
PXB 4330 E
ABM Buffer Configuration
Semiconductor Group
16
Application Note 11.98
threshold. Then cells might not be accepted by the relative threshold even if the absolute
buffer fill is far below maximum.
For each traffic class two main thresholds must be defined: the absolute threshold with
either
BufNrtMax
or
BufNrtEPD
and the relative threshold
TrafClassMax
. The
combined threshold
SbMaxEpdCi
is a relative threshold to protect the total buffer from
the unlimited growth of a single scheduler in case of congestion/errors.
In
Table 3
the counters include nrt cells or rt and nrt cells of the total buffer (see column
"Fill Level" in
Table 3
). That means these counters include all cells of all traffic classes
in all schedulers and all queues per ABM core. So these counters represent absolute fill
levels as they reflect the current occupation by rt/nrt cells of the whole buffer (and not of
a scheduler, traffic class or queue).
Table 3
shows that the thresholds BufNrtMax and BufNrtEPD correspond with the
counters UNRTOC/DNRTOC. These counters represent the current fill level of all nrt
cells in the total buffer (see column "Fill Level" of
Table 3
). BufNrtMax/BufNrtEPD
represent the maximum allowed number of buffered cells of this traffic class i (see
column "Specific Area" of
Table 3
) within the total buffer.
The EPD threshold is depicted with an uncertainty field in
Figure 3
to indicate the
possibility of a delayed response of the EPD threshold. This happens if the EPD
threshold is exceeded just after the first cell of a packet and therefore only the next
packet can be discarded by EPD. The worst case occurs when many VCs of a traffic
class start to send simultaneously barely below the EPD threshold. Then their first
packets are not discarded by EPD as the EPD threshold wasn’t exceeded when their
respective first cells arrived at the ABM. To accommodate this worst case scenario it is
necessary to define in addition to the EPD threshold a maximum fill/PPD threshold per
traffic class. This limits the maximum possible buffer occupation in the worst case
scenario.
Why then is an EPD threshold required The advantage of the EPD threshold is that it
discards the whole packet whereas the PPD threshold discards only a part of a packet
i.e. the leading part of this packet was still sent to the destination. This leads to an
unnecessary switch load as the already sent leading part of a discarded packet is sent
again by the necessary retransmission of the discarded packet. Hence PPD tends to
generate larger switch loads than EPD.
Another essential advantage of EPD consists in making possible a good throughput in
case of an overload situation. For instance, a 10 Mbit/s link is confronted with a
100 Mbit/s load during a certain time period. With only a PPD threshold the throughput
(= rate of whole packets which arrives the destination) probably drops nearly to 0 Mbit/s
due to the fact, that by exceeding a PPD threshold still partial packets are sent to the
destination. An additional EPD threshold (which must has a sufficient distance to the
PPD threshold in order to work properly) preserves the throughput near by 10 Mbit/s as
it discards the whole packet, if the EPD threshold is exceeded. On the other hand even
if the EPD threshold is only barely fallen short of at the arrival of the first cell of a packet,