EndtoEnd Available Bandwidth: Measurement Methodology, Dynamics, and Relation with TCP Throughput

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

• M. Jain and C. Dovrolis
• University of Delaware
• SIGCOMM02
• Presented by Yong Yang

2
Motivation

• The available bandwidth in a network path is of
  major importance in
• Congestion control
• Rate adaptation in Streaming application
• QoS verification
• Mirror server selection
• Setting up routes in overlay networks

3
Overview

• 1. Self-Loading Periodic Streams (SLoPS)
• If a stream rate is higher than the available
  bandwidth, the one-way delay between successive
  packets at the receiver show an increasing trend
• Binary search to estimate the available bandwidth
• 2. Measurement Tool PathLoad
• Implementation issues stream parameters,
  detecting an increasing trend, binary search...
• 3. Performance Verification
• 4. Variability of Available Bandwidth
• 5. TCP Throughput and Available Bandwidth

4
1. Self-Loading Periodic Streams (SLoPS)

• Definitions
• Ci capacity of link i
• End-to-end capacity C is limited by narrow link
  n
• ui utilization of link i (0 ? ui ? 1)
• Available bandwidth of link i
• Available bandwidth A is limited by tight link t

n
5
Major Idea (1)

• SLoPS analyzes One-Way Delays (OWDs) of packets
  from sender s to receiver r
• s sends a periodic stream of K packets to r at
  rate R0
• Packet size is L
• OWD from s to r of packet k
• Relative OWDs between two successive packets k
  and k1

6
Major Idea (2)

• If R0gtA, OWDs show an increasing trend
• If R0ltA, OWDs do not show an increasing trend

7
Proof (1)

• 1. At the first link
• Case I R0ltA1
• The arrival rate is R0u1C1 lt A1u1C1C1
• Thus, packet k is served before packet k1
  arrives
• where Ri-1 (Ri) is defined as the entry-rate
  (exit-rate) of the packets at link i

8
Proof (2)

• Case II R0gtA1
• During interval T L/R0,
• The link receives L u1C1T (R0u1C1)T bytes
• The link serves C1T bytes
• So (R0u1C1)T - C1T(R0-A1)T, and
• Packets k1 departs the first link ? time units
  after packet k
• Thus
• Given R0gtA1 and C1gtA1 , we have

9
Proof (3)

• 2. Induction to subsequent links
• For link i
• Also the queueing delay difference

(1)
(2)
10
Proof (4)

• IF R0gtA,
• There exits a link t such that Rt-1 gt At.
  Otherwise R0 R1 RH, hence R0ltAi,
  contradicting R0gtA.
• Thus, , and so the OWD difference
  between successive packets will be positive
  (i.e., )
• IF R0ltA
• Then Ri lt Ai for every link i
• Thus, , and so the OWD difference

11
Iterative Algorithm to Estimate A

• At source Send a periodic stream n with rate
  R(n)
• At receiver
• Measure Di for i1K
• Check for increasing trend in OWDs and notify
  source
• At source
• If trend is increasing (i.e. R(n)gtA), repeat with
  R(n1)lt R(n)
• If non-increasing (i.e. R(n)ltA ), repeat with
  R(n1)gtR(n)
• Terminate if R(n1) R(n) lt??
• ? estimation resolution

12
2. Measurement Tool PathLoad

• Selection of L, T and K
• L can not be less than certain number of bytes,
  to reduce the effect of Layer-2 headers on the
  stream rate
• L should not be greater than path MTU, to avoid
  fragmentation
• T should be small to complete transmission of
  stream before context switch
• Large K may overflow the queue of the tight link
  when R gt A
• Small K does not give enough samples to infer
  trend robustly

13
Use of Several Streams

• N streams allows us to examine N consecutive
  times whether R gt A or not
• Multiple streams, separated by silence period
  allows queues in network to drain measurement
  traffic
• Duration of a fleet

14
Detecting an increasing trend

• Pairwise Comparison Test (PCT)
• EPCT0.5 for independent OWDs,
  when increasing trend
• Pairwise Difference Test (PDT)
• EPDT0 for independent OWDs,
  when increasing trend

15
Illustration of PCT and PDT Metrics

• Infer increasing trend when PCT or PDT trend ? 1.0

16
Rate Adjustment Algorithm
Increasing trend Rmax R(n) R(n1) (Gmax
Rmax)/2 Non-increasing trend Rmin
R(n) R(n1) (Gmin Rmin)/2 Grey region R(n)
gt Gmax Gmax R(n) R(n1) (Gmax Rmax
)/2 Grey region R(n) lt Gmin Gmin
R(n) R(n1) (Gmin Rmin)/2

• Increasing trend, if more than fN streams in a
  fleet are of increasing trend
• Non-increasing trend, if more than fN streams in
  a fleet are of non-increasing trend
• In grey region, otherwise

Terminate if Rmax Rmin lt ? or
Rmax Gmax lt ? and Gmin Rmin lt ?
17
3. Performance Verification

• From Univ-Oregon to Univ-Delaware
• Tight link is a 155Mbps link
• Compare Pathload estimate with MRTG readings
• MRTG is a tool to display the utilized bandwidth
  of a link based on information that comes
  directly from the router interface
• An MRTG reading is a 5-min average
• Poaload report the weighted average of
  consecutive runs over 5 min (each run takes 10-30
  sec)

18

• PathLoad estimate falls within the MRTG range in
  10 out of 12 runs

19
4. Variability of Available Bandwidth (1)

• Relative variation index
• 1. Heavier tight link utilization leads to higher
  available bandwidth variability

20
4. Variability of Available Bandwidth (2)

• 2. Longer probing streams observe lower available
  bandwidth variability, but also can be more
  intrusive

21
5. TCP Throughput and Available Bandwidth

• IETF (Internet Engineering Task Force) recommends
  the Bulk-Transfer-Capacity (BTC) metric to
  characterize the ability of a path to transfer
  large files using TCP
• BTC of a path is the throughput of a persistent
  (bulk) TCP transfer, which is only limited by the
  network resources and not by buffers or end
  systems.

22
BTC vs. PathLoad

• Partition 25-min measurement into 5 intervals
  (A), (B), (C), (D) and (E)
• Perform a BTC connection (PathLoad) during (B)
  and (D)
• BTC can get more bandwidth than what was
  previously available, grabbing part of the
  throughput of other TCP connections
• BTC causes significant decrease in the available
  bandwidth, while PathLoad does not.

23
6. Conclusion

• Basic idea is simple looking at the trend of
  delays of a periodic stream.
• Algorithm is well designed.
• Actual experiments to verify methodology.
• PathLoad does not cause significant increase in
  network utilization.
• Almost all parameters are empirical.