[NANOG] Comcast latency

Has anyone else noticed a significant increase in latency within Comcast's network?

On one quick test, it looks normal to me from my house.

    --Steve Bellovin, http://www.cs.columbia.edu/~smb

Looks can sometimes be deceiving. :wink: I've seen Comcast drop packets
left and right, but show 8Mbs/2Mbs on speed test sites. Other times
I've seen the opposite, zero PL but extremely high latency (seconds,
double digits!)... all the while non-dynamic web pages (probably
cached upstream) loaded just fine.

Look, I've always been a big fan of Comcast.... but at this point I
suspect they have just about over-engineered their network.

-Jim P.

As luck would have it, I have been running Smokeping from my home
connection (Comcast Business) pointed towards Comcast Residential (and
other broadband providers) routers for a few weeks. I'm not sure when
you saw the latency increase, but if it has been in the past few
weeks, I might actually have useful data for you. If you (or others)
are interested in access, please contact me off-list.


I had mrt running across a Comcast connection that indicated a
14000 ms RTT for one packet out of several hundred. I wasn't sure
that was a trustworthy measurement, but it sure suprised me.

Another time, I had a VOIP conversation fall apart once with somone on a
Comcast link. When the jitter finally fell into a reasonable
range, the call sounded *great* again -- but there was a
4 second latency that had been introduced. It was wild. I kept
wondering -- "where was this buffered? Are we on a satellite
backup route?" We kept the conversation going for a while just for the
sheer novelty.

Overall, my data points on Comcast point to a pretty solid
service -- but the outliers are pretty impressive in their
own right.

The delay was probably introduced by an adaptive jitter buffer on one of the
VoIP end points. It is supposed to reduce the buffer as network conditions
improve. When it "fell apart" the buffer was increased until it was long
enough to reorder all of the packets.

It would be interesting to see if a call between those same end points would
reduce the size of their buffers after the network stabalized again.