# CS 176C: Monitoring Systems for Cable Networks [Slides](https://hackmd.io/@arpitgupta/BJyyVG5sU#/) --- ## Learning Objectives - Proactive Network Management (PNM) - What is PNM, what is pre-equalization, etc. - How is PNM data currently used? - CableMon - design goals, rationale, etc. - how it advances state of the art? --- # Background --- ## HFC Architecture  - Key components - CMTS, fiber optical node, (trunk/line) RF amplifiers, splitters, cable modem (not shown) --- ## Proactive Network Maintenance (PNM) - What? Measurement data collected at cable modem (e.g., SNR, FEC stats, etc.) - Why? Proactively detect, localize, and correct impairments issues - DOCSIS standardizes how to - store it in a local MIB - query it using SNMP --- ## PNM Measurement Data - Previous generation: - Data: error rates, SNRs, etc. - Limits: not suited for root-cause analysis - Next generation: - Data: pre-equalization coefficients, full-band capture, etc. - Limits: hard to scale Analysis of pre-equalization coefficients was a game changer! --- ## Pre-Equalization (1) - Objective - Compensate for RF impairments to improve upstream performance - How? - CMTS analyzes the quality of signals received from cable modem - It computes eqaulizer adjustment values to modem! --- ## Pre-Equalization (2)   Cable modem signal with (top) and without (botton) pre-equalization --- ## Equalizer Taps (Simple)  2-Tap Equalizer --- ## Equalizer Taps (DOCSIS 2.0+) 24-Tap Equalizer  - `Pre-main taps` (b-8 -- b-1): lower is better - `Main tap` (b0): higher is better - `Post-main taps` (b1 - b15): lower is better --- ## Understanding Tap Values (1)  DOCSIS Pre-Equalization Tap Values --- ## Understanding Tap Values (2) Frequency Response from Pre-Equalization Data  - Peak-to-valley: 18 dB > Th (0.5 dB) --- ## Understanding Tap Values (3) - `preMTTER`: - Pre-Main Tap to Total Energy Ratio - `postMTTER`: - Post-Main Tap to Total Energy Ratio - `MRLevel`: - Micro-reflection level - `TDR`: - Time domain reflectometer --- ## Understanding Tap Values (4) Pre-Equalization Table After Complex FFT  - postMTTER > Th --> micro-reflection - TDR = ~180 feet --> distance between modem and reflection source --- ## Localizing Faults  - Correlate tap values across modems - Localize faults using TDR data --- ## Facts - Cable broadband is available to 93 % of US homes - other options: DSL (43 %) and Fiber (29 %) - FCC requires 99.99 % availability, but currently we only have 99 % availability - PNM --> diagnose RF impairments --> avoid future outages --> improve availability --- ## Measurement Systems (Collection) - Approach: - periodically collect PNM data (instantaneous) from cable modems (every four hours) - Questions: - Why such low collection frequency? - How to collect/analyse data at high frequency? --- ## Measurement Systems (Analysis) - Approach: - Scoreboarding (Comcast): - per-signal (pre-specified) thresholds - cumulative score across all signals - MTR example (less than 18 dB 26% of time) - Questions: - How to handle noisy (?) data? - How to adapt thresholds over time? --- # CableMon --- ## Data - Collected for 8 months (2019) from 60 K modems - PNM Data - <ts, channel-freq, SNR, Tx/Rx-power, FEC stats, T3/T4 timeouts, pre-equalization coeffs, etc.>---every 4 hours - Customer Complaint Tickets - <customer-id, creation-time, close-time, etc.> --- ## Design Goals - Accuracy - Ticket prediction accuracy: $\frac{CableMon \cap Customer}{CableMon}$ - Ticket coverage: $\frac{CableMon }{Customer}$ - No manual labeling - No extensive parameter tuning - Efficient --- ## Approach - Ticketing rate - average number of customer tickets created in a unit time (4 hours) - Divide PNM metric into bins - For each bin, compute the average ticketing rate ($\frac{N_{b}}{T_{b}}$) Set threshold based on ticketing rate. --- ## Ticketing Rate vs. PNM Data (SNR)  Ticketing rate increases for low SNR values. --- ## Setting Fault Detection Threshold - **Goal** - Minimizes both FPs and FNs - **Observation** - Ticketing rate higher than normal during a faulty periods - **Approach** - Set threshold that maximizes ticketing rate ratio --- ## Feature Selection Select top-K features with higher ticketing rate ratio  --- ## Minimizing False Positives  - Use a moving window of size y (12, 48 hours) - If x (8) data points in the window are abnormal, then send dispatch to fix problem Response can be much slower! --- ## Results (1) Types of tickets detected  CableMon detected more high-severtiy tickets than state-of-the-art systems --- ## Results (2) Distribution of Ticket Life Time (`end-creation time`)  - A longer ticket life time indicates that the problem that triggered the ticket takes a longer time to fix. --- ## Results (3) - Distribution of Report Waiting Time - time difference between when problem started and when customer reported it  - A shorter report waiting time indicates that the problem triggered by the ticket is more urgent. --- ## Summary - Proactive Network Management (PNM) data - What is PNM, how it is used to proactively detect and diagnose RF impairments in the last mile - CableMon - What is CableMon, and how it uses customer complaint tickets to predict process PNM data for inferring RF impairments