home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.dcom.cell-relay
- Path: sparky!uunet!cs.utexas.edu!qt.cs.utexas.edu!yale.edu!spool.mu.edu!sgiblab!sgigate!sgi!rigden.wpd.sgi.com!rpw3
- From: rpw3@rigden.wpd.sgi.com (Rob Warnock)
- Subject: Re: clarifying the role of SSCOP?
- Message-ID: <s92deoo@sgi.sgi.com>
- Sender: rpw3@rigden.wpd.sgi.com
- Organization: Silicon Graphics, Inc. Mountain View, CA
- Date: Thu, 12 Nov 1992 12:23:56 GMT
- Lines: 88
-
- goldstein@carafe.enet.dec.com (Fred R. Goldstein) writes:
- +---------------
- | Will ATM be that bad? I hope not. I know how to keep it from being
- | very lossy. But I'm not sure the CCITT does. And if it's cheaper in
- | the long run to run the circuits at 95% utilization with say 1% loss
- | (and selective retransmission) than to run them at 90% with 0.001% loss,
- | then why not? And if you run it even lossier (hey, it may not be
- | _planned_), then local retransmission begins to look good.
- +---------------
-
- Fred, I think you're exaggerating the possibilities for raw cell loss due
- to transmission effects [not congestion, but see below]. Remember, this same
- ATM network has got to carry voice traffic with no AAL at all (well, AAL1
- is practically no AAL). Every cell dropped will be a 6ms gap in the conver-
- sation (12ms, if 32 Kb/s ADPCM is used). And depending on such a good error
- rate should not be surprising; today we have *no* retransmission on *any*
- long-haul voice channels. (We do have FEC on satellite circuits, but that's
- not retransmission.)
-
- I am quite confident that -- in the absence of congestion -- the global ATM
- network will have a cell error rate low enough to allow traditional end-to-
- end retransmission of AAL5 packets to be reasonable. And I am also confident
- that the call setup procedures will not allow significant congestion-induced
- cell loss on constant-bit-rate circuits [well, in the absence of TASI...].
- It's just too easy to avoid it by simply not completing the call if the
- requested bandwidth isn't there.
-
- HOWEVER... I do not believe the same to be true if LAN-like bursty traffic
- is allowed to create congestion in small-buffered ATM switches ["small" is
- by comparison with the buffer sizes one would configure for a traditional
- IP router with a similar load on it.]
-
- I fear that because of the small buffers, the congestion (and thus, cell loss)
- will be *far* worse than anything we have ever seen in the Intenet to date,
- and that we *have* to have some form of hop-by-hop flow control on bursty
- traffic so that the net simply does not *allow* congestion-induced cell loss.
- When even *one* full-sized AAL5 PDU (64 KB) sent at the full link rate may
- overflow some switches' output queues (after all, that's 1366 cells!),
- end-to-end congestion avoidance is impossible. [It's always too late.]
-
- We either have to either (1) do "leaky-bucket" rate control to never exceed
- the call's guaranteed bandwidth, which reopens *all* the "dynamic bandwidth"
- issues we have already with IP over X.25 or IP over multiple BRIs, and which
- *never* lets a single connection get the full bandwidth of the ATM access
- link (*boo* *hiss*), or (2) we have to have hop-by-hop flow control (call
- it "back-pressure" if you prefer) on bursty traffic, at a pretty fine-grained
- level (a few cells or a few 10's of cells).
-
- +---------------
- | I'm not saying everyone will need SSCOP. Just that it may be a nice tool
- | when things go wrong, and not a very expensive one to keep around anyway.
- +---------------
-
- But it's the *wrong* tool to try and solve burstiness-induced congestion.
- Hop-by-hop flow control (or reasonable-sized buffers for bursty traffic,
- which I don't hold much hope for) is a much better tool for this particular
- problem.
-
- [See a previous article for a proposed hop-by-hop flow control which would
- be a lot simpler than per-VC credits but just as effective (I think).]
-
-
- -Rob
-
- p.s. Back at Dig. Comm. Assoc. circa 1972, we used hop-by-hop per-VC credits
- (we called them "pre-ACKs") in our statistical multiplexers. Worked like a
- charm! We *never* dropped data due to internal network "congestion" -- we
- simply didn't permit it! But it took a lot of computes *and* large buffers,
- which is why I don't recommend it for ATM.
-
- If the sending host or terminal at the edge of the network wouldn't or
- couldn't respond to the provided flow-control indications (XOFF/XON or CTS),
- we of course were forced to drop data, but we also then "notified" the host
- by dropping the virtual circuit! (*Boom*) "Carrier fail disconnect..."
- [A reasonable policy for the stat-mux market at that time.]
-
- [We also did hop-by-hop retransmission, but that's another issue. Link speeds
- were slow then, and multi-hop latency was a big problem when echoing keyboards
- through the network. Hop-by-hop retransmission, with NACKs, not timeouts,
- helped keep the "stutter" due to error recovery short.]
-
-
- -----
- Rob Warnock, MS-9U/510 rpw3@sgi.com
- Silicon Graphics, Inc. (415)390-1673
- 2011 N. Shoreline Blvd.
- Mountain View, CA 94043
-
-