15 Jan
2016
15 Jan
'16
2:58 p.m.
On 2016-01-15 14:50, Paul_Koning at Dell.com wrote:
TCP deals with it way better than DECnet, though. And various mechanisms
for detecting packet loss exists, which do not rely on timeouts. I've
implemented some in RSX, and it now usually holds up really well.
And with DECnet, a packet loss of, say, 10% will cause about a 99%
degradation (or more) in performance.
On Jan 14, 2016, at 10:33 PM, Johnny Billquist
<bqt at softjar.se> wrote:
...
DECnet does *not* deal gracefully with massive packet loss. :-/
Performance really goes down the drain. Doing morse with a flashlight bouncing off the
moon will be faster.
No ARQ protocol deals gracefully with significant packet loss. There was a paper about
TCP decades ago, way back when the Internet was still called "ARPAnet". It
showed that a 1% packet loss rate would cause a 50% performance drop. With modern link
speeds, the impact is far greater yet. This is unavoidable because the response to packet
loss is timeout and retransmission, and a timeout by definition has to take longer than
the likely round trip latency. There do exist protocols designed for substantial
packet loss, but they are far from mainstream. (Basically, they use FEC -- redundancy in
transmission -- rather than timeout and retry.) You might find them in deep space
satellite downlinks, for example, or shortwave or weak signal radio data networks like
JT65. But DDCMP, X.25, TCP, TP4, and DECnet are all examples of protocols designed with
an assumption that packet loss is fairly rare. It isn't usually stated explicitly,
but as the ARPAnet paper showed, you really need to aim for well under 1 percent loss,
especially on fast links.
Many of these also assume that packet loss is random (caused by link error). The
exception is congestion control, which recent versions of DECnet have, and TCP as well
(originally based in part on the DECnet work, and then extended further).
We can go on about techniques to improve life, as well as how TCP
actually deal with this, which is way more than just timeouts. But my
point was that for DECnet, the problem was so severe that I really had
to do something at the bridge, or else the whole of HECnet turns useless
using the bridge.
Johnny
--
Johnny Billquist || "I'm on a bus
|| on a psychedelic trip
email: bqt at softjar.se || Reading murder books
pdp is alive! || tryin' to stay hip" - B. Idol