[nycbug-talk] Why is DSL-Cable asynchronous?

[Miles Nordin]

>>>>> "il" == Isaac Levy <ike at lesmuug.org> writes:

    il> Why is consumer dsl or cable asynchronous?

for cable, WCDMA/cdma2000/gprs, and maybe FiOS, it is actually more
expensive to move bits upstream at L1.  The problem is that you must
share a band among the many endpoints who'd like to send upstream, and
there's overhead to coordinating this.  There are two general
approaches.

One is straight AlohaNet, or optimisations like ``slotted aloha''
which will get you only 30 - 40% link usefulness in the pure
mathematical model.  (compared to downstream, which has a single
transmitter and thus always 100% link use.)  Half-duplex ethernet can
do better than this because of collision detection---it's able to
listen while transmitting---which radio cannot do, nor can networks
with a large diameter measured in bits: 

  [physical size] / (c * bit-time)

The other approach, which all three in fact use, is to do aloha on a
thin reservation channel to schedule who may talk on a separate fat
upstream channel.  This trick in combination with ``ranging'' lets the
fat channel fill with high efficiency even though the reservation
channel has to stay mostly-quiet.

The fat upstream can thus be about as efficient as downstream.
However I expect there might be other problems for cellular which
cable/FiOS don't have related to band reuse, because cellular does not
have the filters cable/FiOS have to block the signal, and has
endpoints in 2 or 3 dimensions instead of 1 dimension.  I haven't
fully thought it through but the downstream antennas are of
higher-quality, of known 3D location, and are aimed into the ground,
and can be coordinated with each other using GPS while upstream
antennas are omni and coordinated in time with ranging only and
exist in two (or three) unknown dimensions.

anyway...The problem for this thin/fat trick is, the relative sizes of
the reservation channel and the fat channel depend on the size of the
upstream packets you're sending.  There's only a gain if your data
packets are much larger than a reservation packet.  so, honestly,
bittorrenting's upstream is less expensive per kilobit to the L1 than
a leech's stream of ACK's.

The ``unsolicited grant'' I spoke of earlier for cable VoIP is a
stateful, recurring reservation that doesn't need to be discussed on
the reservation channel.  The downside is, if you get an unsolicited
grant and don't use it, then space is wasted on the fat channel.  The
tremendous upsides are: (1) efficient reservation for tiny packets,
(2) no latency to obtain the reservation.  The same trick works for
WiMAX CIR.  I'm not sure if the 802.11 QoS uses it or not---802.11 is
a bit different since it has a small diameter.  Unsolicited grant is
analagous to how circuit-switched celfone voice calls are done
upstream, except on cable it's generalized into a type of QoS for IP
so it can be used with a regular UDP application stack instead of
quaint cumbersome TDM stacks.

for DSL I expect it's all fucking layer 10 market-forces baloney as I
discussed earlier.  I doubt there's anything expensive about upstream
at L1.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 304 bytes
Desc: not available
URL: <http://lists.nycbug.org/pipermail/talk/attachments/20081101/ac759d7a/attachment.bin>