From: smb@research.att.com
Date: Thu, 12 Aug 93 13:58:13 PDT
To: karn@qualcomm.com (Phil Karn)
Subject: Re: Secure voice software issues
Message-ID: <9308122055.AA27140@toad.com>
MIME-Version: 1.0
Content-Type: text/plain


	 I see 160ms round trip times on my SLIP link from home to work, and I
	 can't account for all of this time by just adding up transmission
	 times and store-and-forward delays for the data rates and packet sizes
	 I'm using. And I don't think it can be explained by the trellis
	 decoding in V.32 bis, as that should account for only a few bits of
	 delay.

	 I've since heard of very similar figures for other modems, so It's not
	 just my modem. I'm beginning to suspect the V.42bis packetizing
	 algorithms. Although they're not described in the spec, I suspect that
	 real V.42bis implementations use timers to determine when to send the
	 the currently queued data as a frame. Or maybe there's a Nagle-like
	 algorithm like the one in TCP: immediately send the first byte of data
	 on an idle link, but keep additional traffic pending until the first
	 byte is acknowledged in order to aggregate stream traffic into larger
	 frames.

	 This is all speculation so far, but it does explain the long RTTs I
	 see with packet traffic even though raw character-at-a-time traffic
	 seems to be fast.  Stream traffic would see the worst delays of all,
	 which is ordinarily okay for a file transfer, but death to a real time
	 stream like voice. That's why we may be forced to turn off V.42
	 entirely and speak synchronously to the modem.

	 Time to haul out a protocol analyzer and do some timing measurements.

Real timing measurements would help, but it's not just the V.42bis
algorithms.  A year or two ago, I did some measurements on a number
of different modems.  I saw large -- and sometimes unacceptable -- delays
even without V.42 or MNP in use.  My methodology was to enable loopback
at various points -- hardware loopback plugs, local loopback on my
modem, remote loopback or a plug at the far end, etc.  I sent single
characters, and timed how long they took to show up.  The modems
seem to either buffer up several characters, or have to wait a while
before sending the first one, but the delays appeared to be for the
first character in a ``bunch'' (``packet'' is too strong a word).