Designing an Adaptive Scheduler for Latency Critical Traffic

1
Designing an Adaptive Scheduler for Latency
Critical Traffic

• Hamed Khanmirza
• Advisor Dr. Nasser Yazdani
• Router Laboratory, ECE Dept. Univ. of Tehran
• 31st Shahrivar 1383 Sep. 21st 2004

2
Outline

• QoS Definition
• Differentiated Service Architecture (DiffServ)
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

3
Quality of Service (QoS)

• The capability to control traffic handling
  mechanisms in the network such that the network
  meets the service needs of certain applications
  and users subject to network policies.
• QoS Parameters
• Bandwidth
• Latency
• Jitter
• Loss

4
QoS Components

• Congestion Avoidance (Queue Management)
• Try to inform the sources congestion conditions
• Resource Allocation
• When the congestion occurred which flow should
  use how much of resources

5
Outline

• QoS Definition
• Differentiated Architecture (DiffServ)
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

6
DiffServ Architecture

• Ingress routers (Edge Routers)
• Perform per aggregate shaping or policing
• Mark packets with Code Points, each CP represent
  a Class of Service (DSCP DiffServ Code Point)
• Core routers
• Implement Per Hop Behavior (PHB) for each DSCP
• Process packets based on DSCP

DS-2
DS-1
Ingress
Egress
Ingress
Egress
Edge router
Core router
7
Expedited Forwarding (EF)

• Provides the abstraction of a virtual Leased
  Line between an ingress and an egress router
• Network
• No loss
• low delay jitter
• User
• Send traffic based on SLA
• Excess traffic is delayed, and dropped when
  buffer overflows

8
Assured Forwarding (AF)

• Possible service
• strong assurance for traffic within profile
• Out-of-profile traffic will be marked as lower
  class (i.e. BE)
• Network
• lower loss rate than best-effort
• In case of congestion best-effort packets are
  dropped first

9
EF Traffic

• Due to timing constraints EF traffic has small
  queues (2 10 packets).
• Queue Manager has not significant role on EF
  traffic
• The most important role is Scheduler.
• Schedulers should be
• Simple
• accurate

10
Outline

• QoS Definition
• QoS Standard Frameworks
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

11
Strict Priority Queue (PQ)

• Sends from prioritized queue until it has packet
• Advantages
• Very Simple
• Minimum possible delay and jitter
• Exist in all NPs
• The most wide-spread method
• BUT
• Starvation
• Generating Burst

12
PQ Behavior (Burst Generation)
13
PQ Behavior (Burst Generation)
14
Alternate Priority Queue (APQ)

• Alternates over high priority and low priority
  queues
• Advantages
• Reduces the threat of starvation
• Burst Length is always one!
• BUT
• Very fragile
• Limited range of rates and packet sizes

15
Deficit Round Robin
DC 400
500
200
64
700
DC 500
v
100
100
100
100
DC 700
512
1024
Quantum 1500
16
Deficit Round Robin
DC 400
500
200
64
700
DC 100
DC 700
v
512
1024
Quantum 1500
17
Deficit Round Robin
DC 400
500
200
64
700
DC 100
DC 188
1024
Quantum 1500
18
Deficit Round Robin
DC 1900
v
500
200
64
700
DC 0
DC 1688
1024
Quantum 1500
19
DRR

• Is simple (O(1))
• Fair enough
• BUT
• For Sensitive traffic ?
• DRR
• Give high quality traffic strict priority over
  others
• Have the same problems as PQ!

20
DRR
500
250
QLC 1000
QBE 1500
DC 400
500
200
64
700
DC 500
100
100
100
100
DC 700
512
1024
21
DRR

• Advantages
• Reduces delay comparing with DRR
• Reduces burst comparing with PQ
• BUT
• Sensitive to transient bursts
• Hard to configure (QLC)
• Esp. for TCP traffic in the core!
• worst-case condition configuration

22
Weighted Fair Queue (WFQ)

• Family of GPS-based (General Processor Sharing)
  methods
• In WFQ
• We keep Virtual Time as a service level
• When a packet arrives it is marked with two
  labels
• Si max (Ai , Fi-1)
• Fi Si ( Li / W)
• The packet with the smallest Fi is selected to be
  sent

23
WFQ

• Very bounded guarantees
• Have equivalent Jitter values as PQ
• BUT
• Have larger delay
• To achieve better results the EF queue should
  have larger weights that approaches to PQ
  behavior
• Very complex
• Sorting
• Calculating Virtual Time

24
Outline

• QoS Definition
• QoS Standard Frameworks
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

25
Adaptive Priority Queue (ADPQ)

• ADPQ tries to provide
• Delay guarantee
• Bandwidth Guarantee
• No Starvation!
• And also, generate shortest burst

26
ADPQ

• ADPQ sends K packets from High Priority (HP)
  queue and one quantum from Low Priority (LP)
  queues.
• The amount of K is determined dynamically
• Parameters
• C Line Capacity
• D Max tolerable delay
• LEF Effective Packet Length

27
ADPQ Theory

• Q(t) HP queue length (in packets)
• if Q(t) N K
• QBE LQ quantum

28
ADPQ Theory

• ADPQ is conformant to RFC 3246
• RFC 3246 describes EF schedulers attributes
• Also, we know that EF traffic can not consume
  more than

29
ADPQ Algorithm

• How and When K is updated ?
• Instantaneous queue length
• Increasing K inappropriately
• Configured delay is violated
• Send all the packets, K is calculated for them
• Very fragile!
• Longer bursts in response of relatively long
  arriving burst
• Not aware of previous set of packets

30
ADPQ Algorithm

• Send all the packets, K is calculated for them,
  calculate new K in background and assign it when
  K is increased to dangerous condition or all of
  the previous set was sent (The first version of
  ADPQ)
• Using Instantaneous queue length arise some
  problems ? use average length
• Calculating K with every enqueue has processing
  burden

31
ADPQ Algorithm

• Averaging Period?
• Longer Periods
• We have to provide strict delay guarantee
• Short Periods
• It is the best solutions BUT how much short?
• Averaging Duration?
• Every n enqueue
• From When?
• Last m observations

32
ADPQ Algorithm
Packet Length 1
Packet Length 2
Packet Length 3

• The calculated average is applied to threshold
  array to find appropriate K (the index of the
  array)
• The second part indicates whether the size of the
  queue has increased, decreased or remained
  unchanged

33
ADPQ Algorithm

• ADPQ holds packets (like a shaper) , it means do
  not send packets as soon as their arriving (like
  PQ)
• Then, encountering next relatively long burst,
  ADPQ have to increase K significantly at once
• The generated bursts become worse than PQ!!

34
ADPQ Algorithm

• Solution
• Anticipate the next arriving burst,
• BUT HOW?
• Determine traffic model (CBR, )
• Consider queue length change
• The m indicators give us a pattern. Based on the
  patterns that shows how the queue is changed from
  the previous observation increase K anticipatory.

35
Pattern Codes

• //0 Unchanged - 1 Increased - 2 Decreased
• Category 1 000, 100
• Category 2 010, 001, 210
• Category 3 121, 211, 011, 110, 101, 201
• Category 4 111
• Category 0 Others

36
Patterns
37
Probability Calculation

• Decreasing probability with increasing K
• Considering equal thresholds

38
ADPQ

• Advantages
• Easy configuration
• Very small burst based on delay requirement and
  arriving traffic pattern
• Guarantees rate
• Guarantees delay
• No starvation threat
• Complexity
• Complex operations are in enqueue part
• Dequeue part is still simple

39
ADPQ Flowchart
Calculate Average and Find K
Calculate Probability
Enque count 4
New Enque
Find Patterns Category and its multiplicand
Calculate old and new Patterns
Calculate Final Probability
Increment K
Finish
Increment K
40
Outline

• QoS Definition
• QoS Standard Frameworks
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

41
Simulation Environment

• Simulated with NS2
• EF packet size 60 500
• ADPQ configured delay 3 ms

42
Determining LEF
43
LEF Effect
Maximum Burst length distribution
Maximum Delay
44
Averaging Duration
Rate change with varying average length
45
Averaging Duration
46
UDP Traffic
Comparison of bandwidth usage for EF traffic
47
UDP Traffic
48
UDP Traffic
49
Anticipation Effect
ADPQ treats similar to PQ with the same burst
length distribution, delay and rate
50
TCP Traffic
51
TCP Traffic
52
TCP Traffic
53
TCP Traffic
Number of Packet Drops (TCP)
Two more simulations are inside thesis.
54
Outline

• QoS Definition
• QoS Standard Frameworks
• Scheduling Mechanisms
• ADPQ
• Simulation Results
• Hardware Implementation
• Conclusion Future Works
• Publications

55
Hardware Implementation
56
Increment K
57
Increment K
58
Category Finder (1st Pulse)
59
Category Finder (2nd Pulse)
60
Probability Calculation

• To Calculate
• ROM-Based approach
• Calculate all possible values
• Multiply all in 10 to achieve separate whole
  numbers
• Store them in a ROM
• Delay 3.96 ns / Area 98000 gates
• Very accurate
• Requires 10-bit LFSR and 10-bit comparator

61
Probability Calculation

• Register-based
• Requires a 50-bit LFSR !
• A 50-bit comparator !

62
Comparator

• Ordinary comparator
• Delay 6.17 ns
• Area 14000 gates
• New method
• Delay 2.97 ns
• Area 30000 gates

63
Output Section
64
Priority Encoder

• Ordinary PE
• Delay 3.31 ns
• Area 4800 g.
• Hierarchical PE
• Delay 1.96 ns
• Area 14000 g.

65
Conclusion

• We introduced a new adaptive approach for
  latency-critical traffic which
• Can guarantee maximum delay
• Can guarantee a rate
• Can control output burst
• Starvation Free!
• Has easy configuration
• Relatively simple
• Has better results in face of UDP and TCP traffic
  comparing to other scheduling methods

66
Future Work

• Increasing performance of ADPQ against TCP
  traffic
• Extending ADPQ to multiple EF queues
• Designing of a general hardware framework for a
  variety of scheduling methods

67
Publications

• Hamed Khanmirza, Designing an Adaptive Scheduler
  for Latency-Critical Traffic, M.Sc. Thesis
• H.Khanmirza, S. Zarifzadeh, N.Yazdani,
  Comparison of Alternate Priority Queuing for
  Expedited Forwarding Per Hop Behavior, IST 2003,
  Isfahan
• Hamed Khanmirza, Sajjad Zarifzadeh, Naser
  Yazdani, ADPQ Adaptive Priority Queuing,
  APCC04 Beijing
• Hamed Khanmirza, Sajjad Zarifzadeh, Naser
  Yazdani, ADPQ Adaptive Priority Queue, A
  Simulation-based Analysis, ICCC04 Beijing

68
Publications

• Sajjad Zarifzadeh , Hamed Khanmirza, Naser
  Yazdani, A Multipath Algorithm for Premium
  Traffic Routing in DiffServ Networks, ICON04,
  singapore
• Sajjad Zarifzadeh , Hamed Khanmirza, Naser
  Yazdani,A Distributed Multipath Algorithm for
  Providing Bandwidth-Guaranteed Routing, APCC04,
  Beijing
• Hamed Khanmirza, Sajjad Zarifzadeh, Naser
  Yazdani,E-ADPQ Enhanced Adaptive Priority
  Queue, ready to submit for ICT05
• Hamed Khanmirza, Sajjad Zarifzadeh, Naser
  Yazdani, ADPQ Adaptive Priority Queue,
  Elsevier Journal

69

• Thank you for your attention