slides by zhangyun

VIEWS: 4 PAGES: 23

									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
                                  1
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

                                                   2
    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
                                                            3
1. Self-Loading Periodic
Streams (SLoPS)
   Definitions
       Ci: capacity of link i
       End-to-end capacity C is limited by narrow link n:
                      C  min{Ci}  Cn
                           i 0.. H
       ui : utilization of link i (0  ui  1)
       Available bandwidth of link i : Ai  Ci(1 - ui)
       Available bandwidth A is limited by tight link t:
               A  min Ai  min Ci (1 - ui)  C
                         i  0..H    i 0..H
                                C1       C2    C3
                     n
                                                A
            Source                                  Sink
                                                            4
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 k+1:
                                                  5
    Major Idea (2)
     T= L/R0
1             2           3        4                     K=4
                                                                   At sender


                  1                    2             3         4
                                                                   At receiver
    D1                    D2                                       when R0>A
                                            D3           D4



          1           2        3                 4                 At receiver
                                                                   when R0<=A
D1            D2          D3           D4

        If R0>A, OWDs show an increasing trend
        If R0<=A, OWDs do not show an increasing trend
                                                                            6
Proof (1)


   1. At the first link
   Case I: R0<=A1
       The arrival rate is R0+u1C1 <= A1+u1C1=C1
       Thus, packet k is served before packet k+1
        arrives:


        where Ri-1 (Ri) is defined as the entry-rate (exit-
        rate) of the packets at link i
                                                              7
Proof (2)
   Case II: R0>A1
       During interval T = L/R0,
           The link receives L + u1C1T = (R0+u1C1)T bytes
           The link serves C1T bytes
       So     = (R0+u1C1)T - C1T=(R0-A1)T, and
       Packets k+1 departs the first link Λ time units
        after packet k:
       Thus

       Given R0>A1 and C1>=A1 , we have

                                                             8
Proof (3)

   2. Induction to subsequent links
       For link i:



                                              (1)

       Also the queueing delay difference:

                                              (2)

                                                    9
Proof (4)
   IF R0>A,
       There exits a link t such that Rt-1 > At. Otherwise
        R0 = R1 = …= RH, hence R0<=Ai, contradicting
        R0>A.
       Thus,          , and so the OWD difference
        between successive packets will be positive
        (i.e.,       )
   IF R0<=A
       Then Ri < Ai for every link i
       Thus,         , and so the OWD difference

                                                          10
Iterative Algorithm to Estimate A
   At source: Send a periodic stream n with rate R(n)
   At receiver:
       Measure Di for i=1…K
       Check for increasing trend in OWDs and notify source
   At source:
       If trend is increasing (i.e. R(n)>A), repeat with R(n+1)< R(n)
       If non-increasing (i.e. R(n)<A ), repeat with R(n+1)>R(n)
   Terminate if R(n+1) – R(n) <ω
       ω : estimation resolution



                                                                   11
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 > A
       Small K does not give enough samples to infer
        trend robustly
                                                       12
Use of Several Streams

   N streams allows us to examine N
    consecutive times whether R > A or not
   Multiple streams, separated by silence
    period allows queues in network to drain
    measurement traffic
   Duration of a fleet: U  N  ( K  T  )



                                                13
Detecting an increasing trend
   Pairwise Comparison Test (PCT):

                  
                      K
                              I ( Dj  Dj  1)
         Rpct   
                      j 2
                                                    0  Rpct  1
                              K 1
       E[PCT]=0.5 for independent OWDs,         PCT  1.0when increasing
        trend
   Pairwise Difference Test (PDT):
                  
                      K
                               ( Dj  Dj  1)
         Rpdt   
                       j 2
                                                    1  Rpdt  1
                  
                      K
                      j 2
                               | Dj  Dj  1 |
       E[PDT]=0 for independent OWDs,           PDT  1.0 when increasing
        trend

                                                                         14
    Illustration of PCT and PDT
    Metrics




    Infer increasing trend when PCT or PDT trend  1.0

                                                          15
     Rate Adjustment Algorithm
                                        Increasing trend :
                             Rmax > A        Rmax = R(n)
                                             R(n+1) = (Gmax + Rmax)/2
                             Gmax
      Grey region                       Non-increasing trend:
                             Gmin            Rmin = R(n)
                                             R(n+1) = (Gmin +Rmin)/2
                             Rmin < A   Grey region & R(n) > Gmax:
                                             Gmax = R(n)
   Increasing trend, if more than           R(n+1) = (Gmax + Rmax )/2
    f*N streams in a fleet are of
    increasing trend                    Grey region & R(n) < Gmin:
   Non-increasing trend, if more            Gmin = R(n)
    than f*N streams in a fleet are          R(n+1) = (Gmin + Rmin)/2
    of non-increasing trend
   In grey region, otherwise            Terminate if:
                                            Rmax – Rmin < 
                                          or                                 16

                                            Rmax – Gmax <  and Gmin – Rmin < 
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):


                                                                     17
   PathLoad estimate falls within the MRTG range in 10 out of 12
    runs




                                                                18
4. Variability of Available
Bandwidth (1)
                                      R m ax  R m in
                                  m ax
   Relative variation index:      ( R  R m in ) / 2
   1. Heavier tight link utilization leads to higher
    available bandwidth variability




                                                        19
4. Variability of Available
Bandwidth (2)
   2. Longer probing streams observe lower available
    bandwidth variability, but also can be more intrusive




                                                        20
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.

                                                    21
    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.




                                                                   22
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.
                                                    23

								
To top