1. Reduced TCP Window Size for VoIP in Legacy LAN Environments Nikolaus Färber, Bernd Girod, Balaji Prabhakar

2. Nikolaus FärberLegacy LAN VoIP Scenario  SOHO with < 30 nodes  Switched, full duplex LAN architecture  10/100 BASE-T, single switch WAN Legacy LAN, IP best-effort T1/DSL/Cable “Last Mile” R S 10 BASE-T... Host IP Phone T2  Problems:  Data (TCP) interferes with voice (UDP)  Queuing delay  Loss

3. Nikolaus FärberLegacy LAN VoIP Goal Control TCP traffic from network edge (T2) such that voice delay is reduced

4. Nikolaus FärberLegacy LAN VoIP Overview  TCP flow control basics  Window based flow control  Bandwidth-delay product  TCP’s congestion avoidance  Bandwidth-delay product  LFNs  LANs  Rule of thumb for setting advertised window size  Results: voice delay and data throughput for  File transfer  LAN at different loads  TCP window control by T2  Limitations and work around

5. Nikolaus FärberLegacy LAN VoIP TCP Flow Control Basics  TCP flow control based on window size W (number of packets source is allowed to send without ACK)  Receiver signals advertised window size W max in ACK  Steady state: W  N = B  D  TCP does not know B, D, N !  Use loss as implicit sign for congestion: B = 40 packet/sec D = 0.1 sec N = 2 connections Increase until loss Back off W = 2 (incremental increase) (multiplicative decrease) tx rx

6. Nikolaus FärberLegacy LAN VoIP B  D for LFNs  “Long Fat Networks” (LFNs) have big B  D requiring big windows  Example: cross-country ATM  B = 155 Mbps  D = 70 ms  Original TCP only supports 64 Kbyte (16 bit filed in header)  New “window scaling” option allows up to 1 Gbyte  Common values still 32-64 KByte B  D = 1.3 MByte

7. Nikolaus FärberLegacy LAN VoIP B  D for LANs  W max for LFNs way too big for LANs!  Example, single Ethernet link:  B  D <! 512 bit  Main delay on switched LAN:  Packet transmission d data = P/B  Store-and-forward operation  Queuing delay d Q  Estimate for low loads:  D < 2  H  P/B  N = 1  Rule of thumb based on W  N = B  D W max = 2  H  P  Typical settings: W max = 4 -16 KByte d data = P/B rx sw tx d Q,fwd d Q,bwd d ack = A/B P = 1500 Byte A = 60 Byte B = 10 Mbps

8. Nikolaus FärberLegacy LAN VoIP File Transfer: Scenario  Voice traffic: UDP, 30 ms, 240 Byte, 10 s  Data traffic: TCP, 8 MB file, 1500 Byte packets, start at 3.5 s  Links: 10/100 Mbps, full duplex, 0.1 ms delay  Switch: 30 KByte buffer, Drop-Tail R S File Server 100BASE-T 10BASE-T Host UDP voice TCP data bottle neck

9. Nikolaus FärberLegacy LAN VoIP File Transfer: Simulation Results 2345678910 0 5 15 20 25 30 2345678910 0 2 4 6 8 12 voice delay [ms] data throughput [Mbps] time [s] ftp start W max = 8 KB W max = 32 KB

10. Nikolaus FärberLegacy LAN VoIP File Transfer: Measurements 2345678910 0 20 30 40 50 60 70 2345678910 0 2 4 6 8 voice delay [ms] data throughput [Mbps] time [s] ftp start W max = 32 KB W max = 8 KB

11. Nikolaus FärberLegacy LAN VoIP LAN: Simulation Scenario  Balanced N-N communication  Traffic model:  File size PDF f F (F): Log-Normal [Arlit 99, Douceur 99]  Idle time T ~ F /  Evaluation:  Voice QoS: 95 percentile of voice delay ( d 95 )  Data QoS: goodput G =  F i /  T i R S 1 2 3 L... load [0,1]  10BASE-T only  Average data rate on each link in each direction is ~  10 Mbps UDP voice TCP data

12. Nikolaus FärberLegacy LAN VoIP LAN: Simulation Results  L =4, B sw = 200 KByte 12345678 0 10 20 30 40 50 60 70 80 90 95 percentile of voice delay, d 95 [ms] data goodput, G [Mbps]  = 0.5 0.4 0.3 0.1 0.2 W max = 32 16 8 4 2 1  Low load  Voice uncritical  Data critical  High load  Voice critical  Data uncritical  W=4 is good choice for all loads!

13. Nikolaus FärberLegacy LAN VoIP LAN: Simulation Result (Cont.)  L =16, B sw = 100 KByte 12345678 0 10 20 30 40 50 60 70 80 90 95 percentile of voice delay, d 95 [ms] data goodput, G [Mbps]  = 0.5 W max = 1 2 4 8 16 32 0.4 0.3 0.1 0.2  General behavior also applies for  L = {4,8,16}  B sw = {100,200, 300}  H = {2,4}  For H =4 (each host has T2) optimal window size is W max = 8

14. Nikolaus FärberLegacy LAN VoIP TCP Window Control by T2  Advertised window size W max is signaled in ACK  T2 can intercept all TCP ACKs and reduce W max before butting it back onto LAN  TCP flows of WAN traffic is not changed  This “packet spoofing” technique is also used by Packeteer Inc. for TCP rate control  No need to modify server/client software  Particular simple for single switch LAN  Even for single T2, the connected host can take full advantage of the technique  Allows gradual deployment without need to use T2 for all hosts

15. Nikolaus FärberLegacy LAN VoIP Limitations  Control from network edge is limited for general LAN topology  T2 cannot control traffic through “remote” switches S2 S1 S3 R UDP voice TCP data

16. Nikolaus FärberLegacy LAN VoIP Work Around  Avoid queuing delays in between switches  Keep inter-switch traffic low  Use faster links for inter-switch connections  generate virtual big switch  Allow communication amoung T2s S2 S1 S3 R UDP voice TCP data

17. Nikolaus FärberLegacy LAN VoIP Conclusions  Reduced TCP window size is advantageous for local data traffic  Reduced voice delay and jitter  Improves data goodput  Rule of thumb for switched LAN: W max = 2  H  P  W=4 [packets] is good choice for single switch LAN  Window size can be spoofed by T2 without awareness of client or server  Control from network edge has inherent limitations  For medium sized LANs, optimization of topology may help