Samsung Ubigate iBG3026 Configuration Manual page 226

CHAPTER 19. QoS
If the classes are always serviced in the same order in the first pass(and hence
in the second pass as well) it will lead to unfair servicing of the equal priority
classes. To solve this problem of unfairness, the 'head' of the scheduler list is
advanced to point to the next class(in the circular list) for the next scheduler
interval. As a result each class experiences the same average queuing delay.
Since the scheduling order in the first pass also determines the order in the
second pass, rotation of the list head also happens for the 'xs-bw-list'.
Therefore, each class receives an equal share of the excess bandwidth on an average.
In the first pass, the scheduler services the list at the highest priority level first,
followed by the list at the next lower priority level and so on. Classes in a
circular list(having the same priority) are serviced in the same manner as
explained above, by advancing the list-head around the circular list every
scheduling interval. Classes at a higher priority will always be serviced before
those at a lower priority level. The 'xs-bw-list' formed during this first pass
will contain classes in the same order in which they were serviced in the first
pass. Therefore, a higher priority class will always be serviced prior to a lower
priority class during the both the first and second pass. Consequently its traffic
will experience lower average latency and will also have access to any excess
bandwidth before the traffic in the lower priority class.
When CBQ is enabled, the backpressure from packet driver is automatically
handled in CBQ scheduling mechanism. As the CBQ scheduling happens at
physical port level which as an associated driver queue. The logical driver
queue thresholds and queue depth is maintained at port level. So before
dequeuing packet from CBQ queues, the scheduling mechanism make sure
that driver queue has enough capacity to accommodate it.
Following are the main points of the CBQ scheduling algorithm, which is also
referred to as DRR-P(DRR with Priority).
DRR-P can guarantee bandwidth for each class/queue.
The algorithm offers bandwidth based on bytes and not packets.
Therefore It does not allow flows with bigger packet sizes to dominate.
The scheduler runs periodically every 5 ms. the quantum(in DRR
terminology) to be sent from each class in each pass is accordingly set.
This limits the queuing delay experienced by each queue/class because each
class is serviced at least once every 5ms.
The algorithm supports prioritization in terms of both latency and
bandwidth borrowing.
A class can be rate limited to a configured maximum value.
188
© SAMSUNG Electronics Co., Ltd.