU load or competing processes), the buffer can overflow, and new packets will be dropped.

The OS is notified of incoming packets using hardware interrupts, which trigger the driver to begin processing. To reduce overhead, the OS may use interrupt coalescing or polling mechanisms (like NAPI in Linux) that batch multiple packets before handling them. While this improves CPU efficiency, it can increase delay under bursty or high-throughput conditions. Even if the physical network link has plenty of capacity, performance can degrade if the OS cannot keep up with packet arrival — turning a fast network into a bottleneck.

Scenario

A server receives high-throughput traffic over a fast link. However, due to CPU contention, it cannot process packets immediately. As a result, the NIC ring buffer fills up and packet drops occur. Network performance monitoring shows a sudden spike in application-level latency.

Answer the following questions:

• Where does queuing happen inside the OS when receiving packets?
• What causes buffer overflow even when link capacity is sufficient?
• How does interrupt handling affect packet processing speed?
• Why does high CPU usage worsen network latency?
• What techniques help reduce or absorb incoming bursts?

Hints:

• The NIC uses a ring buffer in kernel memory
• Packet arrival rate can exceed processing rate
• Deferred interrupts or batching can help under load
• Buffer size and CPU availability determine responsiveness
• Drops occur when buffers fill before they are emptied

The Ticking Timeout

Background

Network delay has multiple components. Two fundamental types are:

• Transmission delay: the time to push all bits of a packet onto the wire. It depends on packet size and link bandwidth.
• Propagation delay: the time for a bit to travel across the link, regardless of size. It depends on distance and signal speed.

These two delays dominate in different contexts. In long-distance satellite links, propagation delay is high. In low-bandwidth links, transmission delay dominates. Measuring and comparing them helps explain unexpected latency.

Scenario

Two identical packets are sent: one across a 1 Mbps Wi-Fi link spanning 100 meters, and another across a 10 Gbps transcontinental fiber link (5000 km). Each packet is 1,000 bytes. Assume propagation speed is 2×10⁸ m/s.

Answer the following questions:

• What is the transmission delay for each link?
• What is the propagation delay for each link?
• Which delay dominates in each case?
• How would performance change with larger packet sizes?
• How can systems adapt based on which delay dominates?

Hints:

• Transmission delay = size ÷ bandwidth
• Propagation delay = distance ÷ speed
• Bytes must be converted to bits
• Delays affect buffering and timeout choices
• Applications may batch or pipeline data to reduce overhead

The Buffered Bottleneck

Background

When packets arrive at a network device faster than they can be transmitted, they are queued in a buffer. As the arrival rate gets closer to the link transmission rate, the queuing delay increases. Slowly at first, then sharply. This relationship is non-linear: the delay remains low under light load but rises dramatically as the buffer nears capacity. Understanding this behavior helps explain why networks can feel suddenly “sluggish” even though they are not fully saturated.

Scenario

A router with a 1 Mbps output link handles varying traffic loads. As the average incoming rate increases from 0.1 Mbps to 1.2 Mbps, users notice occasional bursts of latency. This graph plots queuing delay (y-axis) versus traffic arrival rate (x-axis). The curve stays flat until about 0.8 Mbps, then increases steeply, especially near 0.95 Mbps:

Answer the following questions:

• Why is queuing delay low at low arrival rates?
• Why does delay grow rapidly near link capacity?
• What happens when the arrival rate exceeds the link rate?
• How does buffer size affect this curve?
• How could you design systems to handle traffic near capacity?

Hints:

• When arrival rate < link rate, the buffer empties faster than it fills
• Near full load, the buffer starts to build up
• Delay increases non-linearly as service time becomes critical
• Excess packets can be dropped when the buffer overflows
• Techniques like traffic shaping or early drop can manage this growth

The Idle Interface

Background

Link utilization measures how effectively a network link is used. It is defined as the fraction of time the link is actively transmitting data, as opposed to being idle. Low utilization often means the link is underused, possibly due to long propagation delays, small packets, or bursty traffic. Even on high-bandwidth links, poor pipelining or application delays can cause the interface to spend most of its time waiting. Understanding utilization helps diagnose inefficiencies where the network is fast, but performance is still poor.

Scenario

A file transfer sends a single 1500-byte packet every 100 ms over a 10 Mbps link with negligible propagation delay. Students are asked to compute the link utilization and explain why the link feels “slow” despite low latency and decent bandwidth.

Answer the following questions:

• What is the formula for link utilization?
• How long does it take to transmit one packet in this case?
• What is the utilization over each 100 ms interval?
• Why does low utilization lead to underperformance?
• How could the system improve efficiency in this setup?

Hints:

• Utilization = transmission time ÷ total cycle time
• Transmission time depends on packet size and bandwidth
• Idle time dominates if packets are infrequent
• Small or infrequent sends waste bandwidth
• Batching or pipelining can raise efficiency

The Overstuffed Switch

Background

Network switches and routers buffer packets when arrival rate briefly exceeds link capacity. If traffic arrives in bursts, even if the average arrival rate is below capacity, queues can temporarily overflow. This leads to packet drops or increased delay. Queues are stable when the average service rate exceeds the average arrival rate over time. However, burstiness (short, high-volume spikes) can cause instability if the buffer is not large enough or the system is slow to respond.

The graph below illustrates the differences between bursty and smooth traffic, and how bursty traffic can trigger packet loss as the number of arriving packets exceed the router capacity:

Scenario

A switch receives two input streams. Each sends packets in large bursts, followed by idle gaps. The combined average arrival rate is below the link’s output capacity. However, the buffer occasionally overflows, and some packets are lost even though, on paper, the link should handle the load.

Answer the following questions:

• Why can a switch drop packets even if average arrival rate is below capacity?
• What role does burstiness play in queuing instability?
• How does buffer size influence packet loss in this case?
• What mechanisms help absorb bursts without loss?
• How can we balance delay and reliability in queue design?

Hints:

• Instantaneous arrival rate can exceed service rate
• Short-term overload fills the queue before it drains
• Larger buffers delay loss but increase latency
• Traffic shaping and pacing reduce burst impact
• Queue management must balance size and responsiveness

---