LambdaStream

1
LambdaStream
OptIPuter Backplane Workshop UCSD, January 25,
2005

• Chaoyue Xiong, Eric He, Venkatram Vishwanath,
• Jason Leigh, Luc Renambot, Tadao Murata, Thomas
  A. DeFanti

2
Motivation

• Application Characteristics
• Continuous data streaming with intensive
  computing
• Environment Characteristics
• High network bandwidth
• End-computers are bottleneckgtCongestion is not
  the only cause for packet loss
• LambdaStream uses these characteristics to
  transfer continuous streaming data to suit for
  the application needs in the above environment

3
Overview of LS (1)

• Combine the advantage of loss-based protocol and
  rate-based protocol
• Compare sending rate and receiving rate to avoid
  congestion
• Estimate bandwidth to fill available bandwidth
• Use congestion-caused packet loss to quickly
  mitigate congestion

4
Overview of LS (2)

• High throughput
• loss type prediction, respond to each type of
  loss with different update algorithm thus
  avoiding unnecessary throttling of the sending
  rate
• Low packet loss
• detect incipient congestion
• estimate available bandwidth
• Low jitter
• keep sending rate as stable as possible

5
Loss Type Prediction

• When packet loss occurs,
• Average receiving interval

• Loss Types
• Continuous decrease in receiving capability
• Occurrence of congestion in the link
• Sudden decrease in receiving capability or random
  loss

6
Incipient undesirable situations avoidance (1)

• When there is no loss, longer receiving packet
  interval indicates link congestion or lower
  receiving capability.

Sender
Bottleneck router
Receiver
?ts
wi
wi1
?tr
7
Incipient undesirable situations avoidance (2)

• Metric
• Ratio between the sending interval and the
  average receiving interval during one epoch.
• Methods to improve precision
• Use weighted addition of receiving intervals from
  the previous three epochs.
• Exclude unusual samples.

8
Single Stream Experiment Result (1)
9
Single Stream Experiment Result (2)
10
Two LS Streams
11
Conclusions

• Suitable for continuous data intensive
  applications.
• Quickly grasp the underlying bandwidth and
  maintain very low jitter.
• Multiple LS streams can reach fairness.
• With TCP The influence of LS on TCP is not
  greater than one TCP on another TCP flow.

12
Future Work

• Development of XIO driver of LambdaStream is
  underway.
• Supports multiple stream mode and integrate with
  TeraVision and SAGE.
• Support dynamic MN communication mode and apply
  to LambdaRAM.
• Use formal modeling (Petri Net) to improve the
  scalability of the algorithm.
• Improve CPU usage efficiency.

13
Questions and Answers

• ?