Available bandwidth measurement as simple as running wget

D. Antoniades, M. Athanatos, A. Papadogiannakis, P. MarkatosInstitute of Computer Science (ICS), Foundation for Research & Technology Hellas

(FORTH)

C. DovrolisCollege of Computing, Georgia Institute of Technology

Passive and Active Measurement Conference (PAM) 2006

Presented by Ryan

10 July 2006

Outline

Introduction Background Measurement Methodology

Tool - abget Validation Measurement

Introduction

End-to-end available bandwidth Routing and traffic engineering QoS management Overlay network

Introduction

Existing tools and techniques e.g. pathload, IGI/PTR and Spruce Requiring access at both ends of the

measured path Based on UDP and ICMP protocols

Introduction

New tool – abget Requiring access only at the receiving

host The sender can be any TCP-based server

Working with TCP packets Similar estimation methodology to

pathload

Background

The term “available bandwidth” Several definitions

Link capacity Residual bandwidth Achievable bandwidth

Background

Link capacity Maximum data rate a flow that can utilize

when there are no other traffic flows sharing the link

End-to-end capacity, C C = min{C1,C2,…CN}

Ci is the capacity of link i

Background

Residual bandwidth Unutilized capacity of a path End-to-end available bandwidth, U

.

where is the unutilized capacity, Ci is the capacity and ui(t, t +τ) is the average link utilization (in normalized unit from 0 to 1) in the interval [t, t +τ) of the link i

Adopted in this paper (and pathload)

)],(1[),( ttuCttU iii

)},({min),( ...1 ttUttU iNi

Background

Achievable bandwidth Throughput achievable by a TCP (or TCP-

friendly) flow in passing through a network path

End-to-end achievable bandwidth, A .

di(t,t+τ) – the amount of data received in the interval [t, t+τ) by the receiver from sender i

Adopted in our research work (many-to-one data flow analysis)

N

ii ttdttA

1

),(1

),(

Background

pathload – the basic idea Self-Loading Periodic Streams (SLoPS)

A periodic stream consists of K packets sent to the path at a constant rate R

If R > A (available bandwidth), the one-way delay (OWD) of successive packets at the receiver show an increasing trend

M. Jain and C. Dovrolis, “End-to-End Available Bandwidth: Measurement Methodology, Dynamics, and Relation with TCP Throughput,” IEEE/ACM Transactions on Networking, 11(4):537-549, Aug. 2003.

Background

Detection of an increasing OWD trend Partition measured (relative) OWDs = D1, D2,

…,DK into Г= groups of Г consecutive OWDs

Compute the median OWD of each group More robust to outliers and errors

Pairwise Comparison Test (PCT)

,

An increasing trend if SPCT > 0.55

K

kD̂

1

)ˆˆ(2

1

k

kk

PCT

DDIS

0

1)(XI

if X holds

otherwise

Measurement Methodology

Iterative algorithm similar to SLoPS in pathload pathload – the sender transmits periodic

UDP packet streams at a certain rate abget – TCP-based server sends packets

based on TCP’s flow control and congestion control How to send packet streams at a certain

rate?

Measurement Methodology

The basic idea A limited advertised window, “fake” ACKs Receiver – acknowledges only one MMS

with each ACK and advertises a window of only one MSS

Sender – is forced to send one MMS upon receiving each ACK

Measurement Methodology

To achieve a certain rate R, the “fake” ACKs should be generated periodically with a period T = MSS/R Assumption: ACKs arrived at the sender

periodically

Measurement Methodology

Validation

Measurement Methodology

One-Way Delay (OWD) Estimate from the interarrivals of the received

packets s(i) – the time that the sender transmitted the ith packet r(i) – the time that the receiver got the ith packet o – the clock offset between the two hosts t(i) – the interarrival time between packets i and i-1 at the

receiver d(i) – the OWD of packet i T – the (assumed) constant interarrival time between

packets i and i-1 at the sender

Measurement Methodology

OWD Estimation s(i) = s(i-1) + T r(i) = s(i) + d(i) + o t(i) = r(i) – r(i-1)

d(i) = r(i) – s(i) – o = d(i-1) + t(i) - T

Tool – abget

abget, using an iterative algorithm User specifies

Probing range, [Rmin, Rmax] Estimation resolution, w Stream length parameter, K Number of streams per probing rate, N

Probing starts at rate Rmin, gradually increasing the rate in increments of w until Rmax

Tool – abget

In each iteration Connect to the remote server (web server)

and initiates a download operation Start sending K “fake” ACKs (with a period

corresponds to the desired probing rate) Estimate the OWDs and compute the SPCT

(same as pathload) Repeat the previous process N times

Tool – abget

If more than N/2 of the streams are increasing (non-increasing), the corresponding probing rate is higher (lower) than the available bandwidth

Tool – abget

abget reports a variation range [low_bound, high_bound] Low_bound – max probing rate that was

estimated as lower than the available bandwidth

High_bound – min probing rate that was estimated as higher than the available bandwidth

Validation

Parameters Setting N = 5 K = 50 w = 5Mbps Rmin = 0Mbps

Rmax = 100Mbps

Ti = 500ms

Measurement Duration ~ 50s

Validation

In local testbed Cross Traffic

Constant-rate UDP traffic Realistic traffic trace

Web Server

Cross Traffic Source

Cross Traffic Sink

abget Client

Capacity ~ 97Mbps

Validation

Constant rate UDP traffic Realistic traffic trace

Validation

In the monitored network path

Validation

From www.nytimes.com to UoC client From UoC server to Georgia Tech client

Validation

Robustness to reverse path traffic Forward path – LD ~ 1500B

Reverse path – LA ~ 40B

The ratio LD/LA ~ 40 Few paths have such a high degree of

available bandwidth asymmetry?

Measurement

Measurement in the Internet Client hosts

The University of Crete (UoC), Greece The Georgia Institute of Technology, USA

Web servers www.nero.com (in Germany) www.chez.com (in France)

Measurement is performed every 10 minutes during a 24-hour period

Measurement

Conclusion

Available bandwidth measurement tool – abget Single-end TCP Similar to Pathload

Validations and Measurements in different network paths

Duration and Overhead

Trade-offs between measurement duration, overhead and accuracy Parameters

K – stream length N – number of streams w – estimation resolution Ti – idle time between streams

[Rmin, Rmax] – probing range

Duration and Overhead

Measurement Duration

Measurement Overhead (in term of rate)

)(1minmaxi

avg

TR

MSSKN

w

RR

iavg TRMSSK

MSSK

/)(

2/)( minmax RRRavg

idle time between streams

No. of streams per each probing rateNo. of probing rate

K packets transmission time