Queue Scheduling Disciplines

1
Queue Scheduling Disciplines
Ref White Paper from Juniper Supporting
Differentiated Service Classes Queue Scheduling
Disciplines by Chuck Semeria
2
What does a Queue scheduling discipline do?
Objectives of an ideal Queue scheduling
discipline-

• Fair Distribution of bandwidth to each of the
  different service classes.
• Protection between different service classes on
  an output port.
• Provision for a service class to use unused
  bandwidth of another service class.
• Algorithm that can be implemented in the hardware.

3
Classic Queue Scheduling Disciplines

• First-in-first-out Queuing (FIFO)
• Priority Queuing (PQ)
• Fair Queuing (FQ)
• Weighted Fair Queuing (WFQ)
• Weighted Round Robin Queuing (WRR)
• Deficit Weighted Round Robin Queuing (DWRR)

4
FIFO

• Also known as First-come,First-served (FCFS)
• Single Queue
• Packets serviced in the order in which they were
  received.

5

• Benefits
• extremely low computational load
• behavior of the queue is very predictable-maximum
  delay determined by the maximum depth of the
  queue.
• as long as queue depth remains short, it
  provides simple contention resolution for network
  resolution without significant addition of
  queuing delay at each hop

6

• Limitations
• A single FIFO queue doesnt allow organizing of
  packets. (Cannot differentiate packets from one
  class of service from another)
• Impacts all flows equally. (mean queuing delay
  increases as congestion increases- not suitable
  for real time applications)
• During congestion UDP flows are benefited more
  than TCP. (TCP flows may experience increased
  delay, jitter and reduction in the bandwidth
  consumed)
• Bursty flow can consume the entire buffer space
  and deny service to other flows.

7

• Implementations
• generally supported on an output port only when
  no other queue scheduling discipline is
  supported.

8
Priority Queuing (PQ)

• packets are classified and placed into various
  priority queues.
• packets are scheduled from the head of a given
  queue only if all queues of higher priority are
  empty.
• within each PQ, packets serviced in FIFO order.

9

• Benefits
• low computational load
• allows organizing packets and treating of
  different service classes differently.

10

• Limitations
• lower priority traffic may experience excessive
  delay.
• starvation for lower priority traffic.
• misbehaving high-priority flow can affect other
  flows sharing the same queue.
• cannot solve the problem of UDP flows favored
  over TCP flows during congestion, seen in FIFO.

11
Implementations and Applications

• PQ generally configured in 2 modes.
• Strict priority Queuing
• Rate controlled Priority Queuing
• Applications of PQ
• can enhance network stability during periods of
  congestions.
• supports the delivery of a high-throughput,
  low-delay, low-jitter and low-loss service class.
• (provided traffic conditioning is done at edges
  to prevent higher-priority queues to be
  over-subscribed.)

12
Fair Queuing (FQ)/Per-flow Queuing/Flow based
Queuing

• Packets are classified into flows by the system
  and then assigned to a queue that is specifically
  dedicated to that flow.
• Queues are serviced one packet at a time in
  round-robin fashion.

13

• Benefits
• a misbehaving flow does not affect other flows.
• Limitations
• implementations are in software.
• doesnt support flows with different bandwidth
  requirement.
• flows containing large packet sizes get a larger
  size of output port bandwidth than flows with
  smaller packet sizes.
• sensitive to order of packet arrivals, may
  missing a whole cycle
• no mechanism to support real-time services.
• its not easy to classify network traffic into
  well defined flows.
• generally cannot be configured in core routers
  because of the large number of potential queues
  than need to be supported.

14
Implementation and Applications

• typically applied at the edges of networks.
• class based FQ
• Output port divided into number of service
  classes
• Each class allocated a user-configured percentage
  of output port bandwidth.
• FQ is applied within each bandwidth block.

15
Weighted Fair Queuing (WFQ)

• Basis for a class of queue scheduling disciplines
  to address limitations of FQ.
• Supports flows with different bandwidth
  requirements by giving weights.
• Supports variable length packets so that flows
  with larger packets are not allocated more
  bandwidth than flows with smaller packets.

16
WFQ approximates a weighted bit-by-bit
round-robin scheduling discipline.
The order in which each packet is fully
reassembled is determined by the order in which
the last bit of each packet is transmitted, known
as finish time
17

• WFQ approximates this discipline by calculating
  and assigning a finish time to each packet.
• Scheduler selects and forwards the packet which
  has the earliest finish time.

18

• Benefits
• provides protection to each service class by
  ensuring a minimum level of output port bandwidth
  independent of the behavior of other service
  classes.
• With traffic conditioning at the edges of a
  network, WFQ guarantees a weighted fair share of
  bandwidth to each service class with bounded
  delay.

19

• Limitations
• Vendor implementations are software based and
  not hardware based.
• misbehaving flow with in the service class can
  impact other flows with in the same service
  class.
• complex algorithm. Needs maintenance of
  per-service class states and iterative scans of
  states on each packet arrival and departure.
• computational complexity impacts scalability.
• on high speed interfaces, minimizing delay to
  the granularity of a single packet transmission
  considering the computational expense.
• even if delays are bounded they can be very high.

20

• Enhancements
• Class based WFQ
• Self-clocking Fair Queuing (SCFQ)
• Worst case Fair weighted Fair Queuing (WF2Q)
• Worst case Fair weighted Fair Queuing (WF2Q)

21

• Implementations and Applications
• Deployed at edges of networks to provide a fair
  distribution of bandwidth among a number of
  various service classes.
• WFQ can be configured to classify packets into a
  relatively large number of queues using a hash
  function.
• WFQ can be configured to allow the system to
  schedule a limited number of queues that carry
  aggregated traffic flows.
• Class based WFQ can be alternately be used to
  schedule a limited number of queues that carry
  aggregated traffic flows.

22
Weighted Round Robin (WRR)/ Class-based Queuing
(CBQ)

• addresses the limitations of FQ model by
  supporting flows with significantly different
  bandwidth requirements.
• Addresses the limitations of the strict PQ model
  by ensuring that lower-priority queues are not
  denied access to buffer space and output port
  bandwidth.
• Algorithm
• Packets are classified into various service
  classes and then assigned to a queue that is
  specifically dedicated to that service class.
• Each of the queues are serviced in a round robin
  order.

23

• The allocation of various amounts of bandwidth to
  various service classes are done by either
• Allowing higher bandwidth queues to send more
  than a single packet each time that it is visited
  during a service round.
• Allowing each queue to send only one packet each
  time it is visited, but to visit higher-bandwidth
  queues multiple number of times in each service
  round.
• The delay experienced by packets, the jitter
  experienced by packets and the packet loss in
  each queue can be tuned .

24
(No Transcript)
25

• Benefits
• Can be implemented in hardware.
• Provides coarse control over the percentage of
  output port bandwidth allocated to each service
  class.
• Ensures that all service classes have access to
  at least some configured amount to bandwidth to
  avoid starvation.
• Provides an efficient mechanism to support the
  delivery of differentiated service classes to a
  reasonable number of highly aggregated traffic
  flows.
• Classification of traffic by service class
  provides more equitable management and stability
  for network applications than the use of
  priorities or preferences.

26

• Limitations
• provides the right percentage of bandwidth class
  only if all of the packets in all of the queues
  are of the same size or if the mean packet size
  is known in advance.

• Implementations and Applications
• It can be deployed in both the core and edges of
  the network.

27
Deficit Weighted Round Robin(DWRR)

• Addresses limitations of WRR by supporting the
  weighted distribution of bandwidth in case of
  variable-length packets.
• Unlike WFQ this can be implemented in the
  hardware and so can be used on high-speed
  interfaces.
• Each queue is configured with a number of
  parameters.
• A weight that defines the percentage of output
  bandwidth allocated.
• A DeficitCounter that specifies the total number
  of bytes the queue is permitted to transmit each
  time it is visited by the scheduler.
• A quantum of service proportional to the weight
  of the queue. The DeficitCounter for a queue is
  increased by quantum each time it is visited by
  the scheduler.

28

• Algorithm
• Scheduler goes to each non-empty queue and
  determines the number of bytes in the packet at
  the head of the queue.
• The variable DeficitCounter is incremented by the
  value quantum.
• If size of the packet at the head of the queue is
  greater than the variable DeficitCounter
  scheduler moves on to next queue.
• If less than or equal to the DeficitCounter then
  packet is transmitted to the output port and the
  DeficitCounter is reduced by that amount.
• If queue is empty the DeficitCounter is made zero.

29
(No Transcript)
30

• Benefits
• Provides protection among different flows.
• Provides precise control over the percentage of
  output port bandwidth allocated to each service
  class in case of variable-length packets.
• All service classes have access to at least some
  configured amount of output bandwidth.
• Simple and inexpensive algorithm with no
  maintenance of lot of per-service class state .

31

• Limitations
• Because of highly aggregated service classes, a
  misbehaving flow within a service class may
  impact other flows in that service class.
• No end-to-end delay guarantees.
• May not be accurate.