# CS 176C: Measuring Broadband Internet Performance <!-- Put the link to this slide here so people can follow --> [Slides](https://hackmd.io/@arpitgupta/BkFDa7RK8?type=slide) --- ## Learning Objectives - Basics - Overview of measurement infrastructure and tools - Throughput measurements - Learn how to measure/analyse throughput - Latency measurements - Learn how to measure/analyse latency - PowerBoost - Basics of PowerBoost --- # Basics --- ## Why? - Affects (almost) all of us - basic necessity - Is unfair (rural vs. urban) - data-driven policy-making - Is complex - one-time measurements are not enough --- ## Where?  Measure at Gateway - Directly measure access link - Longitudinal measurements - Lesser target diversity --- ## How?  - Volunteer-driven recruiting - distribute gateway routers for free (scale?) - Two infrastructures - SamKnows (6-7K+)--still active - BISMark (500+)--dead now --- # Throughput --- ## How to measure? - single-threaded HTTP - multi-threaded HTTP - passive throughput - #bytes/transfer-time - UDP capacity - [Spacerprobe](https://dl.acm.org/doi/abs/10.1145/2068816.2068860) --- ## Measurement Challenges (1)  Througput measurements sensitive to techniques --- ## Measurement Challenges (2)  Packet losses have a significant impact on throughput --- ## Throughput Performance (1) ### What performance do users achieve?  Considerable variation even within a single service plan --- ## Throughput Performance (2) ### Is performance consistent?  Mostly consistent, with some exceptions --- ## Throughput Performance (3) ### Why inconsistent?  Loss increases during peak hours (why?) --- # Latency --- ## Measuring Latencies - How to measure? - Send 600 UDP packets/hour - Send 5 ICMP ping packets/hour - What to measure? - end-to-end path latency - last-mile latency - latency under load --- ## Why Latency Matters?  Page load time **very** sensitive to latency --- ## What about Throughput?  Page load time saturates at around 6 Mbps --- ## Last Mile Latency  - Last-mile is 40-80 % of end-to-end latency - High variance --> shared medium --- ## Latency Under Load ### Oversized buffers - Types - Downstream buffers (CMTS, DSLAM) - Upstream buffers (cable modems, DSL) - Impact of buffers evident only under load (why?) --- ## Prevalence of Oversized Buffers --- Up to two orders of magnitude inflation - Why is inflation more for upload? - Why variability for ATT higher than Comcast? --- ## Understanding Latency Inflations (1) - Measure Broadband America - Experiment     - Use ICMP `ping` (2-10 packets/s)     - After 30 seconds, use `iperf` to flood uplink with UDP traffic (1-10 Mbps)     - Stop `iperf` after 60 seconds     - Measure latency for another 30 seconds --- ## Understanding Latency Inflations (2)  Different modems affect latency inflations differently (why?) --- # PowerBoost --- ## What is PowerBoost? - Marketing term for traffic shaping - Types - Uncapped - simple token bucket - Capped - token + leaky bucket --- ## Powerboost Deep Dive (1) ### Terms - PBS: Powerboost bucket size - MSTR: Maximum sustained data rate - R: capped/uncapped data rate --- ## Powerboost Deep Dive (1) ### Metrics of Interests - Powerboost duration, D = $\frac{PBS}{R-MSRT}$ - Powerboost bytes, B = $R.D$ - Refill time, RT = $\frac{PBS}{MSRT}$ --- ## Powerboost Deep Dive (2) ### Example - Given R = 15 Mbps, MSTR = 12 Mbps, PBS = 12MB - Duration of powerboost? (32 seconds) - Number of bytes sent at higher rate?(60 MB) - Refill time?(8 seconds) --- ## Impact on Throughput (1)  Different modems see powerboost differently - D-LINK modem - 21 Mbps for 3 seconds - 18.5 Mbps for 14s seconds - 12.5 Mbps for remainder duration --- ## Impact on Throughput (2) - D-LINK has two token buckets - B1 $\rightarrow$ B1 - Link capacity: 21 Mbps - Buckets: - (18.5 Mbps, D=3 s) - (12.5 Mbps, D=14s) --- ## Impact on Throughput (2) - Computing PBS - $PBS=D.(R-MSRT)$ - PBS1 = $\frac{3.(2.5)}{8}\sim 1$ MB - PBS2 = $\frac{14.(6)}{8}\sim 10$ MB --- ## Impact on Latency (Under Load)  Lower latency during powerboost --- ## Summary - We learned - Measurement infrastructure - Throughput measurements - Latency measurements - PowerBoost - Next - Deep dive into throughput measurement challenges