on efficient on-line grouping of flows with shared bottlenecks at loaded servers by o. younis and s....

23
On Efficient On-line Grouping of Flows with Shared Bottlenecks at Loaded Servers by O. Younis and S. Fahmy Department of Computer Sciences, Purdue University Presented by Felix Lam

Post on 21-Dec-2015

216 views

Category:

Documents


2 download

TRANSCRIPT

On Efficient On-line Grouping of Flows with Shared Bottlenecks at Loaded Servers

by O. Younis and S. Fahmy

Department of Computer Sciences, Purdue University

Presented by Felix Lam

Content

Introduction System Design Performance Evaluation Conclusions Comments and Discussions

Introduction

Bottleneck Sharing is common among different connections

Ordinary TCP is simple but NOT efficient.

Internet Internet

FTP/Web/Streaming Server

Clients

Introduction

Coordinated Congestion Management Flows sharing the same bottleneck(s) are

grouped together for congestion control Each flow in a group is given different qu

ality of service by the server (e.g. different sending rates, different error protection… etc.)

Collective performance improved

Introduction

How to group flows that share the same bottleneck(s) ?

FlowMate Make use of packet delay information to

group flows sharing the same bottleneck(s)

Basic Architecture

The FlowMate module is embedded into the TCP implementation of the sender to collect delay samples for correlation test.

Delay Computation

Packet delay information can be collected in two ways Timestamping ACKs

Option fields supported in TCP implementations in most operating systems. TS Value – time at which the data packet / ACK is sent TS Echo Reply – the previously received TS Value, only present in ACK. TS Recv. Time – extension proposed by the paper Clock-skewness does not affect the correlation values as long as it is constant.

Delay = TS Value (Sender) – TS Recv. Time (Receiver)

Delay Computation

Round Trip Time (RTT) samples For each ACK received, TCP sender get a

RTT sample. However, using RTT samples instead of

forward delay may degrade grouping accuracy because: Bottlenecks of reverse path and forward path may

be different Delayed acknowledgement affect RTT

The reduction in grouping accuracy < 5%

Correlation Test

Correlation test is performed after each flow has got N (e.g. N=6) delay samples

Apply Pearson’s correlation function on the delay samples

Packet Sent Time1s 2s 3s 4s 5s 6s 8s 11s 12s 15s

x1x2 x3 x5 x6y1 y2 y3 y4 y5x4 y6

7s 14s

Correlation Test

Cross-measure Mx Select delay samples from the two flow

s that a packet of flow x must precede a packet of flow y immediately

The correlation of the 3 consecutive pairs of delay samples is computed

Packet Sent Time1s 2s 8s 12s

x1x5y1 y4x4 y6

7s 14s

Correlation Test

Auto-measure Ma

Select pairs of delay samples from one of the two flows that the spacing between each pair is larger than the average spacing between the pairs used to compute Mx

(e.g. 1.33)

Packet

Sent Time

1s 5s 6s 12s

x1x2 x3 x5

Correlation Test

If Mx > MaThen the two flows are grouped together (i.e. they share the same bottleneck(s))

Partitioning

If more than one groups succeed in the test, join the one with highest Mx

Each flow will be repartitioned after it gets enough new delay samples

G1

G2Representative Flow

A new flow

Correlation tests

G3

form a new group if fail

Performance Evaluation

kj is the number of splits of the correct group j. correct: {1,2,3},{4,5,6}

result: {1,2},{3,4,5},{6}Accuracy Index =0.67

flowsTotal

k

flowsTotal

flowssharedfalseIndexAccuracy

Groups

jj

#

)1(

#

#1

#

1

Performance Evaluation

Performance Evaluation

Accuracy Index

Performance Evaluation

Effect of Buffer Size

Performance Evaluation

Effect of packet drop policy

Performance Evaluation

Effect of background traffic load

Performance Evaluation

Performance degradation with bursty Telnet traffic

Conclusions

FlowMate is an on-line partitioning algorithm, requires no active-probing

Robust under heavy background traffic, and different router buffer sizes

High burstiness degrades the performance

Useful for many applications that require inter-flow coordination

Comments and Discussions

The high accuracy, robustness under high background traffic load makes FlowMate quite useful in practical network environments.

Although burstiness degrades performance, it does not affect its usefulness in the scenario of TCP video streaming.

Comments and Discussions

Wonder if varying background traffic load will affect the performance

Any better way to choose the representative flow in each group?