Does anyone have real-life experience running TCP over link-level ARQ
with a long delay (i.e. GEO satellite) link? I'm curious about the
impacts on RTO of ARQ-induced jitter that comes in half-second
multiples. What about effects of initial RTO settings?
Well, in the early days of amateur (ham) packet radio, we ran a lot of
TCP/IP over AX.25, a reliable link layer based on LAPB, on 1200 bps
contention channels. Believe me, that was a lot slower than most GEO
satellite links. We had to set our initial RTTs fairly
high. Persistent (cross-connection) RTT caching helped a lot.
LAPB is "reliable", i.e., it'll keep trying to deliver a frame for a
very long time even as TCP duplicates pile up on the send queue behind
it. And since the physical channel has contention, we used an
exponential backoff for link retransmissions. Put TCP's own
exponential retransmission scheme on top of that and things could get
pretty bad pretty quickly.
That experience was one of the things that made me design a more
lightweight RLP for CDMA. It also made me realize that there's simply
no subsitute for well designed physical and link layers, with
efficient coding, modulation, power control, collision
avoidance/detection and so forth. You can't overcome a lousy
physical/link layer by tweaking TCP any more than you can build a
castle on quicksand. ("That one burned down, fell over and THEN sank
into the swamp!")
Phil
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:20 EST