home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.dcom.modems
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!darwin.sura.net!spool.mu.edu!umn.edu!csus.edu!netcom.com!gerg
- From: gerg@netcom.com (Greg Andrews)
- Subject: Re: PEP over SLIP?
- Message-ID: <1992Nov23.082126.29825@netcom.com>
- Organization: Netcom Online Communications Services (408-241-9760 login: guest)
- References: <1992Nov16.213034.27895@eos.arc.nasa.gov> <1992Nov20.083723.11984@netcom.com> <sjsv1fo@rhyolite.wpd.sgi.com>
- Date: Mon, 23 Nov 1992 08:21:26 GMT
- Lines: 79
-
- vjs@rhyolite.wpd.sgi.com (Vernon Schryver) writes:
- >In article <1992Nov20.083723.11984@netcom.com>, gerg@netcom.com (Greg Andrews) writes:
- >> vjs@rhyolite.wpd.sgi.com (Vernon Schryver) writes:
- >> >
- >> >I just don't see how SLIP/PEP can "degrade under [increased] offered
- >> >load." With increased load, the modems turn the line around less
- >> >frequently, thereby wasting less bandwidth. "Collapse" would require
- >> >just the opposite, decreased effecency with increased offered load
- >> >
- >>
- >> If Bob's "offered load" is multiple TCP sessions, he could again be
- >> describing a link that's returning multiple ACKs to the modem before
- >> the line turns around. If the SLIP packets were small, the results
- >> would probably be as bad as he indicates.
- >> ...
- >
- >I do not see this.
- >
- >The word "collapse" is often reserved for situations where a small
- >increase in offered load results in a substantial decrease in
- >throughput. The system, whether a computer virtual memory system
- >during "thrashing" or ethernet in "collision collapse" or rush-hour
- >traffic in gridlock becomes all but useless in "collapse."
- >
- >As far as PEP modems are concerned, a high offered load is just a lot
- >of bytes. It does not matter to them whether those bytes are for 1 TCP
- >stream or a thousand. The modems will be presented with a high load in
- >both directions, and will switch to large packets. Once they have
- >switched to large packets, the modem-to-modem link does not "collapse"
- >any more, does it?
- >
-
- Before I try to answer your last question, let me admit that I didn't
- fully understand the use of the word "collapse" in this context. I'm
- vaguely familiar with data flow patterns over SLIP, but I'm probably not
- aware of the more subtle details. My response may have been about a
- different situation than the one you and Bob have in mind, confusing the
- picture rather than making it clearer...
-
- The type of session I had in mind is where there are multiple TCP sessions
- passing through the same SLIP link. With my vague understanding of how
- SLIP works, I pictured each TCP session returning ACKs independently of
- (and asynchronously from) the other sessions. Perhaps I erred in this,
- but if not, then the additional ACKs could conspire to increase PEP's
- overhead in the reverse direction.
-
- Now for the question you asked. In theory, Yes, the link can "collapse"
- further after the modem begins using Long packets in the reverse direction.
- Remember that the more packets the modem sends in the reverse direction,
- the more time is stolen away from the forward direction. If the modem has
- enough data (ACKs) to overflow one Long packet, it will send two. If it
- has more data than will fit into two, it will send three, and so on up
- to the maximum of 12 Long packets in a single burst. Generally, I would
- not expect ACKs to fill the modem up enough to force additional Longs, but
- it's not impossible.
-
- Also, looking back on your first article (still quoted above), I think
- I see a misconception where you say "...the modems turn the line around
- less frequently, thereby wasting less bandwidth." Bandwidth is wasted
- far less by a line turnaround than it is by transmitting something back
- in the reverse direction. If you count the number of line turnarounds
- per second, you'll think returning a Long packet is good, as the modems
- turn the line around less often. However, extend this concept a little
- further and the problem will rear its head. Return 12 Long packets and
- you'll have enormously reduced the number of line turnarounds, but your
- throughput will drop through the floor. It's not the number of line
- turnarounds that's important in PEP. It's the proportion of forward
- transmit time to reverse transmit time.
-
- (Sorry if you're already aware of all this - I'm just trying to
- clear up some things that appeared to be misleading...)
-
-
- --
- /u3/gerg/signatures/sig.carp-pyramid
- Greg Andrews -- gerg@netcom.com -- ...!{claris,amdahl}!netcom!gerg
-
- ObGuindon: Carp pyramid schemes are fun, but remember to enter
- late. That way, you'll get stuck with fewer carp.
-