CapProbe: Inexpensive and Accurate Estimation of Narrow Link - PowerPoint PPT Presentation

CapProbe: Inexpensive and Accurate Estimation of Narrow Link Capacity Rohit Kapoor Ling-Jyh Chen M. Y. “Medy” Sanadidi Mario Gerla

CapProbe: The Main Idea • Observation: Both expansion and compression of dispersion involve queuing due to cross traffic:  Dispersion expansion => second packet queued  Dispersion compression => first packet queued • Packet pair with minimal end-to-end delay sum, is likely to be dispersed corresponding to narrow link capacity • Looking for packet pair with minimal delay sum is inexpensive • CapProbe appears accurate in most of our experiments, simulations and measurements • CapProbe fails under heavy (~>75%) utilization by non- responsive (UDP) traffic December 2003 UCLA Computer Science 2

Preliminary Simulation Results • A packet pair provides two pieces of information  Dispersion between the two packets  End-to-end/Round trip delay of each packet • CapProbe combines both pieces of information  Calculate delay sum for each packet pair sample  Dispersion at minimum delay sum reflects capacity 0.005 0.0045 Minimum Delay Sum (sec) 0.004 0.0035 0.003 0.0025 Capacity 0.002 0.0015 0.001 0.0005 0 0 1.6 3.2 4.8 6.4 8 Bandwidth Estimates (Mbps) December 2003 UCLA Computer Science 3

CapProbe Filtered the Compression • 6-hop path: capacities {10, 7.5, 5.5, 4, 6, 8} Mbps • PP pkt size = 200 bytes, CT pkt size = 1000 bytes • Path-Persistent TCP Cross-Traffic Bandwidth Estimate Frequency Minimum Delay Sums 1 Over-Estimation 0.01 Cross Traffic Rate 0.9 0.009 1Mbps Cross Traffic Rate 0.8 Min Delay Sums (sec) 0.008 2Mbps 1Mbps 0.7 0.007 4Mbps Frequency 2Mbps 0.6 0.006 4Mbps 0.5 0.005 0.4 0.004 0.3 0.003 0.2 0.002 0.001 0.1 0 0 0 1.6 3.2 4.8 6.4 8 0 1.6 3.2 4.8 6.4 8 Bandwidth Estimate (Mbps) Bandwidth Estimate (Mbps) December 2003 UCLA Computer Science 4

CapProbe Filtered the Expansion • PP pkt size = 500 bytes, CT pkt size = 500 bytes • Non-Path-Persistent TCP Cross-Traffic Bandwidth Estimate Minimum Delay Sums Frequency 1 0.0063 1Mbps 0.9 1Mbps 3Mbps 0.8 Min Delay Sum (sec) 3Mbps 0.7 0.0042 Frequency 0.6 Under-Estimation 0.5 0.4 0.0021 0.3 0.2 0.1 0 0 0 1.6 3.2 4.8 6.4 8 0 1.6 3.2 4.8 6.4 8 Bandwidth Estimate (Mbps) Bandwidth Estimate (Mbps) December 2003 UCLA Computer Science 5

CapProbe Failed Under Intensive Non-Responsive or Deterministic Cross Traffic • Non-Path-Persistent UDP Cross Traffic Estimate Frequency Minimum Delay Sums 1 0.014 0.9 1Mbps 1Mbps 0.012 2Mbps 2Mbps Min Delay Sums (sec) 0.8 3Mbps 3Mbps 0.01 0.7 4Mbps 4Mbps Frequency 0.6 0.008 0.5 0.006 0.4 0.3 0.004 0.2 0.002 0.1 0 0 0 1.6 3.2 4.8 6.4 8 0 1.6 3.2 4.8 6.4 8 Bandwidth Estimate (Mbps) Bandwidth Estimate (Mbps) • Only case where CapProbe does not work: Intensive UDP; No correct sample is obtained December 2003 UCLA Computer Science 6

Internet Measurements Laptop1 PING Source/ Each experiment: 500 PP at 0.5s • Laptop3 Destination intervals Dummy Net • 100 experiments for each {Internet path, Internet nature of CT, narrow link capacity} • OS also induces inaccuracy: being fixed! DummyNet % Measurements % Measurements % Measurements Laptop2 Capacity Within 5% of Within 10% of Within 20% of Cross-Traffic Capacity Capacity Capacity 500 kbps Yahoo 100 100 100 1 mbps Yahoo 95 95 100 5 mbps Yahoo 100 100 100 10 mbps Yahoo 60 100 100 20 mbps Yahoo 75 100 100 500 kbps Google 100 100 100 1 mbps Google 100 100 100 5 mbps Google 95 100 100 10 mbps Google 80 95 100 20 mbps Google 65 100 100 December 2003 UCLA Computer Science 7

Wireless Measurements Laptop1 PING Source/ 802.11b Destination Access Point  Bad channel  Internet retransmission  larger dispersions  lower estimated 802.11b capacity Connectivity •Results for Bluetooth-interfered 802.11b, TCP cross-traffic Laptop2 Cross-Traffic Experiment No. Capacity Capacity Estimated Estimated by by strongest mode CapProbe (kbps) (kbps) 1 5526.68 4955.02 2 5364.46 462.8 3 5522.26 4631.76 4 5369.15 5046.62 5 5409.85 449.73 December 2003 UCLA Computer Science 8

CapProbe: Work in Progress • Enhancements for confidence level , and adaptive probing • Probabilistic analysis to determine number of samples required to get min sample, that is probe length • Implementation in OS Kernel to reduce host inaccuracy • Extensive Internet and Abilene (and NLANR testbed?) measurements experiments • Use within TCP: estimating capacity accurately may help December 2003 UCLA Computer Science 9