A Simulation Based Comparison Between HighSpeed TCP and XCP

1
A Simulation Based Comparison Between HighSpeed
TCP and XCP

• Jae Wook Lee and Gleb Chuvpilo
• 6.829 Final Project
• December 6, 2002

2
Talk at a Glance

• Motivation
• Overview of HSTCP and XCP
• Our Results
• Utilization, Fairness, TCP-friendliness,
  Dynamics, Buffering, Deployment
• Conclusion
• Future Work

3
Motivation

• For a TCP throughput of 10 Gbps
• Need cwnd80,000 packets
• Increase cwnd by 1 packet/RTT
• ? Takes thousands of RTT to ramp up to
  full utilization!
• HighSpeed TCP and XCP are two proposed
  replacements
• ? Which one is better and when?
• ? What are the tradeoffs?

4
HighSpeed TCP (HSTCP)

• Modify TCP response function when cwnd is high
  to
• The whole point is that a(w) increases and b(w)
  decreases as cwnd becomes larger.
• Example behavior when cwnd 80,000 packets

5
eXplicit Control Protocol (XCP)

• Routers provide explicit feedback to senders
• No per-flow state is maintained in routers
• Decoupled fairness and efficiency controllers
• Efficiency controller increases aggregate
  feedback proportionally to spare bandwidth and
  decreases proportionally to the persistent queue
  size
• Fairness controller uses AIMD

6
Simulation Topology
Dumbbell TopologynFlows variableRTT100msBW1
00Mbps (variable)qTypeRED/DropTailqSizeBW?Dela
y
SF0
SR0
SF1
SR1
R0
R1
?
?
SFn-1
SRm-1
7
Utilization

• Do HSTCP and XCP achieve high utilization in high
  bandwidth-delay networks?

SETUP RTT(100ms) Bottleneck BW(200Mbps) QsizeB
WRTT
Bottleneck util. vs. variable BW
Bottleneck util. vs. variable RTT
8
Fairness (Jains Index)

• A fairness index proposed by Jain

• For n flows
• - Best Fairness ? F(x) 1
• - Worst Fairness ? F(x) 1/n

SETUP RTT100ms Bottleneck BW200Mbps QsizeBWRT
T
Fairness Index vs. Number of Flows
9
TCP-friendliness

• HighSpeed TCP flow starves a TCP flow even in
  relatively low/moderate bandwidth (e. g. 50Mbps)

Standard TCP flows fair to each other
SETUP RTT100ms Bottleneck BW50Mbps QsizeBWRTT
QtypeDropTail
2 TCP flows
1 HSTCP and 1 TCP flow
10
Dynamics (Convergence to Fairness)
XCP
TCP
HSTCP
Very short convergence time to fairness
Long convergence time to fairness (4000 RTTs)
Long convergence time to fairness (1700 RTTs)
SETUP RTT100ms Bottleneck BW50Mbps QsizeBWRTT
QtypeDropTail A new flow joins at t40s

• In general, HSTCP takes longer time to converge
  to fairness than XCP
• HSTCP may take longer time to converge than TCP,
  too.

11
Dynamics (Fetching BW)
XCP
HSTCP
TCP
Takes 2550 RTTs to grab suddenly available
bandwidth
Immediately grabs suddenly available bandwidth
SETUP RTT100ms Bottleneck BW50Mbps QsizeBWRTT
QtypeDropTail
HSTCP may rely on additive increase to fetch
suddenly available bandwidth like TCP
12
Buffering (Utilization)

• What is the impact of buffering on utilization?

SETUP RTT100ms Bottleneck BW200Mbps NumFlows10
Buffer size is normalized to bandwidth-delay
product
13
Buffering (Fairness)

• What is the impact of buffering on fairness?

HSTCP
XCP
SETUP RTT100ms Bottleneck BW200Mbps Qsize0.1B
WRTT QtypeDropTail/RED
qSize 0.1 BW ? Delay
14
Effects of Reverse Flows

• Two new phenomena present Zhang, et al
• ACK-compression
• Out-of-phase queue-synchronization
• Reverse flows degrades min-max fairness (by
  factor of 2) of HSTCP flows as well as
  utilization.

15
Deployment

• What are the costs and side effects of
  deployment?
• HSTCP
• Easier to deploy end-to-end protocol, no router
  support needed.
• HighSpeed TCP flows starve TCP flows
• XCP
• Harder to deploy requires router support on the
  path.
• Next generation of IP routers may accommodate XCP
  with a benefit of smaller buffering.

16
Conclusion

• Both HSTCP and XCP demonstrate good utilization
  for high bandwidth-delay environments.
• Fairness index XCP gt HSTCP gt TCP, but all close
  to 1.
• HSTCP starves TCP flows even in low-bandwidth
  networks.
• Dynamics tradeoff convergence to fairness vs.
  convergence to full utilization.
• With reasonably small buffer size of
  0.1BWDelay, both HSTCP and XCP work well in
  terms of utilization and fairness. But, XCP
  outperforms HSTCP for both criteria.
• Both HSTCP and XCP face deployment problems
• HSTCP is TCP-unfriendly,
• XCP needs router support on the path.

17
Questions?
18
Future Work

• Evaluate QuickStart
• Obtain results for TCP friendliness of XCP
• Simulate Web traffic
• Explain specific convergence issues with HSTCP
• Analyze ? and ? in XCP

19
Fairness (Dynamics)

• Are protocols scalable in terms of fairness in a
  homogeneous environment of the future?

XCP
HSTCP
20
Dynamics (HSTCP and TCP)
There is a tradeoff between convergence to
fairness and efficiency in achieving full
utilization.

• Jains analysis for HSTCP
• Convergence to fairness condition holds in
  HighSpeed TCP

• Theta(?) analysis
• For ?the angle between a(w) w?b(w)
• small ? ? quickly fetch available bw
• but, could cause slow conv. to fairness

?TCP
?HSTCP
21
Theta analysis (Details)

• In order to converge to fairness fast

• The larger ?D, the better
• The smaller ?I, the better (?Igt45 always)
• Combining 1 2, maximize tan?D / tan?I.