[69142] in North American Network Operators' Group
Re: Overflow circuit
daemon@ATHENA.MIT.EDU (Steven M. Bellovin)
Sat Mar 27 19:39:45 2004
From: "Steven M. Bellovin" <smb@research.att.com>
To: "Alexei Roudnev" <alex@relcom.net>
Cc: "Michel Py" <michel@arneill-py.sacramento.ca.us>, nanog@merit.edu
In-Reply-To: Your message of "Sat, 27 Mar 2004 11:29:46 PST."
<05bb01c41431$dd522e00$6401a8c0@alexh>
Date: Sat, 27 Mar 2004 19:39:06 -0500
Errors-To: owner-nanog-outgoing@merit.edu
In message <05bb01c41431$dd522e00$6401a8c0@alexh>, "Alexei Roudnev" writes:
>
>Thanks for the answers about Voip usage over satellite (I did not know, that
>it does not cause unacceptable delays and echo).
>Responses (which I received) shows, that many people deployed such system
>successfully.
Define "unacceptable".
Old-style telcos have delay budgets for their designs; if the
round-trip time is too long, people find the call unpleasant. While
VoIP does have its own delay issues (see below), the big problem here
is the satellite link. VoIP doesn't give you an exemption from the
speed of light laws; if you find satellite phone calls unpleasant -- I
do -- you're not going to like satellite VoIP calls, for reasons that
have little to do with the IP. This is one reason why companies have
spent fortunes putting in transoceanic fibers instead of launching more
satellites -- the customers prefer the quality. Satellite calls are
cheaper, but they're noticeably -- and for many people, unacceptably --
worse in quality.
I should note that many VoIP systems make this noticeably worse, though
(of course) less so as a percentage of the total delay than for
domestic US VoIP calls. The problem is the tradeoff between delay and
efficiency. Suppose you're sending 56 Kbps, uncompressed -- the
equivalent of so-called "toll quality" voice. That's 7 kilobytes/sec.
If you want nice, big UDP packets with 1K payloads, you've just
incurred about 143 ms of buffering delay, independent of transmission
time. (And on a DS1 line, transmission time for that packet is
non-trivial.) The total delay budget is, as I recall, about 150 ms.
You can go to nice, short packets -- say, 100 bytes -- but then your
IP, UDP, and RTP headers add a substantial amount of bandwidth
overhead. Apart from line efficiency, on relatively slow lines that's
a lot of serialization time going out over the wire. Compression makes
the packets nice and short; depending on what you do, it may or may not
help with the headers, but you have to spend a chunk of CPU time after
you've collected a large enough voice sample to be worth compressing,
and that means more delay time.
Bottom line: VoIP is inherently costly, either in delay time,
bandwidth, or both. It doesn't mean it's unacceptably bad, but that
150 ms delay budget came from many years of psychoacoustic studies.
Cisco has a good web page on this, with lots of numbers. See
http://www.cisco.com/warp/public/788/voip/delay-details.html
--Steve Bellovin, http://www.research.att.com/~smb