Current fair allocation mechanisms do not increase the
number of receivers with good reception quality. This can be achieved using
SAPRA. In this section we show that SAPRA is still fair allocating resources
for non-scalable sessions with one receiver, approximating the efficiency
of other fair allocation mechanisms.
We compare SAPRA with FIFO, Random Early Detection (RED), RED with preferential
dropping (RED-PD), Flow Random Early
Drop (FRED), Deficit Round Robin (DRR) and CSFQ queueing mechanisms.
We use two variations of FRED, FRED-1 and FRED-2. The latter guarantees a
minimum number of buffers for each flow, having a better performance than
FRED-1 when there is a high number of flows. SAPRA and all these mechanisms,
with the exception of DRR, use first-in-first-out scheduling. DRR has its
packet scheduling algorithm. Although FRED uses first-in-first-out scheduling,
it makes a per-flow management of the queue.
We use the following topology:
scenario
The congested link have 1 ms delay and 10 Mbs of bandwidth. The link
queue has a size of 64 packets, and each packet has 1000 bytes. In these
simulations we use 33 sessions, each one with only one flow and one receiver.
For SAPRA we use 33 RTP sessions with one layer each. For
the other six algorithms we use 33 CBR unicast uni-rate flows. In both
cases sources generate flows (flow 0 to flow 32) where flow
(i) sends (i+1) times more than its fair share of 303 Kbs.
In the case of RED, FRED and RED-PD, the minimum threshold
is set to 16 KB and the maximum to 32 KB. In SAPRA, the variable B
used to estimate the average rate is set to 100 ms and the filter
ter base dropping index is set to 0.5%. In CSFQ the variable K , K
a and K
c are also set to 100 ms.