On 12/4/2012 1:07 AM, Jose Saldana wrote:
Hi all. I am following the discussion about buffers and AQM, so here go my 2
cents:
In our research group in the University of Zaragoza, we have studied the
effect of different buffer policies on some services, mainly real-time ones
based on UDP (VoIP and networked games), when competing with other traffic.
We have also calculated the results in terms of subjective quality
estimators.
Although we have not used AQM, we have used different FIFO buffer
implementations:
- Buffer measured in bytes
- Buffer measured in number of packets
- Time-limited buffer
A summary of some of the conclusions:
- In general, big buffers are bad for real-time traffic (of course).
- If the buffer is measured in bytes, it is better for real-time services:
they use small packets, and the drop probability is higher for big packets.
This depends on whether the resources are byte-based or packet-based.
Byte-based examples include bandwidth (when supporting variable-sized
frames at L2), encryption/authentication of the entire packet, and
memory (using variable-sized buffers).
Packet-based include header processing overheads, fixed-frame network
capacity, and fixed-size buffers.
How different traffic reacts to limited resources depends on how these
resources are implemented, and which resources are limited/constrained.
There's no single answer based solely on traffic.
Joe
