Study on Proportional Differentiated Services by larryp


									Study on Proportional Differentiated Services

                Jan. 26, 2000

              Kun Suk Kim
          Real-Time System Lab.
                CISE, UFL
Key Paper - Proportional Differentiated
Services: Delay Differentiation and Packet

C. Dovrolis, D. Stiliadis, P. Ramanathan
          Internet Traffic Control
• Same-service-to-all model
• Integrated services (Intserv)
• Differentiated Services (Diffserv, DS)
   – Absolute Diffserv
   – Relative Diffserv
        • Strict Prioritization
        • Price Differentiation
        • Capacity Differentiation
        • Additive Differentiation
        • Proportional Diffserv
            – Backlog-Proportional Rate (BPR) scheduler
            – Waiting-Time Priority (WTP) scheduler
                 In Key Paper
• Proportional differentiation model
  - Tuning knobs (independent of the class loads)
  - Queueing-delay differentiation only
    cf) coupled delay and loss differentiation – future work
• Dynamics
• Feasible condition
• Two packet schedulers in heavy-load condition, in short
• Per-hop and class-based mechanism can provide consistent
  end-to-end differentiation to individual flows from
  different classes.
     Integrated Services (Intserv)
• Reservations-based architecture: buffers, link bandwidth
• End-to-end performance guarantees for individual flows
• Call setup (call admission)
   – Traffic characterization and specification of the desired
   – Signaling for call setup: RSVP
   – Per-element call admission
• Two major classes of service
   – Guaranteed Service: provides firm bounds in the queueing delays
     that a packet will experience in a router
   – Controlled-Load Service: receives “a quality of service closely
     approximating the QoS that same flow would receive from an
     unloaded network element”
• Difficulties:
   – Deployment and scalability of RSVP
   – Requirement for an interdomain policy and pricing infrastructure
   – Mapping between application and network service parameters
            Intserv - Call Setup Process

Figure from
    Intserv - Per-element Call behavior

Figure from
       The Essence of RSVP
• Reservations for bandwidth in multicast
• Receiver-oriented
   RSVP: Multicast- and Receiver-Oriented

Figure from
                         RSVP Example


                         3Mbps        Reserve

Modified figure from
  Differentiated Services (Diffserv, DS)

• Per-Hop Behaviors (PHBs): local (per-hop)
  service differentiation for large aggregates of
  network traffic
• Scalability
• Flexible service model
• Better-than-best-effort service to applications, w/o
  the need for host RSVP signaling
          Diffserv - A Simple Scenario

Figure from
                    Diffserv - DS Field
  • Differentiated Services (DS) field
       – IPv4 Type-of-Service (TOS) field
       – IPv6 Traffic Class Field
  • Differentiated Service Code Point (DSCP) subfield:
    determines per-hop behavior
  • CU subfield: Currently Unused

Figure from
     Diffserv - Simple Packet Classification
        and Marking at the Edge Router

Figure from
    Diffserv - Logical View of Packet Classification and
          Traffic Conditioning at the Edge Router

Figure from
   Diffserv - Per-Hops Behavior at the Core Router

• Can result in different classes of traffic receiving different
• Defines differences in performance among classes (but not
  mandate any particular mechanism for achieving these
• Differences in performance must be observable, and hence
 Approach 1: Absolute Differentiated Services

• User receives an absolute service profile (eg. a certain
  bandwidth) from the network
• Premium service (performance level like leased-line)
• Assured service
  - Drop-preference: In (higher priority), Out (discarded
  with higher prob. than In when congestion occurs)
• Open Question: end-to-end performance that users expect?
• Trade-offs: high service assurance, coarse spatial
  granularity, high network utilization
• Route pinning
Approach 2: Relative Differentiated Services

•   N classes of service
•   Class i is better (or at least no worse) than class (i-1)
•   Local metrics for the queueing delay and packet losses
•   Class selector PHBs
    - No admission control and resource reservations
    - Network assures users and applications that higher classes
    will be relatively better than lower classes
    - Users and applications select the class w.r.t. requirements,
    cost, policy constraints
                 Approach 2 (Cont‟)

• End-system adaptation: provides the choice of the service
  class as an additional dim. in the end-system adaptation
• Similar with postal service system
  - several priority classes w/o admission control, end-to-end
  resource reservations, or service guarantees
                 Approach 2 (Cont‟)

• Differentiation based on appropriate pricing, or careful
  capacity provisioning
  - Cannot always provide consistent differentiation between
  classes (not predictable)
• Strict prioritization between classes
  - Consistent differentiation
  - Does not adjust the quality spacing between classes (not
  controllable, no tuning knobs)
• Premise: predictable and controllable
    Other Relative Diff. Models

•   Strict Prioritization
•   Price Differentiation
•   Capacity Differentiation
•   Additive Differentiation
         Strict Prioritization
• The highest backlogged class is serviced
  first (delay aspect)
• The lowest backlogged class is dropped first
  (loss aspect)
• No means for adjusting the quality spacing
  between classes
• Starvation effects
          Price Differentiation
• Pricing mechanism
• Paris Metro Pricing (PMP)
  - higher price will lead to lower loads in the higher
  classes, i.e., better service quality
  - Not effective over relatively short timescales
  - Inconsistent (unpredictable) class diff.
                      Capacity Diff.
• Allocates resources between classes so that higher classes
  have more bandwidth and packet buffers than lower
• Weighted Fair Queueing (WFQ)
   – Distribute the link bandwidth between classes, so that the ratio of
     service-to-arrival rates for higher classes is larger
   – Although the bandwidth diff. is controllable, the delay diff. is not
   – Cannot provide consistent diff. in relatively short timescales
     because the forwarding resources allocated to each class do not
     follow the actual class load variations
• cf) Forwarding mechanisms (packet scheduling and buffer
   – dynamic assignment of forwarding resources
                        Additive Diff.
• Additive constraints for long-term avg. queueing delays in
  heavy load
        d i  d j  Di , j  0( j  i )
        Di , j : delay diff. parameter for classes i and j

• Priority scheduler
        pi (t )  wi (t )  si
        0  si  ...  sN : scheduler diff. parameter
        pi (t ) : priority of a packet in queue i at time t
        wi (t ) : the waiting-time of a packet in queue i at time t
   - Additive delay diff. in heavy-load conditions with
         Di , j  s j  si
   New Approach: Proportional Diff. Model

• The basic performance measures for packet forwarding
  locally at each hop are ratioed proportionally to certain
  class differentiation parameters
  - Queueing delay differentiation only
  - Coupled delay and loss differentiation - future work
• applicable to delay-sensitive applications
  - IP-telephony, video-conferencing, transaction-based
        New Approach (Cont‟)
• Dynamics
• Feasibility
• Packet scheduler (in heavy-load conditions, even
  in short timescales)
  - Backlog-Proportional Rate (BPR) scheduler
  - Waiting-Time Priority (WTP) scheduler
           Proportional Diff. Model
• Spaces certain class performance metrics
  proportionally to the diff. parameters
  qi ci
     (i, j  1,...,N )
  qj cj
  c1  ...  c N

  qi :     performance measure for class i
  ci :     quality differentiation parameters
 Proportional Diff. Model (Cont‟)
di : the avg. queueing delay of the class i packets

 di  i
 dj j
  i : DDPs( 1  ...   N  0)
  Proportional Diff. Model (Cont‟)
• Formation for a short-term queueing delay metric

   d i (t   )  i
   d j (t   )  j
    0 : monitoring timescale

• Unable to derive the feasibility condition for short
• Approximate the proportional diff. model in short
      Dynamics and Feasibility of
       Proportional Delay Diff.
• Lossless and work-conserving packet scheduler
  - N queues (one for each class)
  - Lossless: the offered load is less than the service
  - In practice, sources react to the ECN bit
  Dynamics and Feasibility (Cont‟)
i : avg. arrival rate in class i
  i 1 i : aggregate arrival rate in the system

• Assumption of a scheduling discipline for long-term avg.
  queueing delays (feasible)

    di  i
       (i, j  1,...,N )
    dj j
   Dynamics and Feasibility (Cont‟)

• Conservation law (if the packet length distribution is the
  same in all classes)

            i di   d ( )
      i 1

   d ( ) : avg. queueing delay if the aggregate traffic was
              serviced by a work-conserving FCFS server of
              the same capacity as the scheduler that
              enforces the proportional delay model
  Dynamics and Feasibility (Cont‟)

• Little‟s Law: avg. backlog in a work-conserving
  system is independent of the scheduling discipline
• Generally, not require the same packet length
  distribution in all classes
• d ( ) : depends on the traffic char. (eg, burstiness)
                             i d ( )
              di                           (i  1,..., N )
                      1     2          N
                    1   2  ...   N
                                       
• The average delay of a class i increases with the arrival rate of each
  class j.
• Increasing the load of a higher class causes a larger increase in the
  average delay of a class than increasing the load of a lower class.
• If the delay differentiation parameter of a class increases, the average
  delay of all other classes decreases, while the average delay of that
  class increases.
• Suppose that a fraction of the class i load switches to class j, while the
  aggregate load remains the same. The average delay of each class
  increases if i<j, and decreases if i>j.
•   A set of N avg. delays d  is feasible if and only if the
    following 2 N  2 inequalities hold,

                                   
            i di    i d   i ,   
                                   
           i        i     i  
      : the set of 2 N  2 nonempty proper subsets of  ,2,..., N 
• The average backlog of a subset of the N classes cannot be
  lower than the backlog of these classes in a FCFS server,
  independently of the scheduling discipline.
• DDPs might not be feasible under certain conditions on
  system load, the class load distribution, and the traffic
 Two Packet Schedulers for Relative
            Delay Diff.
• Designed for predictable and controllable relative delay
• Backlog-Proportional Rate (BPR) scheduler
• Waiting-Time Priority (WTP) scheduler
• Approximate the proportional delay diff. model in heavy-
  load conditions, even in short timescales
                 BPR scheduler
• Dynamically readjust the class service rates so that they are
  always ratioed proportionally to the corresponding ratios
  of class loads (class backlogs)
• Proportional constraint
      ri (t ) si qi (t )
      rj (t ) s j q j (t )
      ri (t ) : service rate assigned to queue i at time t
      qi (t ) : backlog of queue i at time t
      {si } : Scheduler Differentiation Parameters (SDPs)
           BPR scheduler (Cont‟)
• Work-conservation constraint

       r (t )  R
       i 1

      R:          Link capacity

• The DDP ratios tending to the inverse of the corresponding
  SDP ratios in heavy-load conditions

       di   i s j
              
       dj    j si
         BPR scheduler (Cont‟)
• When the relative backlog of a queue is quite small, the
  relative service rate given to that queue is also small.
   – The last few packets in a queue (either before the queue gets empty
     or before new arrivals occur) can experience a much larger delay
     than their predecessors
   – Sawtooth-type of variation in the queueing delays of consecutive
• Simultaneous queue clearing property
   – All queues that are backlogged during a busy period become
     empty at the same time
                  WTP scheduler
• Priority scheduler: priority of a packet increases
  proportionally with its waiting-time
• Priority of a packet in queue i at time t
      pi (t )  wi (t ) si
      wi (t ) : waiting-time of the packet at time t
• Time-Dependent-Priorities
    – The utility of this new priority structure is that it provides a
      number of degree of freedom with which to manipulate the relative
      waiting times for each priority group
         WTP scheduler (Cont‟)
• Distributes the service rate between backlogged classes
  dynamically based on the load of each class
• SDPs function as weights in the service rate distribution
• In heavy load conditions,
         i   sj
         j   si
   (more actually than the BPR scheduler)
• Approximates the short-timescale proportional diff. model
  in time intervals of high load
          WTP scheduler (Cont‟)
• Short-term starvation effect can happen for arbitrarily long
  high-class bursts (due to priority nature)
            R si
       1      ( si  s j ) if   R  RI
            RI s j
       RI : peak input rate in the scheduler
       R : output link (or service) rate
 a sequence of  consecutive class j packets that starts arriving
    at time t0 will be serviced before any class i packets that
    arrived at t0 or later, and this is true for arbitrary large
    values of  .
     Evaluation of BPR and WTP

•   The effect of the aggregate load
•   The effect of the class load distribution
•   The behavior of BPR and WTP in short timescales
•   Microscopic views of the behavior of BPR and
                 Simulated Model
• N(=4) packet sources (one source for each service class)
• SDPs (s1=1, s2=2, s3=4, s4=8)
• Packet interarrivals in the same class: Pareto distribution
• Class load distribution (class1: 40%, class2: 30%, class3:
  20%, class4: 10%)
• Packet length distribution (40B: 40%, 550B: 50%, 1500B:
• Avg. packet transmission time (p-unit): 11.2 time units
• utilization factor  : moderate-load (0.7), heavy-load
         Effect of Aggregate Load

• As the aggregate load increases, the WTP scheduler tends
  to proportional average-delay diff.
                i   sj
                j   si

• Both scheduler tends to the proportional delay diff. only
  under sufficiently heavy-load (not in moderate loads)
  Effect of Class Load Distribution

• The WTP scheduler provides the specified avg.-delay diff.
  ratio independent of the load distribution in a very precise
• The BPR can achieve the proportional avg.-delay model
  when all classes have the same load (not when some
  classes are more loaded than others: the highly loaded
  classes get higher delays than what the SDPs specify).
   Behavior in Short Timescales

• WTP: approximates the proportional constraints even with
  a monitored timescale of only tens of p-units.

• BPR: a quite „spread‟ range of avg.-delay ratios in
  timescales of hundreds of p-units or less.
 Microscopic Views of the behavior

• BPR: sawtooth-type of variations in the queueing delays.
  - the queueing delays of consecutive packets gradually
  increase, until they suddenly drop at a certain time, after
  the arrival of new packets in that class

• WTP: approximates more precisely the proportional delay
  diff. model
        The User‟s Perspective
• Issue: Can a local and class-based relative
  differentiation lead to consistent end-to-end and
  flow based relative differentiation, independent of
  the network path and user-flow characteristics?

• Simulation result is that the local and class-based
  diff. translates to consistent end-to-end flow-based
                   Open Issues
• Is the proportional delay diff. model the most appropriate
  means for predictable and controllable differentiation?
• The proportional diff. model has to be extended in the
  direction of coupled delay and loss diff., since both
  performance measures are significant in most applications
  and transport protocols.
• It is important to combine the feasibility conditions with
  experimental procedures and measurements, in order to be
  able to determine efficiently the space of feasible DDPs for
  a certain network link.

To top