Scheduling Algorithms in High Speed Network

1
Scheduling Algorithms in High Speed Network

• 2000-20731 ???

2
Outline

• Introduction
• Goals
• Difficulties
• PFQ Algorithms
• Grouping architecture
• LBT / GBT
• For fixed length cells
• For variable length packets
• Timestamp Implementation
• Delay Result

3
Goals

• Support a large number of sessions with diverse
  bandwidth requirement
• Operate at very high speed
• Maintain important properties of GPS
• Guarantee an end-to-end delay to a
  leaky-bucket-constrained session, regardless of
  other sessions
• Ensure fair allocation of bandwidth

4
Difficulties

• Key difficulty with PFQ algorithms
• require buffering on a per session basis and
  non-trivial service arbitration among all
  sessions
• -gt Not scailable
• Arbitration needed(per session)
• Computation of the system virtual time
• Management of a priority queue
• Management of another priority queue to regulate

5
PFQ Algorithms(1)

• Complexity of calculating virtual time
• Accuracy and complexity trade-offs

6
PFQ Algorithms(2)

• Packet selection policy
• SFF(Smallest virtual Finish time First)
• SSF(Smallest virtual Start time First)
• SEFF(Smallest Eligible virtual Finish time First)
• Maintain a priority queue
• Number of entities in the priority queue is the
  number of active sessions
• Implementation complexity
• on-chip / off-chip

7
Outline

• Introduction
• Goals
• Difficulties
• PFQ Algorithms
• Grouping architecture
• LBT / GBT
• For fixed length cells
• For variable length packets
• Timestamp Implementation
• Delay Result

8
LBT / GBT

• LBT
• Bounds the differences of virtual start times
  between two sessions in the same group
• GBT
• Bounds the differences between the system virtual
  time and virtual start time

9
Grouping Architecture for ATM Network(1)
10
Grouping Architecture for ATM Network(2)

• Restriction
• only a fixed number of guaranteed rates are
  supported by the server
• Sort by start time in the group
• Finish time is calculated
• Fi Si L/ri

11
Grouping Architecture for ATM NetworkChoose Next
Eligible Cell(1)

• Consider only the packets in the scheduler(head)
• SSF
• smallest S(.) in the scheduler
• -gt smallest S(.) among all packet
• SFF
• packet in the scheduler
• -gt smallest F(.) within the group
• -gt smallest S(.) within the group (because of
  same rate)
• SEFF
• eligible packet exist in the group
• -gt head packet is also eligible

12
Grouping Architecture for ATM NetworkChoose Next
Eligible Cell(2)
13
Grouping Architecture for ATM Network
Maintaining Within Groups

• Maintain a priority queue with a simple linked
  list
• Local timestamp bound(LTB)
• Stail Shead lt L/r
• Situations when insertions into the list is
  needed
• (Re-calculate start time, finish time)
• After finishing services head session
• A new session joins
• Idle session becomes active

14
Grouping Architecture for Packet Network

• Problem variable packet size
• Fi - Si L/ri (L is not determined)
• Smallest Si --X-gt Smallest Fi
• 2-D sort is needed
• By Fi Si (service interval Fi)
• By Fi
• Grouping by service interval
• Same service interval -gt same Fi-Si
• Large number of unused group
• Sessions change groups due to change of packet
  length

15
Grouping Architecture for Packet
NetworkDiscretization of Service Interval

• Determine group
• group

16
Grouping Architecture for Packet NetworkProblem
with Discretization

• The maximum gap between time to be eligible and
  computed virtual start time ( )

17
Grouping Architecture for Packet NetworkSorting
Within The Group

• Aproximated finish time

18
Outline

• Introduction
• Goals
• Difficulties
• PFQ Algorithms
• Grouping architecture
• LBT / GBT
• For fixed length cells
• For variable length packets
• Timestamp Implementation
• Delay Result

19
Timestamp Implementation(1)

• Size of timestamp determines
• The range of supportable rates
• The accuracy
• Memory requirements (bandwidth / storage space)
• Size of timestamp
• 1-bit larger than the largest service interval of
  the largest rate in the system
• Pf)
• With globally bounded timestamp property,
• largest gap of timestamps lt L/r (service
  interval)
• (n1 bit) needed for (2n) difference

20
Timestamp Implementation(2)

• Representations of timestamp
• V(t)429496678410 , high rate session i, low rate
  session j
• Fixed point
• Floating point

21
Timestamp Implementation(3)

• Compressed timestamp
• Sj V(t) Fj
• We only store CSj SjMej1ej

22
Delay Result

• ATM scheduling architecture(Proposed in this
  paper)
• A cell of session i will miss its deadline by no
  more than
• (transmit time of a ATM cell
• service interval of session I)
• Packet network architecture
• A packet of session j will miss its deadline by
  no more than
• (transmit time of maximum length packet
• the maximum gap between time to be eligible
  and computed virtual start time
• the maximum inflation of virtual finish time)

23
Summary

• Scalability problems with PFQ Algorithms
• Per session management
• Propose grouping architecture
• For fixed length cell grouping rate
• For variable length packet grouping rate and
  length
• Propose time stamp optimization
• Calculate minimum bit
• Propose compressed timestamp
• Delay Result