lakshminarayanan

Some Findings on the Network Performance of Broadband Hosts : Some Findings on the Network Performance of Broadband Hosts Karthik Lakshminarayanan, UC Berkeley Venkat Padmanabhan, Microsoft Research


Primary focus of our work : Primary focus of our work Increasing number of broadband hosts in “real world” Understand broadband connectivity: Raw characteristics Applicability of traditional measurement techniques Impact on P2P systems: Find ‘good’ peers in a P2P system Impact on applications, in particular overlay multicast


Related work : Related work Direct measurements: (such as NPD, Nimi) Restricted to “well-connected” hosts Indirect inferencing: (UW study, CMU study) Use peers of file-sharing system as vantage points Most vantage points were “well-connected” machines Inability to measure directly between broadband hosts Our work: Run measurement agents on broadband hosts Perform direct measurements Ability to study properties at a micro-scale


Constraints : Constraints Difficulty in recruiting volunteers: Privacy concerns: Could not measure/use existing traffic Restriction on bandwidth consumption: Imposed a limit of 10 kbps (averaged over few minutes) Cannot obtain login access to machines Run as a Windows service (self-starting daemon) NATs: Used techniques similar to IETF STUN proposal for UDP packets traversing NATs


Design of PeerMetric : Design of PeerMetric PeerMetric Server PeerMetric Clients KeepAlive KeepAlive UDP/ICMP ping Traceroute UDP packet trains TCP transfers HTTP transfers Each PeerMetric client performs some basic P2P tests Intelligence about tests to perform is placed in the Action-generator which runs at the server Simple client-helper to restart the client


Initial deployment of PeerMetric : Initial deployment of PeerMetric Initial PeerMetric deployment had 25 hosts 1 1 MA CT Connection Type Cable modem: 13 DSL: 12 Main ISPs AT&T Broadband: 9 Verizon DSL: 8


Summary of results : Summary of results Confirmation of known results: Asymmetry in bandwidth Median Upstream =212 kbps, Downstream = 900kbps Latency between hosts is high Median of 40ms between hosts in the same city compared to 3-4 ms between well-connected hosts Interesting results: Broadband link “management” affects measurements Delay-vector technique picks proximate peers well P2P latency is a poor predictor of P2P throughput Locality-based heuristics for tree construction perform poorly


#1: Impact of broadband link management : #1: Impact of broadband link management Prevalence of asymmetry is not surprising TCP-Down TCP-Up


#1: Impact of broadband link management : #1: Impact of broadband link management Packet-pair throughput >> TCP throughput Observed for cable modem hosts only TCP-Down TCP-Up PP-Down PP-Up


#1: Impact of broadband link management : #1: Impact of broadband link management Observed only for cable-modem hosts Cable-modem routers perform token-bucket rate limiting Modification of measurement techniques Drain the token bucket before packet-pair measurements Date & Time Type Bandwidth observed (kbps) Pair-1 Pair-2 Pair-3 Pair-4 Pair-5 9-18:15:9:26 PKTPAIR 748.86 744.33 465.19 259.63 242.69 9-18:16:4:16 PKTPAIR 749.01 744.28 531.71 253.68 237.58 9-18:16:47:59 PKTPAIR 751.63 743.00 436.02 245.44 247.86 9-18:15:9:42 TCP 242.62 9-18:16:4:32 TCP 241.72 9-18:16:48:16 TCP 241.88 Measurement techniques have to be revisited for broadband hosts


Peer Selection : Peer Selection P2P applications


#2: Peer selection: Latency metric : #2: Peer selection: Latency metric Delay-vector (coordinates) based approach Motivated by GeoPing, GNP Peers ping a set of landmarks and compute a delay vector Delay-vector based approach performs well in finding proximate peers


#3: Peer selection: Throughput metric : #3: Peer selection: Throughput metric Common technique: Ping a set of hosts and pick the best Latency is a poor predictor of TCP throughput (both cable and DSL)


#3: Peer selection: Throughput metric : #3: Peer selection: Throughput metric Using packet-pair to predict TCP throughput (i) Light-weight (ii) Low degree of statistical multiplexing Packet-pair is a good predictor of TCP throughput for DSL


#4: Implications for Overlay Multicast : #4: Implications for Overlay Multicast Geographic clustering: Approximation of network clustering Traditional goal: Mimic IP multicast Minimize repeated traversal of physical links


#4: Implications for Overlay Multicast : #4: Implications for Overlay Multicast Locality-based heuristics for tree construction perform much worse Multicast tree’s root: Symmetric bandwidth: 750 kbps Location: Seattle For achieving delay less than 120ms, max stream is 148kbps Low upstream bandwidth limits out-degree considerably


Conclusions : Conclusions Summary of results: Traditional measurement techniques For example, packet pair techniques need to be revisited Well-accepted design techniques Heuristics for peer selection and multicast tree construction might not work well Some techniques like delay-vector for finding close hosts work well Limitations: Due to operational logistics, we had a modest set of 25 hosts to perform the study