A Measurement Study of Internet Bottlenecks Ningning Hu (CMU) Joint work with Li Erran Li (Bell Lab) Zhuoqing Morley Mao (U. Mich) Peter Steenkiste (CMU) Jia Wang (AT&T) Ningning Hu Carnegie Mellon University 1 Motivation Recent research progress on active probing makes it possible to locate bandwidth bottlenecks 1. How persistent are the Internet bottlenecks? Important for measurement frequency 2. What relationship exists between bottleneck and packet loss and queuing delay? Useful for congestion identification 3. Are bottlenecks shared by end users within the same prefix? Useful for path bandwidth inference 4. What causes intra-AS bottlenecks? Important for traffic engineering Ningning Hu Carnegie Mellon University 2 Pathneck Bottleneck Bottleneck Link: the link with the smallest available bandwidth on a network path Bottleneck Router: the downstream router of a bottleneck link Pathneck An active probing tool that can detect Internet bottleneck location effectively and efficiently For details, please refer to “Locating Internet Bottlenecks: Algorithms, Measurements, and Implications” [SIGCOMM’04] Source code: www.cs.cmu.edu/~hnn/pathneck Pathneck output used in this work Bottleneck link location Route Ningning Hu Carnegie Mellon University 3 Data collection cmu D D D D D D D S 960 Internet Day-1 Destinations Probing Source: a CMU host Destinations: 960 diverse IP addresses 10 continuous probings for each destination (1.5 minutes) Repeat for 38 days (for persistence study) Limitations Pathneck can not cover the last hop 960 << # of Internet paths Ningning Hu Carnegie Mellon University Day-2 … Day-38 4 Outline 1. How persistent are the Internet bottlenecks? Route persistence Bottleneck persistence 2. What relationship is between bottleneck and packet loss and queuing delay? 3. Are bottlenecks shared by end users within the same prefix? 4. What causes intra-AS bottlenecks? Ningning Hu Carnegie Mellon University 5 Terminology probing Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 probing set (persistent) not persistent Consider both AS-level route and locationlevel route Ningning Hu Carnegie Mellon University 6 Route Persistence AS level Location level Route change is very common and must be considered for bottleneck persistence analysis Consistent with the results from Zhang, et. al. [IMW-01] on route persistence Ningning Hu Carnegie Mellon University 7 Bottleneck Persistence Persistence of a bottleneck router # of persistent probing sets R is bottleneck Persist(R) = # of persistent probing sets R appears Bottleneck Persistence of a path Max(Persist(R)) for all bottleneck router R Two views: 1. End-to-end view ― per (src, dst) pair Includes the impact of route change 2. Route-based view ― per route Removes the impact of route change Ningning Hu Carnegie Mellon University 8 Bottleneck Persistence 3 2 1 2 1. Bottleneck persistence in route-based view is higher than end-to-end view 2. AS-level bottleneck persistence is very similar to that from location level 3. 20% bottlenecks have perfect persistence in end-to-end view, and 30% for route-based view Ningning Hu Carnegie Mellon University 9 Outline 1. How persistent are the Internet bottlenecks? 2. What relationship exists between bottleneck and packet loss and queuing delay? 3. Are bottlenecks shared by end users within the same prefix? 4. What causes intra-AS bottlenecks? Ningning Hu Carnegie Mellon University 10 Motivation Possible congestion indication Large queuing delay Packet loss Bottleneck They do not always occur together Packet scheduling algorithm large queuing delay Traffic burstiness or RED packet loss Small link capacity bottleneck Bottleneck ? link loss | large link delay Ningning Hu Carnegie Mellon University 11 Method Collected on the same set of 960 paths, but independent measurements 1. Detect bottleneck location using Pathneck 2. Detect loss location using Tulip Only use the forward path results 3. Detect link queuing delay using Tulip medianRTT – minRTT [ Tulip was developed in University of Washington, SOSP’03 ] Our analysis is based on the 382 paths for which both bottleneck location and packet loss are detected Ningning Hu Carnegie Mellon University 12 Bottleneck Packet Loss Perfectly correlated 30% Ningning Hu 60% |Dist| <= 2 Carnegie Mellon University 13 Bottleneck Link Delay 3% non-bottlenecks have delay > 5ms 15% bottlenecks have delay > 5ms Ningning Hu Carnegie Mellon University 14 More Results 1. How persistent are the Internet There is not much sharing within common cluster bottlenecks? observe clear correlation withbetween link load, while 2. We What relationship exists observing no clear with and link capacity, bottleneck andrelationship packet loss queuing router CPU load, and memory usage. delay? 3. Are bottlenecks shared by end users within the same prefix? 4. What causes intra-AS bottlenecks? Ningning Hu Carnegie Mellon University 15 Related Work Persistence of Internet path properties Zhang [IMW-01], Paxson [TR-2000], Labovitz [TON-1998, Infocom-1999] Congestion points sharing Katabi [TR-2001], Rubenstein [Sigmetrics-2000] Correlation among Internet path properties Paxson [1996] Correlation between router and link properties Agarwal [PAM 2004] Ningning Hu Carnegie Mellon University 16 Conclusion Only 20-30% Internet bottlenecks have perfect persistence Application should be ready for bottleneck location change Bottleneck locations have a fairly strong (60%) correlation with packet loss locations Bottleneck and loss detections should be used together for congestion detection End users within common cluster share bottlenecks only with a low probabilityh End user can not assume common bottlenecks We observe evidence of a correlation between bottleneck and link loads Network engineers should focus on traffic load to eliminate bottlenecks Ningning Hu Carnegie Mellon University 17