Page 1
1
Service Differentiation at Transport Layer via TCP Westwood Low-Priority (TCPW-LP)
H. Shimonishi, M.Y. Sanadidi and M. Geria
System Platforms Research Laboratories, NEC CorporationUCLA Computer Science Department
IEEE Symp on Computers & Communications (ISCC), 2004
Page 2
2
Outline
Introduction TCP Westwood (TCPW) TCP Westwood Low Priority (TCPW-LP) Performance Evaluation
Coexistence with foreground traffic Comparison of TCPW-LP and TCP-LP
Conclusion
Page 3
3
Introduction
TCP Westwood Low-Priority (TCPW-LP) An end-to-end “foreground/background”
priority scheme Objectives
Non-intrusive to coexisting foreground traffic Capable of fully utilizing the unused bandwidth Capable of fairly sharing with other low-priority
flows
Page 4
4
Introduction
Application Web objects pre-fetching (cache) Large bulk transfers, e.g. FTP
Page 5
5
Introduction
Related Works DiffServ (proposed by IETF)
Support from the network router is required End-to-end schemes (TCP-LP and TCP-Nice)
Unused bandwidth cannot be fully utilized Pre-set queuing threshold is required
Page 6
6
Background - TCPW
TCPW – a sender-side only modification Reaction to packet losses
Duplicate ACKs Reno
CWIN = CWIN/2 Westwood
CWIN = (BWE * RTTmin)
Timeout expiration Reno and Westwood
CWIN = 1
Page 7
7
Background - TCPW
BWE – Bandwidth Estimation
Estimated from the rate of ACK b = segment size / (ACKtime - lastACKtime)
segment size = average of last n received segment BWE = αBWE + (1- α)*b
smoothing operator α=0.8
Page 8
8
TCPW-LP
Early Window Reduction (EWR) Congestion window reduction scheme
Dynamic Threshold Adjustment Foreground Traffic Ratio, r
Page 9
9
Early Window Reduction (EWR)
Limit the backlog over the path
Virtual queue length = CWIN – BWE*RTTmin
CWIN = amount of outstanding packets in the path
BWE*RTTmin = amount of packets in the virtual pipe
Page 10
10
Early Window Reduction (EWR)
The virtual queue length exceeds a threshold
CWIN = BWE*RTTmin – BWE*Da
Da – the average queuing delay
BWE*Da – the packets backlogged at the bottleneck
Page 11
11
Dynamic Threshold Adjustment
Foreground Traffic Ratio (FTR), r Ratio of Temporal Minimum Queuing Delay to
Average Queuing Delay When all queued packets belong to foregroun
d traffic r approaches 1
only background flows minimum queuing delay is small due to EWR average queuing delay grows according to the b
acklog threshold
Page 12
12
Dynamic Threshold Adjustment
Dynamic Threshold, Qth = M(1-r) M = 3 (upper bound on backlogged packets)
FTR, r = Dm /(Da+δ) Dm = αDm + (1-α) Dmin
Da = αDa + (1-α) Davg
α= 3/4
δ= 3x10-6/(RTT-RTTmin), ensuring non-zero delay in the calculation of r
Page 13
13
Performance Evaluation
Simulation Topology
End-to-end round trip propagation delay = 74ms
FIFO queuing with drop tail discipline
Page 14
14
Coexistence with foreground traffic
Throughput
Page 15
15
Coexistence with foreground traffic
Congestion Window Behavior
Page 16
16
Coexistence with foreground traffic
Completion time evaluation using FTP traffic
Page 17
17
Coexistence with foreground traffic
Effect of packet losses
Page 18
18
Comparison of TCPW-LP and TCP-LP Throughput
20 identical flows TCP-LP flows utilize only 68% of the link
Page 19
19
Comparison of TCPW-LP and TCP-LP Effect of packet losses
Page 20
20
Comparison of TCPW-LP and TCP-LP Coexistence with UDP traffic
On-off UDP traffic Available Bandwidth = 3.3Mbps(On),
10Mbps(Off) Average available bandwidth = 6.7Mbps
Page 21
21
Comments
Some Questions TCP-LP, one-way delay? Analytical study of Foreground Traffic Ratio? Packet loss improvement? TCP Westwood?
Insight No bandwidth guarantee in both TCPW-LP an
d TCP-LP Protocol between ordinary TCP and TCPW-LP
/TCP-LP Receiver-side only modification scheme
Page 22
22
Conclusion
TCPW-LP – an end-to-end scheme to realize two-class service prioritization
Dynamically adjusting the queuing threshold Evaluation of its performance by simulation Comparison of TCPW-LP and TCP-LP