Measuring Bandwidth

Measuring Bandwidth K. Lai and M. Baker Presented by Cristina Abad Outline  Motivation  Definitions  Current Techniques  Packet Pair  Bandwidth Measurement Algorithms  Potential Bandwidth Filtering  Simulations  Conclusions Motivation  Knowing the bottleneck bandwidth of a route benefits:       Applications in general Developers of network protocols and applications Network clients Mobile computing Congestion control Multicast routing trees Metrics  Bottleneck Bandwidth  ideal bandwidth of the lowest bandwidth link on that route between two hosts. It is not affected by other traffic.  Available Bandwidth  maximum bandwidth at which a host can transmit at a given point in time along that route. It is limited by other traffic along that route. Current Techniques  Use throughput to approximate bandwidth throughput: amount of data a transport protocol can transfer per unit of time  Other metrics may have significant effect on TCP throughput, while not affecting bandwidth  An application’s throughput to a host implies nothing about other transfers, even from the same application to the same host  Current Techniques  TCP’s technique  Measures the bottleneck router’s buffer size too  Wastes network resources by forcing a dropped packet and filling the router’s buffers  Has to increase its sending rate slowly, or else it will overshoot the real bandwidth and cause massive packet loss Current Techniques  pathchar  Measures bandwidth of every link accurately  Requires special software on only one host  Slow  Can consume significant amounts of network bandwidth Packet Pair Measures true bandwidth of network Does not cause packet loss Does not require many packet roundtrips to work Does not send massive amounts of data Packet Pair  Not statistically robust – kernel density estimator  Not scalable – passive implementation  Slow – gradual packet pair implementation  Not robust on all traffic – Potential Bandwidth Filtering (PBF)  Not flexible to bandwidth changes – window  Difficult to deploy – Receiver Only Packet Pair  Not studied under controlled conditions – Simulation Bandwidth Measurement Algorithms pathchar  Number of different packet sizes:  Total time:  p  s  l where h: number of hops, h i 1 i  MTU s  32   1  li: round trip latency from sender to hop i, p: number of packets sent per size (32)  10-hop Ethernet network,10ms avg latency  144s  Avg bandwidth to probe a hop  Total data transferred: p  h   32s  32 avg packet size 2  round trip latency li s i 2 32i  Bandwidth Measurement Algorithms Packet Pair  Two packets queued next to each other at bottleneck link exit the link t seconds apart:  Assuming constant bottleneck separation: bbnl  bbnl  s2 t bbnl s2 : bbnl: size of second packet bottleneck bandwidth  If a packet queues in between: s2  s0 t s2 t Bandwidth Measurement Algorithms Packet Pair Filtering (MBF)  How to filter noise caused by time compressed and extended packets?    Use mean or median of samples – No! Use histogram to find point of greatest density Kernel density estimator algorithm: gives greater weight to samples closer to the point at which we want to estimate the density  Simple and fast to compute  Makes no assumption about distribution it operates on  Bandwidth Measurement Algorithms Receiver and Sender Based PP  RBPP – t is measured at the receiver: s bbnl  2  SBPP – uses round trip time: bbnl  a2  a1 s2 r2  r1 ai: arrival time of packet i ri: arrival time of ACK of packet i  Filtering techniques can be used to reject incorrect estimates  RBPP is more accurate but harder to deploy  ROPP – sacrifices a little accuracy for ease of deployment Bandwidth Measurement Algorithms Timeliness versus Accuracy  PP usually implemented to run over a fixed number of packets before providing estimate  Solution: Calculate bandwidth gradually    Converges to correct bandwidth within 3 packets Problem: slow to detect bandwidth change Solution: use packet window (size: w); BUT, may reduce stability Potential Bandwidth Filtering  Potential Bandwidth Problem: PP cannot measure a higher bandwidth than that at which the sender sends  PBF  correlate the potential bandwidth and measured bandwidth of a sample in deciding how to filter Simulations Simulations Simulations Simulations