[e2e] Re: crippled Internet
vjs at calcite.rhyolite.com
Thu Apr 26 12:25:56 PDT 2001
> From: Randy Bush <randy at psg.com>
> To: Vernon Schryver <vjs at calcite.rhyolite.com>
> Cc: end2end-interest at postel.org
> > Which ISPs? Those running OC 192 simply cannot delay packets enough to
> > matter
> possibly the delays are not always in the circuits. if so, an OC1000000
> circuit will not help you. read the exchanges about what folk are actually
> seeing and MEASURING as opposed to the gossip some seem to be hearing from
> the next cubicle.
I don't always succeed, but I find it helps to use a calculator to
take the most foolish edges off my prejudices and need to flame.
Of course the measurements are of real things. I see that I removed
from my messages my obvious guess that such delays must be caused by
losing sync, router crashes, routing loops (just barely), or similar
high speed hiccups, or they are happening in low speed queues.
As others have said but evidently Mr Bush didn't think through, at high
speeds simple queuing delays cannot explain high delays. An OS 192 link
runs at about 10 Gbit/sec. A 500 millisecond queuing delay implies that
the queue contains about 5 Gbits or 625 MBytes of data. I won't return
Mr. Bush's favor and pretend to assume that he thinks routers might have
half a GByte of buffering per circuit. I won't even assume that he doesn't
know that it is often better to have less than more buffering even for
already small buffers, or that a dropping packet is often better than
queuing it to send next year (where a "year" is extremely relative, if
not preciesly relativistic).
The fact that there cannot be significant (for voice) delays in Tier
1 and even Tier 2 networks that are not (as someone else said) bugs
is important for how VoIP can work and for where the problems lie when
it doesn't. One obvious implication is on the topic at start of this
thread. If the worst case delays that a part of the path might
contribute are insignificantly more than speed of light delays, then
that part does not need to worry about quality of service queuing no
matter what the congestion. Of course, it might need to worry about
quality of service packet drop policies.
Vernon Schryver vjs at rhyolite.com
More information about the end2end-interest