An Analysis of AIMD Algorithm with Decreasing Increases

1
An Analysis of AIMD Algorithm with Decreasing
Increases

• Yunhong Gu, Xinwei Hong, and Robert L. Grossman

National Center for Data Mining
2
Outline

• TCPs inefficiency in grid applications
• Improvements on AIMD
• AIMD with decreasing increases (DAIMD)
• The UDT algorithm
• Experimental result
• Conclusion and future work

3
TCP and AIMD

• AIMD (Additive Increase Multiplicative Decrease)
• Fair max-min fairness
• Stable globally asynchronously stable
• But, inefficient and not scalable
• In grid networks (with high bandwidth-delay
  product)

4
Efficiency of TCP
1 Gb/s link, 200ms RTT, between Tokyo and Chicago
28 minutes
On 10 Gb/s link, 200ms RTT, it will take 4 hours
43 minutes to recover from a single loss.
TCPs throughput model It needs extremely low
loss rate on high bandwidth-delay product
networks.
5
Improvements of TCP

• Fixed parameter (e.g., 1 segment per RTT) is not
  scalable and hence inefficient
• 32 segments per RTT works fine for 1 Gb/s link,
  but how about its performance on 40Gb/s link or
  1.5Mb/s link?
• Increasing the increase parameter as the
  congestion window increases
• E.g., Scalable TCP and HighSpeed TCP
• Cause fairness and convergence problem

6
AIMD with Decreasing Increases

• To reach high efficiency, the increase parameter
  of an AIMD-based algorithm should be correlated
  to the link capacity and the available bandwidth.
• XCP uses available bandwidth and number of
  concurrent flows to calculate next increase
  parameter
• The increase parameter should be large at the
  beginning and decreases as the sending rate
  increases.

7
AIMD with Decreasing Increases
8
UDT - UDP based Transport Protocol

• Application level built above UDP
• End-to-end approach
• Rate based control
• The sending rate is tuned per constant interval
  (SYN).

9
UDT Algorithm

• UDT considers end-to-end link capacity L
• It is hard to estimate the number of concurrent
  flows and real-time available bandwidth
• UDT tunes the increase parameter according to
  L-C, where C is the current sending rate.

10
UDT Algorithm
(1) (2) (3) (4) (5)
11
UDT Algorithm
L 10 Gbps, S 1500 bytes
C (Mbps) L - C (Mbps) Increment (pkts/SYN)
0, 9000) (1000, 10000 10
9000, 9900) (100, 1000 1
9900, 9990) (10, 100 0.1
9990, 9999) (1, 10 0.01
9999, 9999.9) (0.1, 1 0.001
9999.9 lt0.1 0.00067
12
UDT Algorithm
13
UDT Efficiency and Fairness Characteristics

• Takes 7.5 seconds to reach 90 of the link
  capacity, independent of BDP
• Satisfies max-min fairness if all the flows have
  the same end-to-end link capacity
• Otherwise, any flow will obtain at least half of
  its fair share
• Does not take more bandwidth than concurrent TCP
  flow as long as

14
Experiment - Setup
15
Experiment - Results
16
Conclusion

• Standard TCP is inefficient for grid applications
  in high bandwidth-delay product networks.
• We argued that the increase parameter should be
  correlated to such information as link capacity
  and available bandwidth.
• We analyzed a class of AIMD based control
  algorithm whose increase parameter is decreasing
  as the sending rate increases and proved that it
  is fair and stable.
• According to this analysis we designed a new
  control algorithm that uses estimated link
  capacity and the current sending rate as the
  hints to update increase parameter.
• This algorithm has been implemented in our UDT
  protocol and the experiments have demonstrated
  very good performance.

17
Future Work

• Bandwidth Estimation
• Currently UDT uses packet pairs to estimate link
  capacity
• We will consider more methods to deal with cross
  traffic and NIC interrupt coalescence

18
Thank you!

• Questions and comments are welcome!
• For more information, please visit
• http//www.ncdm.uic.edu
• http//udt.sf.net