dupont@inria.inria.fr (Francis Dupont) (03/21/91)
An important component of IGRP metrics is the "delay". Cisco manuals don't explain how to choice the delay of an interface (the defaults are 1000us for Ethernet and 20000us for Serial). How do you compute a delay ? The transit time for one bit is to much bound to bandwidth ... Perhaps the transit time of the first bit (for instance (RTT(200B)-2*RTT(100B))/8*100) is a good approximation of delay ? Francis.Dupont@inria.fr PS : default formula for IGRP metric is : 10000*Bandwidth+Delay/10 with Bandwidth in kb/s and Delay in us
Allen Robel <robel2@mythos.ucs.indiana.edu> (03/26/91)
> metric = [K1*bandwidth + (K2*bandwidth)/(256 - load) + K3*delay] * > [(reliability + K4)/K5] > > If K5 == 0, then there is no reliability term. > The default version of IGRP has K1 == K3 == 1, K2 == K3 == K5 == 0" With K5=0 you're dividing by zero which is a no-no, right? Is K5 really equal 0 or is the equation wrong? I'd like to include the metric in the IGRP decoder I'm working on for Network General's Sniffer. I checked in Hedrick's paper @ ftp.cisco.com ~ftp/igrp.ps and the above is an accurate representation as it appears in that paper. >The Bandwidth must be inverted and scaled by 10E10. Also quoting from >the same source: > >"Bandwidth is inverse bandwidth in bits per sec scaled by a factor of >10e10. The range is from a 1200 bps line to 10 Gbps.... I tried this on my trusty HP: (1/6476)*10E10 (for T1 bandwidth = 1.544Mbs) and get: 15,441,630.6362 My guess is that this should be 10E9? Any info from anyone clarifying these points would be most helpful! thanks, Allen Robel robel2@mythos.ucs.indiana.edu University Computing Services ROBELR@IUJADE.BITNET Network Research & Planning voice: (812)855-7171 Indiana University FAX: (812)855-8299
robel2@mythos.ucs.indiana.edu (Allen Robel) (03/26/91)
>> > If K5 == 0, then there is no reliability term.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Thanks!!! I was reading that to mean that K5, being 0, would somehow
"weight out" the reliability term mathematically rather than
relying on programming ( e.g. if (K5) ...).
allen
robel2@mythos.ucs.indiana.edu (Allen Robel) (03/26/91)
>> > The default version of IGRP has K1 == K3 == 1, K2 == K3 == K5 == 0"
So what's K4, and is K3 really 1 or 0?
thanks,
allen
whp4@cisco.com (Bill Palmer) (03/26/91)
> >> > The default version of IGRP has K1 == K3 == 1, K2 == K3 == K5 == 0" >So what's K4, and is K3 really 1 or 0? It should be K1 == K3 == 1, K2 == K4 == K5 == 0 By default, your metric won't change if your topology doesn't. Only the truly brave who like solving non-linear feedback systems will want to play with K2, K4 and K5, I think. Bill Palmer cisco
forster@cisco.com (Jim Forster) (03/26/91)
Francis,
For the purposes of measuring the delay to be used in IGRP, the simplest
thing to do is run a bunch of 1500 bit ping packets, and use the average delay.
That's 1500 bits, not bytes.
Quoting from ftp.cisco.com:~ftp/igrp.doc
"In principle the composite delay, Dc, could be determined as follows:
Dc = Ds + Dcir + Dt Eq. 2
Where:
Ds = switching delay;
Dcir = circuit delay (propagation delay of 1 bit); and
Dt = transmission delay (no-load delay for a 1500 bit message)."
Walt Prue's method of plotting the delay of various size packets is
interesting, as it yields Ds+Dcir. The slope of the line is the bandwidth.
If you do this over a long path, or a line with some background traffic, you
can deduce the effective bandwidth of the path.
>> 10000*Bandwidth+Delay/10 with Bandwidth in kb/s and Delay in us
The Bandwidth must be inverted and scaled by 10E10. Also quoting from
the same source:
"Bandwidth is inverse bandwidth in bits per sec scaled by a factor of
10e10. The range is from a 1200 bps line to 10 Gbps.... Reliability is
given as a fraction of 255. That is, 255 is 100%. Load is given as a
fraction of 255....
Because of the somewhat wierd units used for bandwidth and delay, some examples
seem in order. These are the default values used for several common media.
Delay Bandwidth
Satellite 200,000 (2 sec) 20 (500 Mbit)
Ethernet 100 (1 ms) 1,000
1.544 Mbit 2000 (20 ms) 6,476
64 Kbit 2000 156,250
56 Kbit 2000 178,571
10 Kbit 2000 1,000,000
1 Kbit 2000 10,000,000
6.3 Metric computations
Here is a description of the way the composite metric is actually computed...
metric = [K1*bandwidth + (K2*bandwidth)/(256 - load) + K3*delay] *
[(reliability + K4)/K5]
If K5 == 0, then there is no reliability term.
The default version of IGRP has K1 == K3 == 1, K2 == K3 == K5 == 0"
-- Jim
robel2@mythos.ucs.indiana.edu (Allen Robel) (03/28/91)
>I have a simple igrp spreadsheet for Microsoft Excel if you'd like to >play with the Kn, occupancy, reliability and delay. Actually, it's >curious to see the results when using values other than 1 for some of >the Ks... Pierre Fortin correctly points out that the Ks can have other values besides 0 and 1. This is something I didn't check for in igrp.c for the sniffer. Where before I did: if(K5==0) no reliability term; if(K5==1) reliability term; now I do: if(K5==0) no reliability term; else reliability term; You won't need the most recent update unless you're planning on changing the K values (something that cisco, at present, doesn't recommend). But, if you *do* get the update, get *both* igrp.h and igrp.c as I've moved the Ks from the former to the latter. Here's the most recent changes for igrp.c Change log. 3/25/91 - Written allen 3/26/91 - Don't use 1st byte of local subnets to determine IP address of Nets in and outside AS. allen 3/27/91 - Changed K1-K5 from defines in igrp.h to unsigned char variables in igrp.c. -allen - Changed test for K5=0 to include other values of K5 besides 0 and 1. -allen thanks Pierre! allen
kwe@bu-it.bu.edu (Kent England) (03/28/91)
> From: whp4@cisco.com (Bill Palmer) > > It should be K1 == K3 == 1, K2 == K4 == K5 == 0 > > By default, your metric won't change if your topology doesn't. Only > the truly brave who like solving non-linear feedback systems will want > to play with K2, K4 and K5, I think. > Does cisco itself fall into the truly brave category? I remember not so long ago when cisco touted IGRP for its vector metric. At that time there was no mention of the fact that most of those values were zero and always have been. (Those truly interested in IGRP should read Chuck Hedrick's description of IGRP from a couple of years ago. I packed all my RFCs for a move next week or I would be able to offer a place to find this document. Perhaps someone else could say? Might be at rutgers in anon FTP... At any rate, a good description of the protocol and an explanation of how the metric equations allegedly work.) --Kent
aggarwal@jvnc.net (Vikas Aggarwal) (04/01/91)
> (Those truly interested in IGRP should read Chuck Hedrick's > description of IGRP from a couple of years ago. I packed all my RFCs > for a move next week or I would be able to offer a place to find this > document. Perhaps someone else could say? Might be at rutgers in > anon FTP... At any rate, a good description of the protocol and an > explanation of how the metric equations allegedly work.) > > --Kent The paper is available for anonymous ftp on nicol.jvnc.net (pub/igrp.ps and pub/igrp.doc in postscript and text versions). -vikas vikas@jvnc.net (609) 258-2403 --------------------------------------------------------------------------