1. Pathload A measurement tool for end-to-end available bandwidth Manish Jain, Univ-Delaware Constantinos Dovrolis, Univ-Delaware Sigcomm 02

2. Overview Capacity and available bandwidth metrics Self-Loading Periodic Streams (SLoPS) methodology Description of pathload Verification experiments

3. Part I Background

4. Definition of capacity Capacity: maximum possible end-to-end throughput End-to-end capacity C is limited by narrow link n: Pathrate: measurement tool based on packet pairs/trains (Infocom’01) »See www.pathrate.org S45 DS3/ATM 100 Fast Ethernet 100 Fast Ethernet R 36(IP) + 9 (ATM)

5. Definition of available bandwidth u i : utilization of link i in time interval T ( 0  u i  1 ) Available bandwidth of link i : Available bandwidth is limited by tight link t: Sink t Source C2C2 C3C3 C1C1 A

6. Measuring per-hop available bandwidth Network managers are very interested in available bandwidth Can be measured at each link from router utilization statistics (via SNMP) - Multi Router Traffic Grapher MRTG graphs: 5-minute averages (each graph one direction) BUT, users do not normally see this data and it is not end- to-end

7. Previous work on available bandwidth estimation 1.Blast the path with UDP traffic (bad and wrong!) 2.Measure throughput of large TCP transfer TCP does not get available bandwidth in under- buffered paths TCP gets more than available bandwidth in over- buffered paths TCP saturates the path (intrusive measurements) 3.Carter & Crovella: Dispersion of packet trains (see cprobe): Does not measure available bandwidth (see Infocom’01) 4.Melander et. al. (Global Internet’00) and Ribeiro et.al.(ITC’00) Correct estimation when queuing only at single link in path

8. Part II Self-Loading Periodic Streams (SLoPS) Methodology

9. Major Idea SLoPS analyzes One-Way Delays (OWDs) of packets from sender S to receiver R OWD: D i = T R arrive -T S send = T arrive - T send + Clock_Offset(S,R) Relative OWDs between successive packets: D i – D i+1 S and R do not have synchronized clocks.

10. Periodic Stream: K packets, size L bytes, rate R = L/T If R>A, OWDs gradually increase due to self-loading of stream D1D2D3 D4 T= L/R 1234 4 123 D1 D2 D3 D4 123 4 K=4 At sender At receiver when R>A At receiver when R<A Basic Idea

11. Experimental result: R > A case K = 100 packets, A= 74Mbps, R=96Mbps, T=100  s

12. K = 100 packets, A= 74Mbps, R=37Mbps, T=100  s Experimental result: R < A case

13. K = 100 packets, A= 74Mbps, R=82Mbps, T=100  s Experimental result: R A case

14. Iterative algorithm in SLoPS At source: Send periodic stream n with rate R(n) At receiver: Measure OWDs D i for i=1…K At receiver: Check for increasing trend in OWDs and notify source At source: If trend is :- increasing (i.e. R(n)>A ),  repeat with R(n+1) < R(n) non-increasing (i.e. R(n) R(n) Terminate if R(n+1) – R(n) <   : resolution of final estimate

15. Part III Description of Pathload

16. Selection of L, T and K L can not be less than certain number of bytes L should not be greater than path MTU, to avoid fragmentation T should be small to complete transmission of stream before context switch Large K may overflow the queue of the tight link when R > A Small K does not give enough samples to infer trend robustly

17. Use of Several Streams N streams allow us to examine for N consecutive times whether R > A or not Disjoint streams, separated by silence period allow queues in network to drain measurement traffic Duration of a fleet of N streams: Effects of stream duration and fleet duration described in a technical report, End-to-end available bandwidth: measurement methodology, dynamics, and relation with TCP throughput

18. How do we detect an increasing trend? Pairwise Comparison Test (PCT): E[PCT]=0.5 for independent OWDs, when increasing trend Pairwise Difference Test (PDT): E[PDT]=0 for independent OWDs, when increasing trend

19. Illustration of PCT and PDT metrics Infer increasing trend when PCT or PDT trend  1.0

20. PCT variation for 3 fleets

21. PDT variation for 3 fleets

22. Rate adjustment algorithm Increasing trend : R max = R(n) R(n+1) = (G max + R max) /2 Non-increasing trend: R min = R(n) R(n+1) = (G max +R min )/2 Grey region & R(n) > G max: G max = R(n) R(n+1) = (G max + R max )/2 Grey region & R(n) < G min: G min = R(n) R(n+1) = (G min + R min )/2 Grey region R max > A R min < A G max G min Terminate if: R max – R min <  or G max – G min < 

23. Detecting sender side context switch Context Switches on sender machine can distort the inter- packet spacing in a stream. Receiver detects the sender side context switches in a stream by testing t(i+1) - t(i) > CS + T

24. Detecting receiver side context switch Receiver side context switch results in packets being accumulated in the buffer in the kernel Receiver detects the receiver side context by checking if a(i+1) - a(i)  0

25. Other pathload features Clock skew between sender and receiver can distort the relative OWD. Clock skew not an issue in pathload due to small stream duration. Pathload aborts the fleet if : –stream encounters excessive loss ( >10 %) –a fraction of streams encounter moderate loss For default tool parameters, and avail-bw  10 Mbps, pathload takes 12 seconds

26. Part IV Experimental Results

27. Verification Approach Use paths from U-Delaware to Greek universities and U- Oregon. Routes through UDel, Abilene, Dante, GRnet MRTG graphs for all links in path report 5-min averages for avail-bw In 5-min interval, pathload runs W times, each for q i secs 5-min average avail-bw R reported by pathload:

28. Verification I Tight link: U-Ioannina to AUTH(C=8.2Mbps), resolution w=1Mbps

29. Verification II Tight link: U-Oregon gigapop-Abilene(C=155Mbps), Resolution w=1 Mbps

30. Summary Avail-bw estimation has numerous application SLoPS: fast, accurate and non-intrusive measurement First release of pathload in Spring’02 Examined avail-bw variability using pathload, and results published in a technical report, End-to-end available bandwidth: measurement methodology, dynamics, and relation with TCP throughput Future work: incorporate avail-bw estimation in transport,QOS and routing