1. IP Network Performance Measurements Bruce Morgan AARNet Pty Ltd

2. Just checking… Why metrics? Metrics are important to identify network related issues especially performance Metrics can be diverse No one metric is suitable for all needs

3. Types of Measurement Active Measurement  Injecting measurement data into the network  E.g. UDP, TCP, ICMP packets Passive Measurement  Measuring what is there already

4. The Problem Measurement of the network cloud is difficult – but is essential if we are to gauge user perception of the internet

5. The World Wide Wait Some problems are host based, while others are network based: Physical latency Network queuing and delays Server processing delay Timeouts and packet loss TCP protocol delays

6. The Dark Cloud Diverse network paths Asymmetric paths Policy routing Committed Access Rates Firewalls and filters

7. IP Performance Metrics Framework spelt out in RFC 2330 from the IPPM Working Group Goal: “to achieve a situation in which users and providers of Internet transport service have an accurate common understanding of the performance and reliability of the Internet component 'clouds' that they use/provide.”

8. On the Standards track… RFC 2678 IPPM Metrics for Measuring Connectivity RFC 2679 A One-way Delay Metric for IPPM. RFC 2680 A One-way Packet Loss Metric for IPPM. RFC 2681 A Round-trip Delay Metric for IPPM.

9. A One-way Delay Metric Type-P-One-way-Delay The P is for protocol A Poisson distribution is chosen to inject packets Both source and destination require time synchronisation

10. A Round-trip Delay Metric Many applications do not perform well with large end to end delays Ease of deployment compared to one-way metrics Ease of interpretation

11. Ping Two way path measurement based on RTTs (return trip times) Choice of monitored address  Host  Router interface  Router Loopback address

12. Packet Loss on ICMP Loss Asymmetry  Loss = 1 – ((1 – Lossfwd).(1-Lossrcv)) Path Asymmetry Possibility of Internet Service Providers (ISPs) or sites or even hosts rate limiting (including complete blocking) ICMP echo and thus giving rise to invalid packet loss measurements.

13. PingER (Ping End-to-end Reporting) is the name given to the Internet End-to-end Performance Measurement (IEPM) project to monitor end-to- end performance of Internet link Uses ICMP RTT for measurement

14. Surveyor Dedicated PC running Unix at key sites GPS for clock synchronization One way delay & loss measurements Community is Internet 2 clients, HEP sites collaborating with Surveyor

15. PingER/Surveyor Comparison PingER uses the ICMP echo facility (ping) and thus only makes round trip measurements. PingER Surveyor uses a GPS system to synchronise time between sites and makes one way measurements.

16. PingER/Surveyor Comparison Surveyor requires a dedicated platform (PC) to be installed at each site that is monitored, whereas PingER uses an existing host with no special software installed at the monitored site. PingER cheaper!

17. PingER/Surveyor Comparison Surveyor is more accurate and better for short term measurement, especially for sites which have good connectivity. PingER is a more light weight solution, requires less management, uses less bandwidth, requires less storage, and nothing needs to be installed at the remotely monitored sites and is good for remote sites with poor connectivity.

18. PingER/Surveyor Comparison SurveyorPingER Method1 way delay2 way ping Hostsdedicatedselected Frequency~2*2/s~ 0.01/s TimingPoisson bursty (30 min intervals) Monitors~3018 Remotes~30 (~full mesh) ~300 (hierarchical) Pairs~900~1200 Storage~38Mbytes / pair / mo ~ 0.6 Mbytes / pair / mo

19. PingER - Surveyor Complementarity Agree well Surveyor has one way measurements, PingER only round-trip Surveyor dedicated platforms & strong central management experience with PingER shows this has benefits. PingER more parsimonious/lightweight (bandwidth, disk space, cpu) but necessarily less accurate especially at small (hourly) time resolution on low loss links. PingER good for looking at long term trends & grouping where statistics are less a problem

20. TCP SYN / ACK tools In order to truly measure Web traffic, which is almost entirely TCP/IP traffic, it is best to probe using TCP/IP rather than ICMP SYN/ACK mechanism proves useful for this purpose

21. TCP SYN/ACK tools 3 way handshake Send SYN seq=x Receive SYN Send SYN seq=y, ACK x+1 Receive SYN +ACK Send ACK y+1 Receive ACK

22. TCP SYN/ACK Connection request by a SYN and measures the time taken by the target to respond with an ACK The connection is promptly cleared by another exchange of packets, this time containing the FIN control flag.

23. TCP SYN/ACK tools

24. MetricPingSYN/ACK Samples30000 Average161.6 ms158.0 ms Standard Deviation 33.0 ms11.6 ms Median154.4 ms153.0 ms Minimum151 ms150 ms Maximum1222 ms610 ms Lost packets 528 (1.76%)469 (1.56%)

25. TCP SYN/ACK tools

26. Sting Sting is a TCP-based network measurement tool that measures end-to-end network path characteristics. sting is unique because it can estimate one-way properties, such as loss rate, through careful manipulation and observation of TCP behaviour. Avoids increasing problems with ICMP-based network measurement (blocking, spoofing, rate limiting, etc). http://www.cs.washington.edu/homes/savage/sting/

27. Current AARNet Measurements MRTG Perf  ICMP RTT measurements  ICMP Packet Loss measurements Wa  Host/endpoint reachability TCP HTTP file transfer measurements Netflow data

28. MRTG Uses SNMP interface statistics Provides multi-functionality from router temperature to throughput Visualisation package Lacks granularity with time Deployed at each RNO

29. MRTG graphs WARNO/ International traffic on June 18 WARNO / VRNO traffic on June 18

30. Perf Tool Perfd – uses a bsd based ping for RTT and packet Loss calculation Perf – web display tool of the data Deployed at each RNO to measure all points of the mesh Used to check SLA agreement with Cable and Wireless Optus

31. Perf – LA Cable 21 June 2000 ICMP Loss

32. Perf – LA Cable 21 June 2000 ICMP RTT

33. Perf – Optus IA3 21 June 2000 Packet Loss

34. Perf – Optus IA3 21 June 2000 ICMP RTT

35. Perf 6 June Optus international ICMP Loss

36. Perf 6 June Optus international ICMP RTT

37. Perf 6 June ACTRNO ICMP Loss

38. Perf 6 June ACTRNO ICMP RTT

39. WA “what’s alive” is based on nocol Checks reachability of hosts/endpoints Uses ICMP echo, but could be easily extended to check on service level availablity Frequent check of all hosts

40. TCP based Measurements Uses an active http file transfer Measure at host Measure from Netflow records  Can detect retransmissions  These may occur from packet loss/out of sequence packets in either direction

41. Load balancing impacts Can use contiguous IP addresses on monitoring machine to monitor per destination load balancing Monitoring machine can determine performance on link but unable to determine which link is used. If a link fails then traffic will divert to other links

42. Load Balancing – round robin

43. Load Balancing – per packet

44. Load Balancing – 14 May

47. Flows… A flow is taken to be either a bidirectional or unidirectional communication between a source and destination host. The communication shares an address/port correspondence. The biggest indicator of scan/DOS attacks are generally flow records!

48. Netflow Records We keep detailed Flow records  Timestamps and durations  Source/destination addresses  Protocol Types  Cumulative IP Flags  ICMP control types

49. Netflow Records Useful for determining metric targets eg top 100 WWW hosts Can derive useful measurements from the netflow data itself Be wary on derived throughput – flows can take a long time.

50. What are the choices? Various tools and methods are available No one tool is good for everything Combinations of tools, both passive and active, leads to interesting and more detailed analysis

51. AARNet futures… Deployment of measurement machines Monitoring and measuring ICMP, TCP and UDP Monitoring QOS Deploying one-way and round-trip metrics To ensure the network does what its supposed to do…