Cell Relay Archive

[Jon Turner <jst@cs.wustl.edu>]

Couple things. In response to the original question, yes it
makes sense to use EPD and PPD simultaneously. In general PPD
is implied by EPD. It is possible with EPD to get partial
discards if the threshold is too high. When this happens,
the rest of the packet is dropped too.

So how to set the threshold? This is not a simple question
in the general case, but a rough rule of thumb is this.
If the per VC rate is r and the link rate is R and the
packet size is C cells, then the threshold should be
at least C*r/R below the full level to avoid partial discards.
It should also be this far above "empty" to avoid underflows.
If you want to look at this more closely, there are some
papers on my web site (www.arl.wustl.edu/~jst). Click on
the publications link and then look at paper 5 in the
journal pubs section and papers 12 and 16 in the conference
papers section.

Also,
Dominic Richens wrote:
> 
> Where things started getting a little out of hand, IMHO, was when
> Weighted-Random Early Discard ...

Here, I think we are confusing two things. Understandable because
of the similarity of the names. Early Packet Discard is an ATM
mechanism for avoiding congestion collapse due to partial discards
of packets when links get overloaded.

Random Early Discard (RED) is an IP mechanism which is
intended to improve the performance of TCP congestion control.
TCP adjusts its rate continually trying to make use of available
network bandwidth, while avoiding overloading links too much.
The result is that it forces links to live on the edge of overload
all the time. Router vendors have been forced to configure routers
with large buffers so that they can keep the link busy, even when
TCP flows cut back their sending rate in response to detected
congestion. Buffers in WAN routers generally have enough data
to keep the link busy for about half a second - that's about
500 MB of buffer for a 10 Gb/s link. RED attempts to reduce the
required buffer size by randomly discarding IP packets when the
average buffer occupancy gets above a certain level. It does
not discard all packets, just a random subset. The objective is
to get some of the flows to slow down, but not so many that the
buffer drains too fast and underflows. Since high rate flows
send more packets, the high rate ones tend to detect congestion
earlier (because their packets are lost) and slow down. This
is generally a good thing. Weighted RED is an attempt to allow
different flows to have different discard probabilities during
congestion. In my opinion, RED is a kludge that attempts to
fix a fundamentel weakness in TCP congestion control. A better
approach is to move to something like ABR explicit rate control.
The performance is much better and it can make routers cheaper,
since it reduces the buffer requirements by orders of magnitude.
But it is not likely in the IP world, where the assumption of
a dumb network layer is so intrinsic, that it's very hard to
move forward, even when there are obviously superior technical
solutions.

Jon