[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [RRG] Route Flap Damping parameters
Hi Sriram,
Thank-you for pointing out these references. I have read the first two,
but the third one (which I assume is your paper?) is new to me, and it
looks like it touches on what I'm getting at - especially the discussion
surrounding equations (1) to (3) and figure 7. I need to spend some time to read
it through.
I
was trying to see if by examining the equation that defines the penalty
algorithm, one could come up with a parameter set(s) that would make RFD behave
'better'. By that I mean, suppress early into a sequence of flaps, but
unsuppress fairly soon after the sequence had passed thereby preventing the long
suppressions that are detrimental to convergence. It is very early days on
this work, but it looks like some approximate equations that connect the
configurable parameters to flapping characteristics might be derivable, which
would then point the way to how to configure the parameters. For example, figure
7 in your paper is quite interesting because it looks like work I was doing on
an approximate equation linking the time-to-half amplitude to suppression
threshold, flapping penalty and flapping frequency to get suppression to occur
in 3 flaps. This equation 'might be':
Suppression
threshold
= n * Flap
penalty, where
n >= 1 and,
n = 1 + exp(ln(1/2)*tf / thalf) + sqr(exp(ln(1/2)*tf /
thalf))
where, thalf is the configurable
time-to-half amplitude
tf is the flapping frequency.
{I hope I didn't leave out any brackets when I
transcribed this}
I have said 'might
be' instead of 'is' because extensive verification with BGP simulations and
testing against actual flapping data is still not done. As well, it also seems
possible to develop other simple, approximate equations linking other RFD
parameters to flapping charateristics. I can't say this is the last word on the
subject yet, but it looks possible to give RFD 'better' damping characteristics
through different parameters, but, to repeat myself, simulation and use in the
saddle is still very much required.
I look forward to
reading your paper.
Cheers,
Jim.
Jim:
I am not sure these papers provide all the insights you
are looking for,
but they generally address some of the questions you've
posed.
http://www.ensc.sfu.ca/~ljilja/papers/spects2005_steve.pdf
http://www.eecs.umich.edu/~zmao/Papers/sig02.pdf
http://www.antd.nist.gov/~ksriram/BGP_Security_Sriram_IEEE_JSAC.pdf
(The third one includes an analytical sensitivity study w.r.t RFD
parameters;
comparing parameter values used by two different vendors;
however, the study
primarily reports simulation results of BGP vulnerability
to attacks on BGP sessions,
including aggravating influence due to
RFD.)
Sriram
At 08:15 AM 6/26/2007, LOWE Jim wrote:
Hello,
I have been following the
discussion about BGP path hunting ('[RRG] BGP path hunting, MRAI timer and
Path Length Damping') on this list with some interest. With Route Flap Damping
(RFD) being a possible factor in the observed behavior, I had some questions
about determining values for its configurable parameters (Suppression
threshold, Reuse threshold, Time-to-half amplitude, and Maximum suppression
time).
I understand that prior to the
recommendation made in RIPE-378 to not use RFD, RIPE-229 made recommendations
about configurable parameter settings. In reading RIPE-229, it seems the
recommendations were not strongly based on observations of actual dynamics
(although I could be dead wrong on this since I'm not familiar with all the
work that lead to the creation of RIPE-229). As well, from reading a selection
of papers on the subject of RFD preventing timely convergence, there doesn't
seem to be an analysis of what might happen if RFD parameter sets other than
the common defaults were applied, nor does there appear to be a history of
literature on this subject. My reading in this area has been by no means
exhaustive, so my questions are: has there been any work done on determining
what the impact of other sets of RFD parameters might be; how have (or might)
such sets be chosen; have the dynamical properties of the RFD algorithm been
worked out - with a view towards assessing whether they would even be
potentially capable of improving observed dynamics? I have been thinking about
these areas, but wasn't sure if it had all been done before and these were
dumb questions.
Cheers,
Jim.
K. Sriram, Ph.D.
Advanced Network Technologies
Division
National Institute of Standards and Technology
100 Bureau Drive,
Stop 8920
Gaithersburg, MD 20899-8920
Phone: (301) 975-3973
E-Mail:
ksriram@nist.gov
Web: http://www.antd.nist.gov/~ksriram/