WRED and MaxThreshold

[tomek0001]

Hello,

I was reading in one of my training manuals that if the average queue size is above the MaxThreshold but less then the total queue size packets will get dropped with MarkPrabobiltity and if the average queue size is above the max threshold and the max queue size all packets are tailed dropped.

I've never heard of the total queue size having to do with anything, but only that after the average queue size reaches the max threshold all packets are dropped.

Can anyone confirm of comment on this?

Thank you,

[Giuseppe Larosa]

Hello Tomacs,

the objective of WRED is that of preventing indifferentiated tail drops and to provide differentiated treatment near congestion to different traffic classes.

However, the software queue used to hold packets has a finite size, so if the queue size is full even if WRED is configured indifferentiated tail drops happen.

I remember a test performed on POS STM-1 interface where very high WRED thresolds were used for all classes.

The test was performed in a lab with a traffic generator not  acting as true TCP speaker.

The end result was tail drop on all classes without better treatment for some traffic types.

There is a limit that is given by the packet memory size in a linecard or associated to an interface.

So I would read that comment just as a warning that remembers that WRED to be effective has to use thresholds low enough to prevent the total queue to be filled.

To be noted testing WRED is among the most difficult tests that could be done specially if trying to emulate a population of true TCP sessions

Results vary from test to test when looking at TCP goodput

Hope to help

Giuseppe