Performance Evaluation of new TCP variants for HighSpeed Networks

1
Performance Evaluation of new TCP variants for
High-Speed Networks
Graduate Workshop on Networking 2005

• 2005.10.20
• Yusung Kim and Kilnam Chon
• yskim and chon_at_cosmos.kaist.ac.kr

2
Contents

• Introduction
• Objective
• Methodology
• Experimental Results
• Discussion
• Concluding Remarks
• Reference

3
1. Introduction

• TCP Reno eventually underutilizes network
  bandwidth as bandwidth-delay product continues
  to grow 2.
• A TCP connection with 1250Byte packet size and
  100ms RTT is running over a 10Gbps link.

Packet loss
Packet loss
Packet loss
Packet loss
cwnd
10Gbps
big decrease
slow increase
Time (RTT)
Congestion avoidance
Slow start
4
1. Introduction (cont.)

• New TCP variants are studied to address the
  low-utilization problem of TCP Reno.(e.g.
  HSTCP2, STCP3, FAST4, BIC5, CUBIC6)
• While the window growth of new TCP variants is
  scalable, evaluating the overall performance,
  especially regarding the fairness issues, of new
  TCP variants has remained as a major challenge
  6.

5
2. Objective

• We evaluate the important functionalities of new
  TCP variants with various network conditions and
  background traffic on an experimental testbed.
• We do NOT argue what a TCP variant is the best.
  Instead, we try to help users and researchers
  know the strong point and weak point of each
  variant according to network conditions.

6
3. Methodology3.1 Experimental Setup
Sender 1
TCP Variants CUBIC, BIC, FAST, HSTCP, STCP,
RENO
Receiver 1
Sender 2
Receiver 2
Router 2
Router 1
Long-lived flows sender receiver
Background TrafficGeneration
Short-lived flows Sender receiver
Short-lived flows sender receiver
1 Gbps link
Linux
Free BSD Dummynet
7
3.2 Methodology

• Each test is run for 10 minutes.All tests are
  repeated at least 3 times.
• We consider
• RTT in the range of 10ms 300ms,
• bandwidth of 250Mbps and 800Mbps
• queue size at 20 100 of BDP
• There are 3 scenarios of background traffic
• no background traffic,
• small background traffic (around 10 of
  bandwidth)
• big background traffic (more than 50 of
  bandwidth)

8
3.3 Metrics

• Link Utilization
• Intra-Protocol Fairness
• RTT Fairness
• TCP Friendliness
• Multiple Flows with different RTT

9
4. Experiment Results4.1 Link Utilization Test
CUBIC BIC showedgood link utilization. FAST
had burstnessproblem at RTT 300ms.

• Link utilization of a single stream.
• A bottleneck bandwidth is 800 Mbps and queue
  size is 50 of BDP when background traffic is
  around (10) 80 Mbps.

10
4.2 Intra-Protocol Fairness Test
CUBIC BIC showedgood fairness index. FAST
STCP showedunstable results.

• Fairness index of two flows that have a same
  protocol on the same RTT.
• 250 Mbps bandwidth and 20 of BDP queue size
  with background traffic of 2550 Mbps

11
4.3 RTT Fairness Test
CUBIC was the best. HSTCP STCP wasworse than
RENO. FAST seemed to beabnormal.

• Fairness index of two flows when the RTT of the
  first flow is fixed (162ms) and that of the
  second flow is varied.
• 250 Mbps bandwidth and 20 of BDP queue size
  with background traffic of 2550 Mbps

12
4.4 TCP Friendliness Test
CUBIC showed goodTCP Friendliness atshort
RTT. STCP was the most aggressive. FAST seemed
to beabnormal.

• Fairness index between a standard RENO and a
  new TCP variant.
• 250 Mbps bandwidth and 20 of BDP queue size
  with background traffic of 2550 Mbps

13
4.5 Multiple flows with different RTT Test
b. Average packet interval time of
short-lived flows
a. Utilization ratio between multiple new TCP
variant flows and background traffic
14
4.6 Result Summary

• Most new TCP variants had good link utilization
  but still had unfairness aggressiveness
  problem over Reno.
• b. CUBIC had the best RTT Fairness.and good TCP
  friendliness at short RTT too.
• c. FAST used low queue size (low queuing
  delay)and the least aggressive but seemed to be
  sensitive to network conditions.
• d. STCP was the most aggressive and sometimes
  unstable.
• e. CUBIC, BIC, and STCP increased more queuing
  delay than those of Reno, HSTCP, and FAST.

15
5. Discussion

• a. Most new TCP variants showed good link
  utilizationhowever had still unfairness and
  aggressiveness problems over standard TCP Reno.
• b. When we using multiple flows of a new TCP
  variantwith different RTTs, queuing delay can be
  a issue as much as a bandwidth aspect.
• c. Some new TCP variants seemed to be sensitive
  to network conditions.

16
6. Concluding Remarks

• a. We described an experimental testbed and
  metrics to evaluate new TCP variants.
• b. We presented the strong point weak point of
  each new TCP variant according to network
  conditions.
• c. Our works can be the basis of new TCP variants
  evaluation and we plan to elaborate the
  evaluationfurther more.

17
Reference

• 1. V. Jacobson Congestion avoidance and
  control. Proceedings of SIGCOMM 88, ACM, August
  1988
• 2. S. Floyd HighSpeed TCP for Large Congestion
  Windows. IETF RFC3649, December 2003.
• 3. Tom Kelly Scalable TCP Improving
  performance in highspeed wide area networks.
  Submitted for publication, http//www-lce.eng.cam.
  ac.uk/ctk21/scalable/, December 2002
• 4. C. Jin, D. X. Wei and S. H. Low FAST TCP
  Motivation, Architecture, Algorithms,
  Performance. In Proceedings of IEEE INFOCOM 2004,
  March 2004
• 5. L. Xu, K. Harfoush, and I. Rhee Binary
  Increase Congestion Control (BIC) for Fast
  Long-Distance Networks. In Proceedings of IEEE
  INFOCOM 2004, March 2004
• 6. Injong Rhee and Lisong Xu CUBIC A New
  TCP-Friendly High-Speed TCP Variant. PFLDnet
  2005, February 2005
• 7. M. Mathis, J Heffner and R Reddy Web100
  Extended TCP Instrumentation for Research,
  Education and Diagnosis. ACM Computer
  Communications Review, July 2003.
• 8. How to achieve Gigabit speeds with
  Linux.http//datatag.web.cern.ch/datatag/howto/tcp
  .html. 15th August 2005
• 9. J. Ajkat, J. Kaur, FD Smith, and K. Jeffay
  Variability in TCP Roundtrip Times. In SIGCOMM
  Internet Measurement Conference, 2003
• 10. P. Barford, M. Crovella Generating
  Representative Web Workloads for Network and
  Server Performance Evaluation. In SIGMETRIC 1998
• 11. Yee-Ting Li, Douglas Leith and Robert N.
  Shorten Experimental Evaluation of TCP
  Protocols for High-Speed Networks. Technical
  Report, 2005.
• 12 D. Chiu and R. Jain Analysis of the
  Increase and Decrease Algorithms for Congestion
  Avoidance. in Computer Networks, Computer
  Networks and ISDN Systems, vol. 17, pp. 114,
  1989.
• 13 Ravi S. Prasad, Manish Jain, and Constantinos
  Dovrolis On the Effectiveness of Delay-Based
  Congestion Avoidance. In PFLDnet 2004, February
  2004