Tail Drop vs. Full Drop

[b.gamble]

Preparing for 642-642 QoS exam.

When WRED passes its maximum threshold are packets tail dropped or full dropped? The QoS book by Odom suggests they're "technically" full dropped, but other studying shows they might actually be tail dropped. I realize it's splitting hairs, but on exams, you spend a great deal of time splitting hairs.

[Joseph W. Doherty]

I don't recall ever seeing a "technically" full dropped, so unsure what Odom means by this. Also just did a quick Google for this term, didn't find a reference.

Wred (or RED) drops all packets that exceed max theshold, and since the are being added to the tail of the queue, I would say they are tail dropped. The only real difference between this and FIFO tail drop, RED packet count, that's being compared to the threshold, it's a moving average, not an actual queue count.

[b.gamble]

Pg. 427, 2nd Ed.

"Full Drop - State in which RED's average queue depth exceeds the maximum threshold. All packets are dropped."

Pg. 425, 2nd Ed.

"...to distinguish between true Tail Drop, and the case when the RED average queue depth exceeds the maximum threshold, RED calls this action category Full Drop."

[Joseph W. Doherty]

Did some more searching and did find references to RED "full drop", at least with Cisco equipment. Also found one reference that consider RED "full drop" also "tail drop".

It seems the "full drop" is used to distinguish between RED no-drop or random-drop. Also used because, at least in what I was just reading, "tail drop" implies a full queue (although I think that's debatable).

If this terminology is unique to Cisco(?), and if you're working to pass a Cisco QoS test, then it makes sense to call such drops "full drops".

[c.captari]

The full drop is the same thing as tail drop.

"When the average queue

length is greater than the user-specified minimum threshold, WRED begins to randomly drop

packets (both TCP and User Datagram Protocol [UDP] packets) with a certain probability. If

the average length of the queue continues to increase so that it becomes larger than the userspecified

maximum threshold, WRED reverts to a tail-drop packet discard strategy, in which all

incoming packets may be dropped."

However note that by the nature of how WRED works is hard to reach the tail drop state. WRED in itself tries to avoid going to full drop/tail drop by dropping packets between Min/Max Thereshold margins.

Therefore i think:

tail drop = full drop.

Early drop = Weighted Random Early drop

[Giuseppe Larosa]

Hello Benjamin,

when average queue depth > max threshold for a traffic class the packet of this class has 100% probability to be dropped.

Another packet in another traffic class sitting near that first packet if the average queue depth is less then max threshold for its traffic class hasn't 100% probability to be dropped as the first one.

With WRED Weighted RED the key is differentiating the thresholds for different traffic classes in a DiffServ QoS Model.

full drop = 100% drop probability

but it is class dependent.

tail drop = 100% drop probability for every packet arriving with an interface queue >= max queue

this is indifferentiated.

From this they introduced the term full drop for RED.

Hope to help

Giuseppe

[Joseph W. Doherty]

Giuseppe touches upon an interesting point when he writes: "Another packet in another traffic class sitting near that first packet if the average queue depth is less then max threshold for its traffic class hasn't 100% probability to be dropped as the first one." This because WRED unlike RED can have application of drops when avg. queue depth exceeds max theshold where there can be multiple and different max thesholds (e.g. one per ToS marking). Perhaps this is why it's called full drop since it's possible other (marked) packets are still be added to the queue's tail.