1 How is the Internet Performing? Les Cottrell – SLAC Lecture # 2 presented at the 26 th International Nathiagali Summer College on Physics and Contemporary Needs, 25 th June – 14 th July, Nathiagali, Pakistan Partially funded by DOE/MICS Field Work Proposal on Internet End-to-end Performance Monitoring (IEPM), also supported by IUPAP

2 Overview Internet characteristics –packet sizes, protocols, hops, hosts … –complexity, flows, applications Application requirements How the Internet worldwide is performing as seen by various measurements and metrics How well are requirements met? Many sources of measurements CAIDA/Skitter PingER/IEPM Matrix Surveyor

3 Packet size primarily 3 sizes: –close to minimum=telnet and ACKs, 1500 (max Ethernet payload, e.g. FTP, HTTP); ~ 560Bytes for TCP implementations not using max transmission unit discovery Packet size (bytes) Cu,mulative probability % Packets Bytes Mean ~ 420Bytes, median ~ 80Bytes Measured Feb 2000 at Ames Internet eXchange ~ 84M packets, < 0.05% fragmented

4 Internet protocol use There are 3 main protocols in use on the Internet: –UDP (connectionless datagrams, best effort delivery), –TCP (Connection oriented, “guaranteed” delivery) –ICMP (Control Message protocol) Time Feb-May 2001 Flows/10min In Out TCP dominates today SLAC protocol flows TCP UDP ICMP

5 Web use characteristics Size of web objects varies from site to site, server to server and by time of day. –Typical medians vary from 1500 to 4000 bytes Also varies by object type, e.g. medians for –movies few 100KB to MBs, postscript & audio few 100KB –text, html, applets and images few thousand KB Bytes Big peaks for error messages

6 Hops Hop counts seen from 4 Skitter sites (Japan, S. Cal, N. Cal, E. Canada, i.e hops on average Hop Count Weak RTT dependence on hop count 95% 50% 5% RTT Hops

7 Autonomous Systems (AS) Disperson Color indicates the AS responsible for the router at the hop, height is number of probes for that route Seen by Skitter at Palo Alto US (F root name server) Hop number

8 Country dispersion Seen from Japan After 3 to 4 hops most goes to US. –In some cases goes US & back to jp –Some goes to UK & onto other European countries Hops Probes

9 Route maps Simple routes from TRIUMF, Canada to several sites already gets quite complex TRIUMF SLAC KEK UW FNAL DESY CERN

10 Getting more complex PingER Beacon sites in US seen from TRIUMF, Vancouver (from Andrew Daviel, TRIUMF)

11 Connections by country Unknown US UK NL DE IT JP RU

12 Richness of connectivity Angle = longitude of AS HQ in whois records Radius=1-log(outdegree(AS)+1)/(maxoutdegree + 1) –Outdegree = number of next Hops As’ accepting traffic Deeper blue & red more connections All except 1 of top 15 AS’ are in US, exception in Canada Few links between ISPs in Europe and Asia

13 Notes: Many.com are in N. America S. Asia = in (36K), pk (6K), lk, bd E. Asia= jp, cn, my, sg, tw, hk, th, id, bn, mm Mid East= il, kw, lb, ae, tr, sa TLDs with hosts~238 Total TLDs~258 Hosts by regions Jan 2001, 109 Million hosts –Source: Internet Software Consortium ( see web site also for hosts/population

14 Backbone utilization Shows utilization of I2/Abilene backbone links, NB Backbone < 30% loaded Most losses at exchange points & edges

15 Flow sizes Heavy tailed, in ~ out, UDP flows shorter than TCP, packet~bytes 75% TCP-in < 5kBytes, 75% TCP-out < 1.5kBytes (<10pkts) UDP 80% < 600Bytes (75% < 3 pkts), ~10 * more TCP than UDP Top UDP = AFS (>55%), Real(~25%), SNMP(~1.4%) SNMP Real A/V AFS file server

16 Flow lengths 60% of TCP flows less than 1 second Would expect TCP streams longer lived –But 60% of UDP flows over 10 seconds, maybe due to heavy use of AFS at SLAC –Another (CAIDA) study indicates UDP flows are shorter than TCP flows TCP outbound flows Active time in secs Measured by Netflow flows tied off at 30 mins

17 Typical Internet traffic by Application CERFnet link Dominated by WWW (http) WWW FTP RealAudio Mail

18 SLAC Traffic profile SLAC offsite links: OC3 to ESnet, 1Gbps to Stanford U & thence OC12 to I2 OC48 to NTON Profile bulk-data xfer dominates SSH FTP HTTP Mbps in Mbps out Last 6 months 2 Days bbftp iperf

19 SLAC Internet Application usage Ames IXP: approximately 60-65% was HTTP, about 13% was NNTP Uwisc: 34% HTTP, 24% FTP, 13% Napster

20 What does performance depend on? End-to end internet performance seen by applications depends on: –round trip times –packet loss –jitter –reachability –bottleneck bandwidth –implementation/configurations –application requirements Data transmitted in packets

21 Application requirements Based on ITU Y1541 The VoIP loss of 10^-3 used to be 0.25 but that assumed random flat loss –actual loss is often bursty Tail drop in routers Sync loss in circuits, bridge spanning tree reconfiguration, route changes

22 RTT from ESnet to Groups of Sites ITU G ms RTT limit for voice 20%/year RTT ~ distance/(0.6*c) + hops * router delay Router delay = queuing + clocking in & out + processing

23 RTT Region to Region OK White 0-64ms Green ms Yellow ms NOT OK Pink ms Red > 512ms OK within regions, N. America OK with Europe, Japan

24 RTT from California to world Longitude (degrees) 300ms RTT (ms.) Frequency RTT (ms) Source = Palo Alto CA, W. Coast E. Coast US W. Coast US Europe & S. America Europe 0.3*0.6c Brazil E. Coast Data from CAIDA Skitter project

25 Longitude RTT(ms) Seen from Japan RTT from Japan to world

26 Cumulative RTT distributions Gives quality measure Seen from San Diego, US Skitter Steeper = less jitter, i.e. better Small values better RTT ms Cumulative %

27 Routes are not symmetric Min, 50% & 90% RTT measured by Surveyor Notice big differences in RTTs May be due to different paths in the 2 directions or to different loading Advanced to U. Chicago RTT ms U. Chicago to Advanced

28 Loss seen from US to groups of Sites ETSI DTR/TIPHON V1.2.5 threshold for good speech 50% improvement / year

29 Detailed example of improvements Increase of bandwidth by factor of 460 in 6 years, more than kept pace - factor of 50 times improvement in loss Note valleys when students on vacation

30 Loss to world from US Using year 2000, fraction of world’s population/country from

31 How are the U.S. Nets doing? In general performance is good (i.e. <= 1%) ESnet holding steady, still better than others Edu (vBNS/Abilene) &.com improving

32 Losses for 28 days in May 2001 Measured by MIDS to 583 DNS services, 383 Web services, 1367 Internet (ping) hosts, & 1225 ISPs (routers) DNS WWW Internet ISP % Loss

33 Losses between Regions

34 Bulk throughput Important for long TCP flows where we want to copy large amounts of data from one site to another in a relatively short time, e.g. file transfer Depends on RTT, loss, timeouts, window sizes

35 Throughput quality TCP BW < 1/(RTT*sqrt(loss)) Note E. Europe catching up Macroscopic Behavior of the TCP Congestion Avoidance Algorithm, Matthis, Semke, Mahdavi, Ott, Computer Communication Review 27(3), July 1997

36 Throughput also depends on window Optimal window size depends on: –Bandwidth end to end, i.e. min(BW links ) AKA bottleneck bandwidth –Round Trip Time (RTT) –For TCP keep pipe full Window (sometime called pipe) ~ RTT*BW –Can increase bandwidth by orders of magnitude If no loss Throughput ~ Window/RTT Src Rcv ACK t = bits in packet/link speed RTT

37 “Jitter” from N. America to W. Europe “Jitter” = IQR(ipdv), where ipdv(i) =RTT(i) – RTT(i-1) 214 pairs ETSI: DTR/TIPHON V1.2.5 ( ) good speech < 75ms jitter

38 “Jitter” between regions 75ms=Good 125ms=Med225ms=Poor ETSI: DTR/TIPHON V1.2.5 ( ) Jitter varies with loading

39 SLAC-CERN Jitter ETSI/TIPHON delay jitter threshold (75 ms)

40 Reachability Within N. America, & W. Europe loss, RTT and jitter is acceptable for VoIP But what about reachability

41 Reachability – Outage Probability Surveyor probes randomly 2/second Measure time (Outage length) consecutive probes don’t get through Heavy tailed outage lengths (packet loss not Poisson)

42 Europe seen from U.S. 650ms 200 ms 7% loss 10% loss 1% loss Monitor site Beacon site (~10% sites) HENP country Not HENP Not HENP & not monitored

43 Asia seen from U.S. 3.6% loss 10% loss 0.1% loss 640 ms 450 ms 250ms

44 Latin America, Africa & Australasia 4% Loss 2% Loss 350 ms 700ms 170 ms 220 ms

45 Animated monthly % loss 200ms RTT 20% unreachable Big is Bad

46 RTT worldwide from the Matrix

47 More Information IEEE Communications, May 2000, Vol 38, No 5, pp IEPM/PingER home site –www-iepm.slac.stanford.edu/www-iepm.slac.stanford.edu/ CAIDA/Skitter home site – Matrix Net home site – Surveyor home site: –