🚀 go-pugleaf

> > I don't think so.  Upstream oversubscription is really an issue only on
> > cable, since (1) DSL has dedicated upstream to the DSLAM, and (2)
> > backhaul links are symmetric (as compared to asymmetric subscriber
> > connections).

I disagree on one finer point.

This is where I think we agree:

* TCP itself will limit downstream throughput when the upstream channel
becomes saturated

* Cable upstream saturation (between customer and CMTS) is highly likely
because of the shared nature of the link

User A can be degraded because B, C, D, and E are uploading at max rate.
-or-
User A can be degradged because just the cumulative affect of all users
ACK traffic causes saturation.  (this would seem to occur only where
*many* users share the same local segment).

* ADSL upstream saturation (between the customer and the DSLAM) is
highly
unlikely because it is a private link. (Though the customer could
saturate it himself with some kind of outbound datastream...e.g. UDP
broadcast video).

It's NOT going to saturate just due to his outbound ACK traffic.
Assuming download rate of 150,000Bytes/sec, with 1500 bytes
packets...that's about 100packets/sec.  That should result in upstream
50packets/sec of 64byte ACKs...that's 25Kbps, far less than the 256Kbps
typical ADSL upstream bandwidth.

Here is where I disagree:

Downstream degradation occurs WHENEVER the uplink is saturated.  Granted
the more limited (and shared) upstream bandwidth of cable makes this an
easy choke point.  However, ANY uplink that saturates can have the same
affect.

Even though bandwidth between the DSLAM and aggregation router is
assymetric (e.g. 45Mbps full-duplex), you could still saturate upstream
before downstream, depending on what users are doing.

For example, if 5% (175) of the ~4000 ADSL users aggregated into a DS3
ran active web servers (uploading at constant 256Kbps), you'd saturate
the upstream.  I'm not saying that this is extremely likely, but I
believe the best way to quantify the probablility is to quantify:
upstream oversubscription rate

To do an apples to apples comparison, you need to compare both
technologies at the same layers: Customer to Local Aggregation (I think
we agree cable and ADSL infrastructure upstream from the aggregation
layer use similar architectures).

ADSL: Subscriber to Aggregation router
Cable: Subscriber to CMTS router


I suspect the upstream oversubscription rate (at the local aggregation
layer) of cable is 10x-20x more than ASDL.  I think this is a more
generic way to quantify the probability of upstream saturation and
provide metric to compare different technologies on equal ground.

In fact we already know that the way to compare different ISPs to each
other is to understand their downstream oversubscription policies.  What
I find intriguing, is that you need to look at upstream as well...this
is a real eye-opener for me.

Taking this concept a bit further, my real suspicion is the
oversubscription at the *backbone* layer is what is causing the most
differences in experienced throughput variances between different
providers (whether ADSL or cable, or whatever).  Wish I had the data to
back that up though.  

Upstream saturation might actually be an even bigger issue at backbone
interconnect layer.  Think about it, how many ISPs merge ADSL traffic
(downstream traffic) with commercial T1 traffic (primarily upstream) at
the backbone layer.

I suspect that few considered the impacts of upstream saturation on ADSL
throughput...or ADSL downstream traffic saturating the reverse channel
of their T1 customers...hmmmmm.

lb.
-- 
-------------------------
Lawrence Baldwin
Protocol Scientist
678-624-0924
http://www.mynetwatchman.com - I'm under attack, now what?