Review on
Stratified Round Robin: A Low Complexity Packet Scheduler with Bandwidth
Fairness and Bounded Delay
Through a background survey, we see there are a group of fair scheduling
algorithms all sharing same basic ideas:
Put flows into classes based on their weight approximating to some scale,
schedule flow classes using deadline based algorithms and then schedule
flows in each class using some variant of Round Robin. By hybriding these
two types of algorithms, the new scheduler can achieve approximate WFQ
performance (low delay and low burstiness) and RR complexity(O(1)). The
approximate sorting can be done in different method. This paper shows a
method fitting for simple hardware implementation.
Though the paper provides a practical implementation for the scheduler,
the intrinsic idea is not quite new. Some previous research papers have
alreay developed the similiar method with only a slight variant (Smoothed
RR, Aliquem, Bin sorting etc.). The paper does refer
these papers but doesn't compare with these works. A newer paper (FRR)
pursues the same
approach but uses other variant of the inter and intra packet scheduling
algorithms and proves to achieve better delay performance.
The paper claims that the timestamp-based algorithms need to sort packet
deadline and therefore suffer from complexity logarithmic in the number
of flow N. The paper fails to mention an efficient implementation of such
kind of algorithms with only O(1) complexity: approximate radix sorting
or timing wheels.
Generally, all deadline-based scheduling algorithms bound the queue head
packet's delay. I am not sure why the paper emphasizes the so called single
packet delay bound as if it is a unique feature of SRR. For example, the
paper said that its worst case single packet delay is independent of the
number of flow. Actually, the worst case delay is still decided by the
number
of flow. What's more, the bound appears to be too large comparing with the
GPS model
or WFQ.
Burstiness of Stratified RR is not as good as other similar algorithms
either.
In same class, the largest flow weight can be as large as two times of the
smallest flow weight. This means in one schedule slot this flow can send 2
times of size of the longest packet.
The paper is well written, easy to understand and practical for
implementation.
Considering its impact and originality to the research area, I vote it to be
ranked in the lower high class among the papers reviewed so far.
Haoyu Song