Can coexisting overlays inadvertently step on each other Ram Keralapura, ChenNee Chuah, Nina Taft, G - PowerPoint PPT Presentation


PPT – Can coexisting overlays inadvertently step on each other Ram Keralapura, ChenNee Chuah, Nina Taft, G PowerPoint presentation | free to download - id: a991f-NzkwY


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation

Can coexisting overlays inadvertently step on each other Ram Keralapura, ChenNee Chuah, Nina Taft, G


WASHINGTON UNIVERSITY IN ST LOUIS. Mike Wilson 29 June 2006. Can overlays step on each other? ... WASHINGTON UNIVERSITY IN ST LOUIS. Mike Wilson 29 June ... – PowerPoint PPT presentation

Number of Views:82
Avg rating:3.0/5.0
Slides: 31
Provided by: mlw2
Learn more at:


Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Can coexisting overlays inadvertently step on each other Ram Keralapura, ChenNee Chuah, Nina Taft, G

Can coexisting overlays inadvertently step on
each other? (Ram Keralapura, Chen-Nee Chuah,
Nina Taft, Gianluca Iannaccone)
  • The problem
  • Coexisting overlay networks usually have separate
    routing mechanisms. Experience teaches us that
    this can lead to routing update synchronization.
  • The authors of this paper analyze
  • What conditions are necessary for synchronization
  • Probability of synchronization
  • Maximal length of synchronization
  • How best to tune overlays to avoid synchronization

Routing Oscillation
Overlay 1
Overlay 2
Everything is fine, links are utilized well
We have a congested link
Both overlays detect the congestion and make the
same routing decision change to an alternate
We have a new congested link. Both overlays
detect the congestion and switch back to the
former link.
Why do oscillations happen?
  • In essence, synchronizations can happen when
    multiple control paths react, after a delay, to
    the same trigger.
  • Multiple overlay networks
  • …each using periodic probing
  • …each detecting congestion via probes
  • …sharing links in the current routing path
  • Oscillations occur when the overlays share links
    in both primary and alternate paths.

Modeling overlays
  • The model only addresses overlays with periodic
    probing fitting the following model
  • Probes are sent every P seconds
  • Probes have a timeout value of T seconds
  • After a timeout, probes are sent at a higher
    rate, every Q seconds
  • After N unsuccessful high frequency probes, a new
    route is sought

Analytic results
  • Authors provide an analysis where they derive the
    conditions for synchronization as

Region of Conflict
  • From the conditions for synchronization and the
    overlay probing parameters ltPi, Qi, Ti, Ni, Rigt,
    the authors can compute the potential region of
  • From ltPi, Rigt we can define a rectangle enclosing
    all possible values for t1, t2 given a link
    failure at time tl0.
  • Finally, using the rest of the parameters and the
    conditions for synchronization, we can determine
    which portions of the bounding rectangle can
    result in synchronization.

Region of Conflict
Computing the Probability
  • The actual computation is complicated by details
    resulting in nine special-case formulae that
    depend on which sides of the bounding rectangle
    border the region of conflict.
  • In all cases, the general method is the same
    find the area of the region of conflict, divide
    by the area of the bounding rectangle.

Duration of Oscillations
  • Oscillations are subject to the same condition
    for synchronization, that is, b lt t1 t2 lt a.
    Thus, our stop condition is t1 t2 gt a b.
  • The spacing between t1 and t2 will grow with
    every oscillation.
  • The upper bound on oscillations can be derived as

  • The authors attempt to validate their analytic
    conclusions with simulations.
  • The underlying network is modeled as IP using IGP
    for the routing. Routing overlays are used to
    attempt a faster response than IGP to link
  • Despite author claims of strong support, there is
    at least one serious question on the agreement of
    the simulations with the analysis.
  • On the whole, the results mostly agree with the

One Clash Simulation/Analysis
  • According to the paper, the first set of
    oscillations terminated with self-disentanglement.
  • Using the formula derived in the analysis, the
    maximum number of oscillations we should have
    seen for these overlays is 1!

Note the length of time in oscillation is close,
but may not match the X-axis is misleadingly
labeled with ms x 104. E.g., Overlay 1 should
detect failure in 5350 ms the chart shows
approximately 1250 ms.
Validation Results
These results are much better. I still question
the value of the simulation used to prove that a
parameter, Q, is not a factor (lower left).
Effect of Probing Parameters
  • The authors attempt to determine probing
    parameters that reduce the probability of
    conflict with any coexisting overlays.
  • Since the complete 2-overlay analysis is
    unfeasibly complex (10 parameters, yielding a
    10-dimensional probability plot), the authors
    focus on the following questions
  • Is P(S) significant with typical Internet values?
  • Will natural variation in RTT reduce P(S) to
  • If not, what parameter settings drive low P(S)

Study of Parameter Impact Caveat
  • The results of studies on the probing parameter
    impact are purely based on the theory, not
  • Since the authors claim to have validated their
    model against simulations, this should be
    acceptable however, the study range of
    parameters is much greater than the validation
  • Before being put into practice, these results
    should be validated against simulations (or
    experimental testbeds).

  • Consider the case of identical parameters. P(S)
    T(2P-T)/P2, which depends only on P and T.
  • It is also common for overlays to use PT, in
    which case P(S) 1. That is, synchronization is
  • The authors characterize this by aggressiveness
    of an overlay, defined as a Ti/Pi.
  • For identical overlays, P(S) 2a a2, thus,
    P(S) increases with aggression. Obviously,
    response time also increases with aggression, so
    there is incentive for overlays to be aggressive.

Effect of RTT
  • For non-identical overlays, RTT can be
  • The authors assume that most overlay creators
    will use an approach similar to TCP that is,
  • Since TQP, it makes sense to select P,Q as
    multiples of T (and thus RTT).
  • This leaves the same formula for P(S), but will
    impact the formula for since different paths
    will have different values of P(S).
  • The authors also investigate the result of fixed

Effect of RTT
  • The results are unsurprising. As the RTTs
    approach each other, P(S) increases.
  • Also as predicted, as aggression increases, P(S)
  • Result Natural RTT variation can help, but may
    not be enough.

Disparate Aggressiveness
  • The authors demonstrate that a non-aggressive
    overlay has potential benefit even in the
    presence of an aggressive overlay.

Fixed RTT, Varying Aggressiveness
  • As expected, P(S) is at the worst when both
    overlay networks are aggressive.
  • Since aggressiveness is one of the few tuneable
    factors, avoiding synchronization seems to imply
    less aggressive overlays.

Fixed parameters Effects of RTT
Unlike proportional parameter systems, absolute
RTT now matters. Best results occur with small
RTT values. (This chart varies from 20ms to
Fixed Parameters Effect of Q
As Q approaches P, P(S) goes down. However, it
is clear that P (and therefore aggressiveness) is
much more significant than Q.
Maximum number of Oscillations
Note the authors make no statement about the RTT
of the alternate path I assume that both the
primary and alternate RTTs are the same.
Author Conclusions
  • Natural RTT variation is not an adequate method
    of avoiding synchronization. Overlay designers
    should take the possibility of synchronization
    into account explicitly.
  • Sound overlay design is hard. There is no safe
    range of parameters, although the authors believe
    that proportional parameter systems can reduce
    the range of relative RTTs with high
    synchronization probability.
  • The best approach to avoid synchronization is to
    be non-aggressive. However, the entire point to
    routing overlay networks is faster response to
    events. Intuitively, the trend will be toward
    more aggressive overlays, not less aggressive

An observation for the TechX team
  • Most of the Network Diversification Architecture
    is designed to provide isolation between
  • With provisioned links and capacity guarantees,
    this entire problem goes away. However, this
    relies on end-to-end capacity negotiation, a hard
  • The problem becomes easier when we only attempt
    to maintain isolation between meta-networks, not
    within them. That is, our overlays can only step
    on themselves, not each other.

Supplemental Diagram
Analytic Derivation
  • Overlays 1 and 2 have both detected a link
    failure on probes sent at times t1 and t2,
    respectively. After following their respective
    probe procedures, the final (high-frequency)
    probes are sent at f1 and f2, respectively, where
    fi ti NiQi.
  • These probes time out at f1 T1 and f2 T2. If
    both probes are sent before either probe expires,
    then the networks will synchronize. Thus, for
    the case where overlay 1 moves first, we
    synchronize if

The case where overlay 2 moves first is analogous
Analytic Derivation
  • The derivation of the synchronization condition
    continues by combining our two inequalities and
    continuing as follows

Supplemental Diagram All Scenarios