PPrate: A Passive Capacity Estimation Tool

1
PPrate A Passive Capacity Estimation Tool
Taoufik En-Najjary 17 march 2006
2
Definition

• Capacity The capacity of a path is defined as
  the maximum IP-layer throughput that a flow can
  get on the path, without any cross traffic.
• Available Bandwidth The maximum rate at which a
  new flow can send without impacting the rate
  achieved by the existing flows.

3
Previous work

• Jacobsons Pathchar
• Relies on ICMP time exceeded messages from
  routers
• Applicability and accuracy?
• Packet Pair Dispersion
• Crovella
• Capacity is reflected by the packet pair
  dispersion that occurs with highest frequency
• Lai
• Filters samples whose dispersion reflects a
  capacity greater than their potential bandwidth
• Both techniques assume unimodal distributions
• Paxson showed distribution can be multimodal

4
Previous work

• Dovrolis Work
• Analysed under/over estimation of Capacity
• Designed Pathrate
• Based on packet pair/train dispersion
• Packet pair estimates Set of possible capacity
  modes
• Packet train estimates ADRLower bound on
  capacity
• Version 1 Capacity (next mode gt ADR)
• Version 2 Capacity (Strongest and narrowest
  mode gt ADR)

5
Drawbacks

• All these techniques are active inject a probe
  packets in the network
• Not suitable for large scale measurement
• Probe overhead !
• Need not to bias the characteristics being
  measured
• Need to access to both the sender and the
  receiver
• Cannot be applied to the large set of Internet
  traces
• A set of accurate, passive capacity
  measurement tool is required

6
PPrate approach

• PPrate is based on packets dispersion principle
• PPrate extracts the capacities from TCP
  connections
• TCP is the most popular transport protocol
• TCP sender often injects packets in pair in the
  network delayed acks startegy
• Work with measurement collected at the receiver
  or the sender side.
• Use as imput packets or acks inter-arrival times
  from TCP connections
• PPrate is based on Pathrate algorithm

7
Receiver side algorithm

• Compute inter-arrivals from packet trace file
• Compute packet pair dispersion B
• Find the modes
• IF there is only one mode, at value M THEN
• output capacity CMSS X 8/M
• Exit
• ELSE
• WHILE Number of mode gt 1 DO
• Group Inter-arrivals N by N
• Compute packet train dispersion B(N)
• N N 1
• END
• Output single mode M
• Compute ADR MSS X8/M
• Find the mode Mc larger than ADR among modes of
  B
• Output narrow link capacity CMSS X 8/Mc
• END

8
Sender side algorithm

• Use the dispersion of pure acks as approximation
  of data packets dispersion
• Cancel the effect of delayed acks by dividing by
  two the inter-arrivals of acks that acknowledge
  two MSS worth of bytes
• Construct the multimodal distribution of the path
• Apply the receiver side algorithm to select the
  mode capacity

9
Limitation?

• TCP Layer
• Receiver advertise window
• Packet train size required is relatively small
• Application on top of TCP
• Application can send data in both directions
  (BitTorrent)
• PPrate can only be applied at the receiver side
• Application can generate idle periods on the TCP
  channel (HTTP)
• Filter Inter-arrival times
• Application can enforce rate limitations
  (E-Donkey)
• In this work we focus on applications that
  should not apply rate limitation FTP, SCP and
  HTTP

10
Validation

• Validation on simulated network
• Cross-traffic effect
• TCP advertise window effect
• Planetlab Validation
• Comparison between PPrate and Pathrate
• The effect of the number of samples
• Application on a large ADSL traces

11
Simulation configuration

• NS-2 simulator, 9 hop path
• 30 second FTP transfer from Source to Sink
• LRD cross traffic
• Pareto sources with alpha 1.9
• Cross-traffic is both path persistent and
  non-persistent

12
Cross-traffic effect
Estimation on Data Inter-arrivals
Estimation on ACKs Inter-arrivals
13
TCP advertise window effect

• Ratio of PPrate estimates made with x and
  without receiver window limitations
• Load ? 50

14
Planetlab validation

• Selection of 33 Paths between Planetlab nodes
• On each path
• SCP transfer of a 20 Mbytes file
• Collection of tcpdump traces at the receiver side
• PlanetLab does not allow collecting traces at
  both end points
• Run Pathrate on the same path
• Repeat procedure 10 times
• It takes few hours for each path
• 15 to 30 minutes for Pathrate to output an
  estimate
• 1 to 5 minutes to perform an scp

15
PPrate vs. Pathrate

• Bandwidth reported by PPrate (o) and Pathrate ()

16
Impact of the number of samples

• Boxplots of ratio between estimates obtained with
  x packets and the ones obtained with the full FTP
  transfer

Version 1 of PPrate
Version 2 of PPrate
17
ADSL traces analysis

• Public ADSL traces http//m2c-a.cs.utwente.nl/re
  pository
• Traffic collected close to ADSL clients
• The only information we have
• The access link capacities range from 256 kbits/s
  to 8 Mbits/s
• We apply PPrate on the ack stream of FTP and HTTP
  connections
• We consider only the ack stream with more than
  300 samples
• 20 of FTP connections
• 8 of HTTP connections

18
Capacities estimated by PPrate
19
Train size needed to the ADR

• 80 of the ADR lt 5 packets
• 93 of tha ADR lt 10 packets
• Advertise window of 65 Kbytes ? 43 packets of
  1500 bytes

N values used to estimate the ADR
20
Conclusion

• PPrate an accurate passive capacity estimation
  tool
• Matlab code of PPrate can be downloaded from
  http//www.eurecom.fr/ennajjar/
• Future work
• Accuracy enhancing Apply PPrate to Bulk transfer
  periods
• Estimation of the capacities before and after the
  measurement point

21
Thanks !
22
Motivation

• Measuring network characteristics is important
  for wide area Internet services.
• ISPs can acquire a good picture about
  characteristics of their own links and links
  outside the boundaries of their network and plan
  their capacity upgrades.
• Multimedia servers would be able to detrmine
  appropriate codecs and streaming rate

23
Packet Pair Dispersion

• Ideal Packet dispersion No cross traffic
• Capacity (Packet Size)/ (Dispersion)

24
Packet Pair Dispersion

• Compression of dispersion
• First packet queuing
• Over estimation
• Expansion of dispersion
• Second packet queues due to Cross Traffic
• Under estimation