The effect of layer-2 store-and-forward devices. per-hop capacity estimation. Ravi S. Prasad, Constantinos Dovrolis

Transcription

1 The effect of layer-2 store-and-forward devices on per-hop capacity estimation Ravi S. Prasad, Constantinos Dovrolis and Bruce A. Mah University of Delaware Packet Design Ravi S. Prasad - rsp@udel.edu, June of 26

2 Capacity estimation tools Tool Author Measurement objective Methodology bprobe Carter End-to-End Capacity Packet Pairs nettimer Lai End-to-End Capacity Packet Pairs pathrate Dovrolis End-to-End Capacity Packet Pairs & Trains pathchar Jacobson Per-Hop Capacity Variable Packet Size clink Downey Per-Hop Capacity Variable Packet Size pchar Mah Per-Hop Capacity Variable Packet Size pipechar Guojun End-to-End Bottleneck Packet Trains cprobe Carter End-to-End Avail-BW Packet Trains pathload Jain End-to-End Avail-BW Self-Loading Periodic Streams cat Allman Bulk-Transfer-Capacity Standardized TCP throughput IPerf NLANR-DAST Maximum TCP throughput Parallel TCP streams Ravi S. Prasad - rsp@udel.edu, June of 26

3 Ravi S. Prasad - rsp@udel.edu, June of 26 What is the capacity from A to B? A and B both have Fast Ethernet cards. A B FastEthernet Switch A single-hop path

4 And we get... Tool Capacity estimate pathchar Mbps clink Mbps pchar Mbps pipechar Mbps pathrate Mbps bprobe Mbps Ravi S. Prasad - rsp@udel.edu, June of 26

5 Questions What is wrong with pathchar, pchar and clink? The tool? The methodology used? Ravi S. Prasad - rsp@udel.edu, June of 26

6 Overview Variable Packet Size (VPS) methodology Effect of L2 store-and-forward devices Some experimental results Other sources of error Conclusions Ravi S. Prasad - rsp@udel.edu, June of 26

7 d f 1 SND L/C 1 C1 τ 1 f τ 1 r VPS methodology TTL d 2 f L/C 2 C 2 τ f 2 TTL L ICMP / C 1 r d r 1 ==0 ==0 L ICMP ICMP C r 1 τ r 2 C r I I th Hop ICMP : Serialization delay (1) Ravi S. Prasad - rsp@udel.edu, June of 26

8 Ravi S. Prasad - rsp@udel.edu, June of 26 and (3) Linear fit to the minimum RTT obtained experimentally, (2) queuing delays. VPS tools assume : Minimum RTT for each packet size, doesn t include any Components of RTT

9 Minimum RTT (msec) Ravi S. Prasad - rsp@udel.edu, June of Packet size (bytes) 0.25 α α 2 β =0.47usec/byte β =0.84usec/byte First hop (Estimated capacity : 17Mbps) Second hop (Estimated capacity : 22Mbps) 2 An example for 2-hop path

10 L2 segments Ravi S. Prasad - rsp@udel.edu, June of 26 (4) The capacity of L3 hop consisting Different L2 segments may have different capacities May have more than one L2 segment However, If an L3 hop has intermediate L2 devices, Each L3 hop has at-least one L2 segment Each L3 device has an L2 interface Physical Layer Segment 1,1 Segment 2,1 Segment 2,2 Segment 2,3 Layer 2 C1,1 L2 C2,1 L2 Switch C2,2 L2 Switch C L2 2,3 Hop 1 Router Hop 2 Router Layer 3 C 1 L3 C 2 L3 Higher Layers SND Links : Layer3 (L3) vs Layer2 (L2)

11 Ravi S. Prasad - rsp@udel.edu, June of 26 and capacity change relation between increase RTT proportional to the packet size Affect capacity estimated with VPS tools do not generate ICMP packets do not decrease TTL field Can not be detected by upper layers L2 store-and-forward devices

12 L2 store-and-forward devices & serialization delay L2 L /C 1,1 L / L2 C 1,2 L α Probing packet size Minimum RTT SND RCV Layer 2 Device Segment 1,1 Segment 1,2 2 α 1 L2 C 1,1 L2 C 1,2 L2 L /C 1,1 L / L2 C 1,2 (5) (6) Ravi S. Prasad - rsp@udel.edu, June of 26

13 Ravi S. Prasad - rsp@udel.edu, June of 26 (8) The estimated capacity of the hop Path up to hop may include L2 devices. (7) No intermediate L2 store-and-forward devices in the hop Does this error propagate?

14 Ravi S. Prasad - rsp@udel.edu, June of 26 Nominal capacity Mbps pchar Mbps clink Mbps pathchar Mbps Tool Capacity estimate = 50 Mbps = 100Mbps Single-hop path with two L2 segments Experimental results : LAN path

15 Campus paths orion.pc.cis 100Mbps 100Mbps Mbps 100Mbps chp br4 g nss chp br4 f nss Host 1Gbps 100Mbps Incoming Interface FastEthernet Switch chp rt1 v 29.nss chp rt1 v 9.nss Ethernet Switch 10Mbps 10Mbps newark gw chp 7k e 2 4.nss 10Mbps 10Mbps tsunami.coastal Ravi S. Prasad - rsp@udel.edu, June of 26

16 Experimental results : Campus path 1 L3 hop capacity pathchar clink pchar from orion.ps.cis to Mbps from to chp-br4-f nss 100Mbps from chp-br4-f nss to chp-rt1-v-9.nss 100Mbps from chp-rt1-v-9.nss to chp-7k-e-2-4.nss 10Mbps from chp-7k-e-2-4.nss to tsunami.coastal 10Mbps Table 1: Capacity estimates for the path from orion.pc.cis to tsunami.coastal. Ravi S. Prasad - rsp@udel.edu, June of 26

17 Experimental results : Campus path 2 L3 hop capacity pathchar clink pchar from tsunami.coastal to newark-gw 10Mbps from newark-gw to chp-rt1-v-29.nss 10Mbps from chp-rt1-v-29.nss to chp-br4-g nss 1Gbps from chp-br4-g nss to Mbps from to orion.pc.cis 100Mbps Table 2: Capacity estimates for the path from tsunami.coastal to orion.pc.cis. Ravi S. Prasad - rsp@udel.edu, June of 26

18 Experimental results : WAN path 2 L3 hop capacity pathchar clink pchar from abilene-wash-gsr.nss.udel.edu to atla-wash.abilene.ucaid.edu 2480Mbps 460 Table 3: Capacity estimates for an Abilene OC-48 core link. Ravi S. Prasad - rsp@udel.edu, June of 26

19 Other sources of errors in VPS tools Traffic load Non-zero queuing delays Limited clock resolution Error propagation from the previous hop ICMP generation latency? Ravi S. Prasad - rsp@udel.edu, June of 26

20 Ravi S. Prasad - rsp@udel.edu, June of 26 (10) packets The probability of not observing any queuing in hops by at least 1 out of where is the utilization of the link. (9) The probability of not observing any queuing for a packet in link High network traffic High probability of observing queuing delays. Traffic load

21 so that Ravi S. Prasad - rsp@udel.edu, June of 26 too few for remote hops under heavy load VPS tools use same number of probes (default 32) for each hop Table 4: minimum number of packets Path length =0.2 =0.4 =0.6 =0.8 Traffic load (contd.)

22 Ravi S. Prasad - rsp@udel.edu, June of 26 For OC-48, 1 resolution and can result in 25% error. (13) The estimated capacity would be (12) (11) If clock resolution is, Limited clock resolution

23 Ravi S. Prasad - rsp@udel.edu, June of 26 Error in a Gigabit hop after an Ethernet hop gets magnified by a factor of 100 (15) estimated capacity of second hop (14) if measured RTT slopes are Any probabilistic error will propagate to next hop. Error propagation from previous hop

24 ICMP generation latency Latency of ICMP generation not related to probing packet size doesn t affect RTT slope measurement Variation of these latencies may affect RTT slope Minimum ICMP generation latency in high traffic load large number of probes required to catch this effect is similar to that of non-zero queuing delays Ravi S. Prasad - rsp@udel.edu, June of 26

25 Conclusions Methodology used by VPS tools can introduce large errors Errors due to L2 store-and forward devices consistent and hard to identify Probabilistic errors can be detected by repetitive run of the tools Ravi S. Prasad - rsp@udel.edu, June of 26

26 Thank you! Ravi S. Prasad - rsp@udel.edu, June of 26

27 and are minimum queue sizes and Ravi S. Prasad - rsp@udel.edu, June of 26 Non-zero queuing delays cause a multiplicative error in capacity estimate (18) Estimated capacity will be where (17) (16) Minimum RTT measurement for packet sizes Non-zero queuing delays

28 Effect of ATM switches on RTT L L / L / L / L / L /L L Segment 1,2 SND RCV Switch Probing packet size Minimum RTT ATM Without intermediate L2 device c C L2 1,1 C L2 1,2 c C L2 1,1 C L2 1,1 c C L2 1,2 c C L2 1,2 c C L2 1,1 With intermediate L2 device 1 c c Segment 1,1 Ravi S. Prasad - rsp@udel.edu, June of 26

29 Experimental results : WAN path 1 L3 hop capacity pathchar clink pchar from chp-br4-f nss.udel.edu to delaware-gw-f2-0.voicenet.net 45Mbps from delaware-gw-f2-0.voicenet.net to delaware2-gw-h2-0-t3.voicenet.net 45Mbps Table 5: Capacity estimates for the Univ-Delaware access link to VoiceNet, and for a VoiceNet edge link. Ravi S. Prasad - rsp@udel.edu, June of 26