End-to-End Available Bandwidth: Measurement Methodology, Dynamics, and Relation with TCP Throughput

1
End-to-End Available Bandwidth Measurement
Methodology, Dynamics, and Relation with TCP
Throughput

• Manish Jain
• Constantinos Dovrolis
• SIGCOMM 2002

Presented by Jyothi Guntaka
2
Definitions

• Path capacity C Maximum possible end-to-end
  throughput. It is defined as C mini0H Ci,
  where, Ci is capacity of link i.
• Available bandwidth (termed as avail-bw) Spare
  capacity in the path. In other terms, maximum
  end-to-end throughput given cross traffic load.
  It is a time-varying metric, defined as average
  over a certain time interval.
• Narrow link The link with minimum capacity.
• Tight link The link with minimum available
  bandwidth.

3
Capacity vs. Avail-bw
4
Previous work

• Measure throughput of bulk TCP transfer
• A bulk TCPs throughput is not avail-bw.
• TCP saturates path (i.e., intrusive measurements)
• Carter Crovella dispersion of long packet
  trains (cprobe)
• Ribeiro et al. estimation technique for
  single-queue paths (Delphi)
• Melander et al. attempt to estimate capacity
  avail-bw of every link in path (TOPP)

5
Self-Loading Periodic Streams (SLoPS)

• Basic idea
• Periodic stream (probing packets) which consists
  of K packets of size L at a constant rate R is
  sent from sender to receiver.
• When RgtA, the one-way delays of successive
  packets at the receiver show an increasing trend.

6
SLoPS (2)

• Periodic stream K packets, period T, packet size
  L, rate RL/T

7
SLoPS with Fluid Cross Traffic

• For a path P
• One-way delay (OWD) of packet k
• where is the queue size at link i upon ks
  arrival

SLoPS Stream
Cross Traffic
8
SLoPS with Fluid Cross Traffic (2)

• The OWD difference between two successive packets
  k and k1 is
• where
• Proposition 1 if R gt A, then for
  k1,,K-1. Else, if R lt A, for
  k1,,K-1

9
SLoPS algorithm

• Iterative algorithm
• Sender send a periodic stream n at rate R(n)
• Receiver determine whether or not R(n) gt A
• Receiver notify sender
• If R(n) gt A, R(n1) lt R(n)
• Else, R(n1) gt R(n)
• Specifically
• Initially
• If R(n) gt A, then
• The algorithm terminate when

10
Check with Proposition 1

• A74Mbps (MRTG), R96Mbps (K100packets, T100?s,
  L1200B)

R96 Mbps
R 37 Mbps
11
Refinement of SLoPS algorithm
R82 Mbps

• Refinement
• Watching the increasing
• trend during the entire
• stream
• Accept the possibility of
• variation of A during a
• probing stream, no strict
• ordering between R and A
• which is called
• grey-region

12
PATHLOAD Implementation

• No timing issue consider the variation of OWD
• Parameters
• a stream consists of K packets, each has size L,
  sent at a constant rate R, inter-spacing time T
  L/R,
• Stream duration VKT

13
Detection of increasing OWD trend

• OWD of a stream, can be
• grouped into groups, find
  median in each group , Pathload analyzes
  the set
• Two metrics to determine the trend
• Pairwise Comparison Test (PCT)
• PCT Measures the fraction of consecutive OWD
  pairs that are increasing (between 0 and 1).

14
Detection of increasing OWD trend (2)

• Pairwise Difference Test (PDT)
• PDT Quantifies how strong is the start-to-end
  OWD variation, relative to the OWD absolute
  variations during the stream (between 1 and 1).

15
Fleets of streams

• N streams
• idle time between streams
• Duration of a fleet
• Average rate of a fleet

packets
16
Rate-adjustment algorithm

• If either metrics shows an increasing trend, the
  stream is typed as type-I, otherwise type-N.
• If a fraction f of the streams in a fleet are
  type-I, the fleet has a rate gt A.
• If a fraction f of the streams in a fleet are
  type-N, the fleet has a rate lt A.
• If less than Nf streams are type-I, and also less
  than Nf streams are type-N, then the fleet is in
  grey-region.

17
Grey region

• Measurement stream rate can fall into avail-bw
  variation range.
• Pathload reports grey-region boundaries Gmin,
  Gmax.
• Relative width of grey-region quantify avail-bw
  variability.

18
Experimental Verification

• Simulation scenario
• Path tightness factor

19
Simulation Results

• Pathload produces a range that includes the
  average avail-bw in the path, in both light and
  heavy load conditions at the tight link.

20
Simulation Results (2)

• Pathload estimates a range that includes the
  actual avail-bw in all cases, independent of the
  number of non-tight links or of their load.

21
Simulation Results (3)

• Pathload succeeds in estimating a range that
  includes the actual avail-bw when there is only
  one tight link in the path, but it underestimates
  the avail-bw where there are multiple tight links.

22
Dynamics of Available Bandwidth

• Relative variation metrics
• To compare the variability of the avail-bw across
  different operating conditions and paths.
• Each experiment has 110 runs, plot the
  5,15,,95 percentiles of .

23
Different Load Condition

• Variability of the avail-bw increases
  significantly as the utilization u of the tight
  link increases (i.e., as the avali-bw A
  decreases).

24
Effect of Stream Length K

• Variability of the avail-bw decreases
  significantly as the stream duration increases.

25
Effect of Fleet Length

• As the fleet duration increases, the variability
  in the measured avail-bw increases. Also, as the
  fleet duration increases, the variation across
  different pathload runs decreases.

26
TCP and intrusiveness

• A Bulk Transfer Capacity (BTC) connection using
  TCP can get more bandwidth than what was
  previously available in the path, grabbing part
  of the throughput of other TCP connections.
• Pathload is not intrusive.

27
TCP and intrusiveness (2)
28
TCP and intrusiveness (3)
29
Applications

• Bandwidth-Delay-Product in TCP
• Overlay networks and end-system multicast
• Rate adaptation in streaming applications
• End-to-end admission control
• Server selection and anycasting

30
Comments

• Works well when there is only one tight link.
• Almost all parameters are empirical.
• Could be difficult to tune them under different
  scenarios.
• Difficult to draw general conclusions.
• Difficult to predict converge time.
• In their reported experiments, converge time for
  a single fleet of streams is 10, 30 seconds.
• Not intrusive?
• Only gives a single experiment. Difficult to
  justify.
• How about if lots of users are using pathloads?

31
Acknowledgements

• Some of the slides are taken from
• The presentation by Honggang Zhang
  (http//gaia.cs.umass.edu/measurement/slides/avbw.
  ppt)
• http//lion.cs.uiuc.edu/seminar.ppt

32
Questions?