RFC 4128

Network Working Group                                             W. Lai
Request for Comments: 4128                                     AT&T Labs
Category: Informational                                        June 2005

                   Bandwidth Constraints Models for
  Differentiated Services (Diffserv)-aware MPLS Traffic Engineering:
                        Performance Evaluation

Status of This Memo

   This memo provides information for the Internet community.  It does
   not specify an Internet standard of any kind.  Distribution of this
   memo is unlimited.

Copyright Notice

   Copyright (C) The Internet Society (2005).


   The content of this RFC has been considered by the IETF (specifically
   in the TE-WG working group, which has no problem with publication as
   an Informational RFC), and therefore it may resemble a current IETF
   work in progress or a published IETF work.  However, this document is
   an individual submission and not a candidate for any level of
   Internet Standard.  The IETF disclaims any knowledge of the fitness
   of this RFC for any purpose, and in particular notes that it has not
   had complete IETF review for such things as security, congestion
   control or inappropriate interaction with deployed protocols.  The
   RFC Editor has chosen to publish this document at its discretion.
   Readers of this RFC should exercise caution in evaluating its value
   for implementation and deployment.  See RFC 3932 for more


   "Differentiated Services (Diffserv)-aware MPLS Traffic Engineering
   Requirements", RFC 3564, specifies the requirements and selection
   criteria for Bandwidth Constraints Models.  Two such models, the
   Maximum Allocation and the Russian Dolls, are described therein.
   This document complements RFC 3564 by presenting the results of a
   performance evaluation of these two models under various operational
   conditions: normal load, overload, preemption fully or partially
   enabled, pure blocking, or complete sharing.

Lai                         Standards Track                     [Page 1]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

Table of Contents

   1. Introduction ....................................................3
      1.1. Conventions used in this document ..........................4
   2. Bandwidth Constraints Models ....................................4
   3. Performance Model ...............................................5
      3.1. LSP Blocking and Preemption ................................6
      3.2. Example Link Traffic Model .................................8
      3.3. Performance under Normal Load ..............................9
   4. Performance under Overload .....................................10
      4.1. Bandwidth Sharing versus Isolation ........................10
      4.2. Improving Class 2 Performance at the Expense of Class 3 ...12
      4.3. Comparing Bandwidth Constraints of Different Models .......13
   5. Performance under Partial Preemption ...........................15
      5.1. Russian Dolls Model .......................................16
      5.2. Maximum Allocation Model ..................................16
   6. Performance under Pure Blocking ................................17
      6.1. Russian Dolls Model .......................................17
      6.2. Maximum Allocation Model ..................................18
   7. Performance under Complete Sharing .............................19
   8. Implications on Performance Criteria ...........................20
   9. Conclusions ....................................................21
   10. Security Considerations .......................................22
   11. Acknowledgements ..............................................22
   12. References ....................................................22
       12.1. Normative References ....................................22
       12.2. Informative References ..................................22

Lai                         Standards Track                     [Page 2]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

1.  Introduction

   Differentiated Services (Diffserv)-aware MPLS Traffic Engineering
   (DS-TE) mechanisms operate on the basis of different Diffserv classes
   of traffic to improve network performance.  Requirements for DS-TE
   and the associated protocol extensions are specified in references
   [1] and [2] respectively.

   To achieve per-class traffic engineering, rather than on an aggregate
   basis across all classes, DS-TE enforces different Bandwidth
   Constraints (BCs) on different classes.  Reference [1] specifies the
   requirements and selection criteria for Bandwidth Constraints Models
   (BCMs) for the purpose of allocating bandwidth to individual classes.

   This document presents a performance analysis for the two BCMs
   described in [1]:

   (1) Maximum Allocation Model (MAM) - the maximum allowable bandwidth
       usage of each class, together with the aggregate usage across all
       classes, are explicitly specified.

   (2) Russian Dolls Model (RDM) - specification of maximum allowable
       usage is done cumulatively by grouping successive priority
       classes recursively.

   The following criteria are also listed in [1] for investigating the
   performance and trade-offs of different operational aspects of BCMs:

   (1) addresses the scenarios in Section 2 of [1]

   (2) works well under both normal and overload conditions

   (3) applies equally when preemption is either enabled or disabled

   (4) minimizes signaling load processing requirements

   (5) maximizes efficient use of the network

   (6) minimizes implementation and deployment complexity

   The use of any given BCM has significant impacts on the capability of
   a network to provide protection for different classes of traffic,
   particularly under high load, so that performance objectives can be
   met [3].  This document complements [1] by presenting the results of
   a performance evaluation of the above two BCMs under various
   operational conditions: normal load, overload, preemption fully or
   partially enabled, pure blocking, or complete sharing.  Thus, our
   focus is only on the performance-oriented criteria and their

Lai                         Standards Track                     [Page 3]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   implications for a network implementation.  In other words, we are
   only concerned with criteria (2), (3), and (5); we will not address
   criteria (1), (4), or (6).

   Related documents in this area include [4], [5], [6], [7], and [8].

   In the rest of this document, the following DS-TE acronyms are used:

      BC    Bandwidth Constraint
      BCM   Bandwidth Constraints Model
      MAM   Maximum Allocation Model
      RDM   Russian Dolls Model

   There may be differences between the quality of service expressed and
   obtained with Diffserv without DS-TE and with DS-TE.  Because DS-TE
   uses Constraint Based Routing, and because of the type of admission
   control capabilities it adds to Diffserv, DS-TE has capabilities for
   traffic that Diffserv does not.  Diffserv does not indicate
   preemption, by intent, whereas DS-TE describes multiple levels of
   preemption for its Class-Types.  Also, Diffserv does not support any
   means of explicitly controlling overbooking, while DS-TE allows this.
   When considering a complete quality of service environment, with
   Diffserv routers and DS-TE, it is important to consider these
   differences carefully.

1.1.  Conventions used in this document

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   document are to be interpreted as described in RFC 2119.

2.  Bandwidth Constraints Models

   To simplify our presentation, we use the informal name "class of
   traffic" for the terms Class-Type and TE-Class, defined in [1].  We
   assume that (1) there are only three classes of traffic, and that (2)
   all label-switched paths (LSPs), regardless of class, require the
   same amount of bandwidth.  Furthermore, the focus is on the bandwidth
   usage of an individual link with a given capacity; routing aspects of
   LSP setup are not considered.

   The concept of reserved bandwidth is also defined in [1] to account
   for the possible use of overbooking.  Rather than get into these
   details, we assume that each LSP is allocated 1 unit of bandwidth on
   a given link after establishment.  This allows us to express link
   bandwidth usage simply in terms of the number of simultaneously
   established LSPs.  Link capacity can then be used as the aggregate
   constraint on bandwidth usage across all classes.

Lai                         Standards Track                     [Page 4]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   Suppose that the three classes of traffic assumed above for the
   purposes of this document are denoted by class 1 (highest priority),
   class 2, and class 3 (lowest priority).  When preemption is enabled,
   these are the preemption priorities.  To define a generic class of
   BCMs for the purpose of our analysis in accordance with the above
   assumptions, let

      Nmax = link capacity; i.e., the maximum number of simultaneously
             established LSPs for all classes together

      Nc = the number of simultaneously established class c LSPs,
           for c = 1, 2, and 3, respectively.

   For MAM, let

      Bc = maximum number of simultaneously established class c LSPs.

   Then, Bc is the Bandwidth Constraint for class c, and we have

      Nc <= Bc <= Nmax, for c = 1, 2, and 3
      N1 + N2 + N3 <= Nmax
      B1 + B2 + B3 >= Nmax

   For RDM, the BCs are specified as:

      B1 = maximum number of simultaneously established class 1 LSPs

      B2 = maximum number of simultaneously established LSPs for classes
           1 and 2 together

      B3 = maximum number of simultaneously established LSPs for classes
           1, 2, and 3 together

   Then, we have the following relationships:

      N1 <= B1
      N1 + N2 <= B2
      N1 + N2 + N3 <= B3
      B1 < B2 < B3 = Nmax

3.  Performance Model

   Reference [8] presents a 3-class Markov-chain performance model to
   analyze a general class of BCMs.  The BCMs that can be analyzed
   include, besides MAM and RDM, BCMs with privately reserved bandwidth
   that cannot be preempted by other classes.

Lai                         Standards Track                     [Page 5]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   The Markov-chain performance model in [8] assumes Poisson arrivals
   for LSP requests with exponentially distributed lifetime.  The
   Poisson assumption for LSP requests is relevant since we are not
   dealing with the arrivals of individual packets within an LSP.  Also,
   LSP lifetime may exhibit heavy-tail characteristics.  This effect
   should be accounted for when the performance of a particular BCM by
   itself is evaluated.  As the effect would be common for all BCMs, we
   ignore it for simplicity in the comparative analysis of the relative
   performance of different BCMs.  In principle, a suitably chosen
   hyperexponential distribution may be used to capture some aspects of
   heavy tail.  However, this will significantly increase the complexity
   of the non-product-form preemption model in [8].

   The model in [8] assumes the use of admission control to allocate
   link bandwidth to LSPs of different classes in accordance with their
   respective BCs.  Thus, the model accepts as input the link capacity
   and offered load from different classes.  The blocking and preemption
   probabilities for different classes under different BCs are generated
   as output.  Thus, from a service provider's perspective, given the
   desired level of blocking and preemption performance, the model can
   be used iteratively to determine the corresponding set of BCs.

   To understand the implications of using criteria (2), (3), and (5) in
   the Introduction Section to select a BCM, we present some numerical
   results of the analysis in [8].  This is intended to facilitate
   discussion of the issues that can arise.  The major performance
   objective is to achieve a balance between the need for bandwidth
   sharing (for increasing bandwidth efficiency) and the need for
   bandwidth isolation (for protecting bandwidth access by different

3.1.  LSP Blocking and Preemption

   As described in Section 2, the three classes of traffic used as an
   example are class 1 (highest priority), class 2, and class 3 (lowest
   priority).  Preemption may or may not be used, and we will examine
   the performance of each scenario.  When preemption is used, the
   priorities are the preemption priorities.  We consider cross-class
   preemption only, with no within-class preemption.  In other words,
   preemption is enabled so that, when necessary, class 1 can preempt
   class 3 or class 2 (in that order), and class 2 can preempt class 3.

   Each class offers a load of traffic to the network that is expressed
   in terms of the arrival rate of its LSP requests and the average
   lifetime of an LSP.  A unit of such a load is an erlang.  (In
   packet-based networks, traffic volume is usually measured by counting
   the number of bytes and/or packets that are sent or received over an
   interface during a measurement period.  Here we are only concerned

Lai                         Standards Track                     [Page 6]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   with bandwidth allocation and usage at the LSP level.  Therefore, as
   a measure of resource utilization in a link-speed independent manner,
   the erlang is an appropriate unit for our purpose [9].)

   To prevent Diffserv QoS degradation at the packet level, the expected
   number of established LSPs for a given class should be kept in line
   with the average service rate that the Diffserv scheduler can provide
   to that class.  Because of the use of overbooking, the actual traffic
   carried by a link may be higher than expected, and hence QoS
   degradation may not be totally avoidable.

   However, the use of admission control at the LSP level helps minimize
   QoS degradation by enforcing the BCs established for the different
   classes, according to the rules of the BCM adopted.  That is, the BCs
   are used to determine the number of LSPs that can be simultaneously
   established for different classes under various operational
   conditions.  By controlling the number of LSPs admitted from
   different classes, this in turn ensures that the amount of traffic
   submitted to the Diffserv scheduler is compatible with the targeted
   packet-level QoS objectives.

   The performance of a BCM can therefore be measured by how well the
   given BCM handles the offered traffic, under normal or overload
   conditions, while maintaining packet-level service objectives.  Thus,
   assuming that the enforcement of Diffserv QoS objectives by admission
   control is a given, the performance of a BCM can be expressed in
   terms of LSP blocking and preemption probabilities.

   Different BCMs have different strengths and weaknesses.  Depending on
   the BCs chosen for a given load, a BCM may perform well in one
   operating region and poorly in another.  Service providers are mainly
   concerned with the utility of a BCM to meet their operational needs.
   Regardless of which BCM is deployed, the foremost consideration is
   that the BCM works well under the engineered load, such as the
   ability to deliver service-level objectives for LSP blocking
   probabilities.  It is also expected that the BCM handles overload
   "reasonably" well.  Thus, for comparison, the common operating point
   we choose for BCMs is that they meet specified performance objectives
   in terms of blocking/preemption under given normal load.  We then
   observe how their performance varies under overload.  More will be
   said about this aspect later in Section 4.2.

Lai                         Standards Track                     [Page 7]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

3.2.  Example Link Traffic Model

   For example, consider a link with a capacity that allows a maximum of
   15 LSPs from different classes to be established simultaneously.  All
   LSPs are assumed to have an average lifetime of 1 time unit.  Suppose
   that this link is being offered a load of

   2.7 erlangs from class 1,
   3.5 erlangs from class 2, and
   3.5 erlangs from class 3.

   We now consider a scenario wherein the blocking/preemption
   performance objectives for the three classes are desired to be
   comparable under normal conditions (other scenarios are covered in
   later sections).  To meet this service requirement under the above
   given load, the BCs are selected as follows:

   For MAM:

   up to 6 simultaneous LSPs for class 1,
   up to 7 simultaneous LSPs for class 2, and
   up to 15 simultaneous LSPs for class 3.

   For RDM:

   up to 6 simultaneous LSPs for class 1 by itself,
   up to 11 simultaneous LSPs for classes 1 and 2 together, and
   up to 15 simultaneous LSPs for all three classes together.

   Note that the driver is service requirement, independent of BCM.  The
   above BCs are not picked arbitrarily; they are chosen to meet
   specific performance objectives in terms of blocking/preemption
   (detailed in the next section).

   An intuitive "explanation" for the above set of BCs may be as
   follows.  Class 1 BC is the same (6) for both models, as class 1 is
   treated the same way under either model with preemption.  However,
   MAM and RDM operate in fundamentally different ways and give
   different treatments to classes with lower preemption priorities.  It
   can be seen from Section 2 that although RDM imposes a strict
   ordering of the different BCs (B1 < B2 < B3) and a hard boundary
   (B3 = Nmax), MAM uses a soft boundary (B1+B2+B3 >= Nmax) with no
   specific ordering.  As will be explained in Section 4.3, this allows
   RDM to have a higher degree of sharing among different classes.  Such
   a higher degree of coupling means that the numerical values of the
   BCs can be relatively smaller than those for MAM, to meet given
   performance requirements under normal load.

Lai                         Standards Track                     [Page 8]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   Thus, in the above example, the RDM BCs of (6, 11, 15) may be thought
   of as roughly corresponding to the MAM BCs of (6, 6+7, 6+7+15).  (The
   intent here is just to point out that the design parameters for the
   two BCMs need to be different, as they operate differently; strictly
   speaking, the numerical correspondence is incorrect.)  Of course,
   both BCMs are bounded by the same aggregate constraint of the link
   capacity (15).

   The BCs chosen in the above example are not intended to be regarded
   as typical values used by any service provider.  They are used here
   mainly for illustrative purposes.  The method we used for analysis
   can easily accommodate another set of parameter values as input.

3.3.  Performance under Normal Load

   In the example above, based on the BCs chosen, the blocking and
   preemption probabilities for LSP setup requests under normal
   conditions for the two BCMs are given in Table 1.  Remember that the
   BCs have been selected for this scenario to address the service
   requirement to offer comparable blocking/preemption objectives for
   the three classes.

   Table 1.  Blocking and preemption probabilities

   BCM     PB1      PB2      PB3      PP2      PP3    PB2+PP2  PB3+PP3

   MAM   0.03692  0.03961  0.02384     0     0.02275  0.03961  0.04659
   RDM   0.03692  0.02296  0.02402  0.01578  0.01611  0.03874  0.04013

   In the above table, the following apply:

   PB1 = blocking probability of class 1
   PB2 = blocking probability of class 2
   PB3 = blocking probability of class 3

   PP2 = preemption probability of class 2
   PP3 = preemption probability of class 3

   PB2+PP2 = combined blocking/preemption probability of class 2
   PB3+PP3 = combined blocking/preemption probability of class 3

   First, we observe that, indeed, the values for (PB1, PB2+PP2,
   PB3+PP3) are very similar one to another.  This confirms that the
   service requirement (of comparable blocking/preemption objectives for
   the three classes) has been met for both BCMs.

Lai                         Standards Track                     [Page 9]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   Then, we observe that the (PB1, PB2+PP2, PB3+PP3) values for MAM are
   very similar to the (PB1, PB2+PP2, PB3+PP3) values for RDM.  This
   indicates that, in this scenario, both BCMs offer very similar
   performance under normal load.

   From column 2 of Table 1, it can be seen that class 1 sees exactly
   the same blocking under both BCMs.  This should be obvious since both
   allocate up to 6 simultaneous LSPs for use by class 1 only.  Slightly
   better results are obtained from RDM, as shown by the last two
   columns in Table 1.  This comes about because the cascaded bandwidth
   separation in RDM effectively gives class 3 some form of protection
   from being preempted by higher-priority classes.

   Also, note that PP2 is zero in this particular case, simply because
   the BCs for MAM happen to have been chosen in such a way that class 1
   never has to preempt class 2 for any of the bandwidth that class 1
   needs.  (This is because class 1 can, in the worst case, get all the
   bandwidth it needs simply by preempting class 3 alone.)  In general,
   this will not be the case.

   It is interesting to compare these results with those for the case of
   a single class.  Based on the Erlang loss formula, a capacity of 15
   servers can support an offered load of 10 erlangs with a blocking
   probability of 0.0364969.  Whereas the total load for the 3-class BCM
   is less with 2.7 + 3.5 + 3.5 = 9.7 erlangs, the probabilities of
   blocking/preemption are higher.  Thus, there is some loss of
   efficiency due to the link bandwidth being partitioned to accommodate
   for different traffic classes, thereby resulting in less sharing.
   This aspect will be examined in more detail later, in Section 7 on
   Complete Sharing.

4.  Performance under Overload

   Overload occurs when the traffic on a system is greater than the
   traffic capacity of the system.  To investigate the performance under
   overload conditions, the load of each class is varied separately.
   Blocking and preemption probabilities are not shown separately for
   each case; they are added together to yield a combined
   blocking/preemption probability.

4.1.  Bandwidth Sharing versus Isolation

   Figures 1 and 2 show the relative performance when the load of each
   class in the example of Section 3.2 is varied separately.  The three
   series of data in each of these figures are, respectively,

Lai                         Standards Track                    [Page 10]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   class 1 blocking probability ("Class 1 B"),
   class 2 blocking/preemption probability ("Class 2 B+P"), and
   class 3 blocking/preemption probability ("Class 3 B+P").

   For each of these series, the first set of four points is for the
   performance when class 1 load is increased from half of its normal
   load to twice its normal.  Similarly, the next and the last sets of
   four points are when class 2 and class 3 loads are increased

   The following observations apply to both BCMs:

   1. The performance of any class generally degrades as its load

   2. The performance of class 1 is not affected by any changes
      (increases or decreases) in either class 2 or class 3 traffic,
      because class 1 can always preempt others.

   3. Similarly, the performance of class 2 is not affected by any
      changes in class 3 traffic.

   4. Class 3 sees better (worse) than normal performance when either
      class 1 or class 2 traffic is below (above) normal.

   In contrast, the impact of the changes in class 1 traffic on class 2
   performance is different for the two BCMs: It is negligible in MAM
   and significant in RDM.

   1. Although class 2 sees little improvement (no improvement in this
      particular example) in performance when class 1 traffic is below
      normal when MAM is used, it sees better than normal performance
      under RDM.

   2. Class 2 sees no degradation in performance when class 1 traffic is
      above normal when MAM is used.  In this example, with BCs 6 + 7 <
      15, class 1 and class 2 traffic is effectively being served by
      separate pools.  Therefore, class 2 sees no preemption, and only
      class 3 is being preempted whenever necessary.  This fact is
      confirmed by the Erlang loss formula: a load of 2.7 erlangs
      offered to 6 servers sees a 0.03692 blocking, and a load of 3.5
      erlangs offered to 7 servers sees a 0.03961 blocking.  These
      blocking probabilities are exactly the same as the corresponding
      entries in Table 1: PB1 and PB2 for MAM.

   3. This is not the case in RDM.  Here, the probability for class 2 to
      be preempted by class 1 is nonzero because of two effects.  (1)
      Through the cascaded bandwidth arrangement, class 3 is protected

Lai                         Standards Track                    [Page 11]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

      somewhat from preemption.  (2) Class 2 traffic is sharing a BC
      with class 1.  Consequently, class 2 suffers when class 1 traffic

   Thus, it appears that although the cascaded bandwidth arrangement and
   the resulting bandwidth sharing makes RDM work better under normal
   conditions, such interaction makes it less effective to provide class
   isolation under overload conditions.

4.2.  Improving Class 2 Performance at the Expense of Class 3

   We now consider a scenario in which the service requirement is to
   give better blocking/preemption performance to class 2 than to class
   3, while maintaining class 1 performance at the same level as in the
   previous scenario.  (The use of minimum deterministic guarantee for
   class 3 is to be considered in the next section.)  So that the
   specified class 2 performance objective can be met, class 2 BC is
   increased appropriately.  As an example, BCs (6, 9, 15) are now used
   for MAM, and (6, 13, 15) for RDM.  For both BCMs, as shown in Figures
   1bis and 2bis, although class 1 performance remains unchanged, class
   2 now receives better performance, at the expense of class 3. This is
   of course due to the increased access of bandwidth by class 2 over
   class 3.  Under normal conditions, the performance of the two BCMs is
   similar in terms of their blocking and preemption probabilities for
   LSP setup requests, as shown in Table 2.

   Table 2.  Blocking and preemption probabilities

   BCM      PB1      PB2      PB3      PP2      PP3    PB2+PP2  PB3+PP3

   MAM    0.03692  0.00658  0.02733     0     0.02709  0.00658  0.05441
   RDM    0.03692  0.00449  0.02759  0.00272  0.02436  0.00721  0.05195

   Under overload, the observations in Section 4.1 regarding the
   difference in the general behavior between the two BCMs still apply,
   as shown in Figures 1bis and 2bis.

   The following are two frequently asked questions about the operation
   of BCMs.

   (1) For a link capacity of 15, would a class 1 BC of 6 and a class 2
       BC of 9 in MAM result in the possibility of a total lockout for
       class 3?

   This will certainly be the case when there are 6 class 1 and 9 class
   2 LSPs being established simultaneously.  Such an offered load (with
   6 class 1 and 9 class 2 LSP requests) will not cause a lockout of
   class 3 with RDM having a BC of 13 for classes 1 and 2 combined, but

Lai                         Standards Track                    [Page 12]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   will result in class 2 LSPs being rejected.  If class 2 traffic were
   considered relatively more important than class 3 traffic, then RDM
   would perform very poorly compared to MAM with BCs of (6, 9, 15).

   (2) Should MAM with BCs of (6, 7, 15) be used instead so as to make
       the performance of RDM look comparable?

   The answer is that the above scenario is not very realistic when the
   offered load is assumed to be (2.7, 3.5, 3.5) for the three classes,
   as stated in Section 3.2.  Treating an overload of (6, 9, x) as a
   normal operating condition is incompatible with the engineering of
   BCs according to needed bandwidth from different classes.  It would
   be rare for a given class to need so much more than its engineered
   bandwidth level.  But if the class did, the expectation based on
   design and normal traffic fluctuations is that this class would
   quickly release unneeded bandwidth toward its engineered level,
   freeing up bandwidth for other classes.

   Service providers engineer their networks based on traffic
   projections to determine network configurations and needed capacity.
   All BCMs should be designed to operate under realistic network
   conditions.  For any BCM to work properly, the selection of values
   for different BCs must therefore be based on the projected bandwidth
   needs of each class, as well as on the bandwidth allocation rules of
   the BCM itself.  This is to ensure that the BCM works as expected
   under the intended design conditions.  In operation, the actual load
   may well turn out to be different from that of the design.  Thus, an
   assessment of the performance of a BCM under overload is essential to
   see how well the BCM can cope with traffic surges or network
   failures.  Reflecting this view, the basis for comparison of two BCMs
   is that they meet the same or similar performance requirements under
   normal conditions, and how they withstand overload.

   In operational practice, load measurement and forecast would be
   useful to calibrate and fine-tune the BCs so that traffic from
   different classes could be redistributed accordingly.  Dynamic
   adjustment of the Diffserv scheduler could also be used to minimize
   QoS degradation.

4.3.  Comparing Bandwidth Constraints of Different Models

   As is pointed out in Section 3.2, the higher degree of sharing among
   the different classes in RDM means that the numerical values of the
   BCs could be relatively smaller than those for MAM. We now examine
   this aspect in more detail by considering the following scenario.  We
   set the BCs so that (1) for both BCMs, the same value is used for
   class 1, (2) the same minimum deterministic guarantee of bandwidth
   for class 3 is offered by both BCMs, and (3) the blocking/preemption

Lai                         Standards Track                    [Page 13]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   probability is minimized for class 2.  We want to emphasize that this
   may not be the way service providers select BCs.  It is done here to
   investigate the statistical behavior of such a deterministic

   For illustration, we use BCs (6, 7, 15) for MAM, and (6, 13, 15) for
   RDM.  In this case, both BCMs have 13 units of bandwidth for classes
   1 and 2 together, and dedicate 2 units of bandwidth for use by class
   3 only.  The performance of the two BCMs under normal conditions is
   shown in Table 3.  It is clear that MAM with (6, 7, 15) gives fairly
   comparable performance objectives across the three classes, whereas
   RDM with (6, 13, 15) strongly favors class 2 at the expense of class
   3.  They therefore cater to different service requirements.

   Table 3.  Blocking and preemption probabilities

   BCM      PB1      PB2      PB3      PP2      PP3    PB2+PP2  PB3+PP3

   MAM    0.03692  0.03961  0.02384     0     0.02275  0.03961  0.04659
   RDM    0.03692  0.00449  0.02759  0.00272  0.02436  0.00721  0.05195

   By comparing Figures 1 and 2bis, it can be seen that, when being
   subjected to the same set of BCs, RDM gives class 2 much better
   performance than MAM, with class 3 being only slightly worse.

   This confirms the observation in Section 3.2 that, when the same
   service requirements under normal conditions are to be met, the
   numerical values of the BCs for RDM can be relatively smaller than
   those for MAM.  This should not be surprising in view of the hard
   boundary (B3 = Nmax) in RDM versus the soft boundary (B1+B2+B3 >=
   Nmax) in MAM.  The strict ordering of BCs (B1 < B2 < B3) gives RDM
   the advantage of a higher degree of sharing among the different
   classes; i.e., the ability to reallocate the unused bandwidth of
   higher-priority classes to lower-priority ones, if needed.
   Consequently, this leads to better performance when an identical set
   of BCs is used as exemplified above.  Such a higher degree of sharing
   may necessitate the use of minimum deterministic bandwidth guarantee
   to offer some protection for lower-priority traffic from preemption.
   The explicit lack of ordering of BCs in MAM and its soft boundary
   imply that the use of minimum deterministic guarantees for lower-
   priority classes may not need to be enforced when there is a lesser
   degree of sharing.  This is demonstrated by the example in Section
   4.2 with BCs (6, 9, 15) for MAM.

   For illustration, Table 4 shows the performance under normal
   conditions of RDM with BCs (6, 15, 15).

Lai                         Standards Track                    [Page 14]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   Table 4.  Blocking and preemption probabilities

   BCM      PB1      PB2      PB3      PP2      PP3    PB2+PP2  PB3+PP3

   RDM    0.03692  0.00060  0.02800  0.00032  0.02740  0.00092  0.05540

   Regardless of whether deterministic guarantees are used, both BCMs
   are bounded by the same aggregate constraint of the link capacity.
   Also, in both BCMs, bandwidth access guarantees are necessarily
   achieved statistically because of traffic fluctuations, as explained
   in Section 4.2.  (As a result, service-level objectives are typically
   specified as monthly averages, under the use of statistical
   guarantees rather than deterministic guarantees.) Thus, given the
   fundamentally different operating principles of the two BCMs
   (ordering, hard versus soft boundary), the dimensions of one BCM
   should not be adopted to design for the other.  Rather, it is the
   service requirements, and perhaps also the operational needs, of a
   service provider that should be used to drive how the BCs of a BCM
   are selected.

5.  Performance under Partial Preemption

   In the previous two sections, preemption is fully enabled in the
   sense that class 1 can preempt class 3 or class 2 (in that order),
   and class 2 can preempt class 3.  That is, both classes 1 and 2 are
   preemptor-enabled, whereas classes 2 and 3 are preemptable.  A class
   that is preemptor-enabled can preempt lower-priority classes
   designated as preemptable.  A class not designated as preemptable
   cannot be preempted by any other classes, regardless of relative

   We now consider the three cases shown in Table 5, in which preemption
   is only partially enabled.

   Table 5.  Partial preemption modes

   preemption modes         preemptor-enabled     preemptable

   "1+2 on 3" (Fig. 3, 6)   class 1, class 2        class 3
   "1 on 3"   (Fig. 4, 7)       class 1             class 3
   "1 on 2+3" (Fig. 5, 8)       class 1         class 3, class 2

   In this section, we evaluate how these preemption modes affect the
   performance of a particular BCM.  Thus, we are comparing how a given
   BCM performs when preemption is fully enabled versus how the same BCM
   performs when preemption is partially enabled.  The performance of
   these preemption modes is shown in Figures 3 to 5 for RDM, and in
   Figures 6 through 8 for MAM, respectively.  In all of these figures,

Lai                         Standards Track                    [Page 15]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   the BCs of Section 3.2 are used for illustration; i.e., (6, 7, 15)
   for MAM and (6, 11, 15) for RDM.  However, the general behavior is
   similar when the BCs are changed to those in Sections 4.2 and 4.3;
   i.e., (6, 9, 15) and (6, 13, 15), respectively.

5.1.  Russian Dolls Model

   Let us first examine the performance under RDM.  There are two sets
   of results, depending on whether class 2 is preemptable: (1) Figures
   3 and 4 for the two modes when only class 3 is preemptable, and (2)
   Figure 2 in the previous section and Figure 5 for the two modes when
   both classes 2 and 3 are preemptable.  By comparing these two sets of
   results, the following impacts can be observed.  Specifically, when
   class 2 is non-preemptable, the behavior of each class is as follows:

   1. Class 1 generally sees a higher blocking probability.  As the
      class 1 space allocated by the class 1 BC is shared with class 2,
      which is now non-preemptable, class 1 cannot reclaim any such
      space occupied by class 2 when needed.  Also, class 1 has less
      opportunity to preempt, as it is able to preempt class 3 only.

   2. Class 3 also sees higher blocking/preemption when its own load is
      increased, as it is being preempted more frequently by class 1,
      when class 1 cannot preempt class 2.  (See the last set of four
      points in the series for class 3 shown in Figures 3 and 4, when
      comparing with Figures 2 and 5.)

   3. Class 2 blocking/preemption is reduced even when its own load is
      increased, since it is not being preempted by class 1.  (See the
      middle set of four points in the series for class 2 shown in
      Figures 3 and 4, when comparing with Figures 2 and 5.)

   Another two sets of results are related to whether class 2 is
   preemptor-enabled.  In this case, when class 2 is not preemptor-
   enabled, class 2 blocking/preemption is increased when class 3 load
   is increased.  (See the last set of four points in the series for
   class 2 shown in Figures 4 and 5, when comparing with Figures 2 and
   3.)  This is because both classes 2 and 3 are now competing
   independently with each other for resources.

5.2.  Maximum Allocation Model

   Turning now to MAM, the significant impact appears to be only on
   class 2, when it cannot preempt class 3, thereby causing its
   blocking/preemption to increase in two situations.

Lai                         Standards Track                    [Page 16]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   1. When class 1 load is increased.  (See the first set of four points
      in the series for class 2 shown in Figures 7 and 8, when comparing
      with Figures 1 and 6.)

   2. When class 3 load is increased.  (See the last set of four points
      in the series for class 2 shown in Figures 7 and 8, when comparing
      with Figures 1 and 6.)  This is similar to RDM; i.e., class 2 and
      class 3 are now competing with each other.

   When Figure 1 (for the case of fully enabled preemption) is compared
   to Figures 6 through 8 (for partially enabled preemption), it can be
   seen that the performance of MAM is relatively insensitive to the
   different preemption modes.  This is because when each class has its
   own bandwidth access limits, the degree of interference among the
   different classes is reduced.

   This is in contrast with RDM, whose behavior is more dependent on the
   preemption mode in use.

6.  Performance under Pure Blocking

   This section covers the case in which preemption is completely
   disabled.  We continue with the numerical example used in the
   previous sections, with the same link capacity and offered load.

6.1.  Russian Dolls Model

   For RDM, we consider two different settings:

   "Russian Dolls (1)" BCs:

   up to 6 simultaneous LSPs for class 1 by itself,
   up to 11 simultaneous LSPs for classes 1 and 2 together, and
   up to 15 simultaneous LSPs for all three classes together.

   "Russian Dolls (2)" BCs:

   up to 9 simultaneous LSPs for class 3 by itself,
   up to 14 simultaneous LSPs for classes 3 and 2 together, and
   up to 15 simultaneous LSPs for all three classes together.

   Note that the "Russian Dolls (1)" set of BCs is the same as
   previously with preemption enabled, whereas the "Russian Dolls (2)"
   has the cascade of bandwidth arranged in reverse order of the

Lai                         Standards Track                    [Page 17]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   As observed in Section 4, the cascaded bandwidth arrangement is
   intended to offer lower-priority traffic some protection from
   preemption by higher-priority traffic.  This is to avoid starvation.
   In a pure blocking environment, such protection is no longer
   necessary.  As depicted in Figure 9, it actually produces the
   opposite, undesirable effect: higher-priority traffic sees higher
   blocking than lower-priority traffic.  With no preemption, higher-
   priority traffic should be protected instead to ensure that it could
   get through when under high load.  Indeed, when the reverse cascade
   is used in "Russian Dolls (2)", the required performance of lower
   blocking for higher-priority traffic is achieved, as shown in Figure
   10.  In this specific example, there is very little difference among
   the performance of the three classes in the first eight data points
   for each of the three series.  However, the BCs can be tuned to get a
   bigger differentiation.

6.2.  Maximum Allocation Model

   For MAM, we also consider two different settings:

   "Exp. Max. Alloc. (1)" BCs:

   up to 7 simultaneous LSPs for class 1,
   up to 8 simultaneous LSPs for class 2, and
   up to 8 simultaneous LSPs for class 3.

   "Exp. Max. Alloc. (2)" BCs:

   up to 7 simultaneous LSPs for class 1, with additional bandwidth for
      1 LSP privately reserved
   up to 8 simultaneous LSPs for class 2, and
   up to 8 simultaneous LSPs for class 3.

   These BCs are chosen so that, under normal conditions, the blocking
   performance is similar to all the previous scenarios.  The only
   difference between these two sets of values is that the "Exp. Max.
   Alloc. (2)" algorithm gives class 1 a private pool of 1 server for
   class protection.  As a result, class 1 has a relatively lower
   blocking especially when its traffic is above normal, as can be seen
   by comparing Figures 11 and 12.  This comes, of course, with a slight
   increase in the blocking of classes 2 and 3 traffic.

   When comparing the "Russian Dolls (2)" in Figure 10 with MAM in
   Figures 11 or 12, the difference between their behavior and the
   associated explanation are again similar to the case when preemption
   is used.  The higher degree of sharing in the cascaded bandwidth
   arrangement of RDM leads to a tighter coupling between the different
   classes of traffic when under overload.  Their performance therefore

Lai                         Standards Track                    [Page 18]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   tends to degrade together when the load of any one class is
   increased.  By imposing explicit maximum bandwidth usage on each
   class individually, better class isolation is achieved.  The trade-
   off is that, generally, blocking performance in MAM is somewhat
   higher than in RDM, because of reduced sharing.

   The difference in the behavior of RDM with or without preemption has
   already been discussed at the beginning of this section.  For MAM,
   some notable differences can also be observed from a comparison of
   Figures 1 and 11.  If preemption is used, higher-priority traffic
   tends to be able to maintain its performance despite the overloading
   of other classes.  This is not so if preemption is not allowed.  The
   trade-off is that, generally, the overloaded class sees a relatively
   higher blocking/preemption when preemption is enabled than there
   would be if preemption is disabled.

7.  Performance under Complete Sharing

   As observed towards the end of Section 3, the partitioning of
   bandwidth capacity for access by different traffic classes tends to
   reduce the maximum link efficiency achievable.  We now consider the
   case where there is no such partitioning, thereby resulting in full
   sharing of the total bandwidth among all the classes.  This is
   referred to as the Complete Sharing Model.

   For MAM, this means that the BCs are such that up to 15 simultaneous
   LSPs are allowed for any class.

   Similarly, for RDM, the BCs are

   up to 15 simultaneous LSPs for class 1 by itself,
   up to 15 simultaneous LSPs for classes 1 and 2 together, and
   up to 15 simultaneous LSPs for all three classes together.

   Effectively, there is now no distinction between MAM and RDM.  Figure
   13 shows the performance when all classes have equal access to link
   bandwidth under Complete Sharing.

   With preemption being fully enabled, class 1 sees virtually no
   blocking, regardless of the loading conditions of the link.  Since
   class 2 can only preempt class 3, class 2 sees some blocking and/or
   preemption when either class 1 load or its own load is above normal;
   otherwise, class 2 is unaffected by increases of class 3 load.  As
   higher priority classes always preempt class 3 when the link is full,
   class 3 suffers the most, with high blocking/preemption when there is
   any load increase from any class.  A comparison of Figures 1, 2, and
   13 shows that, although the performance of both classes 1 and 2 is
   far superior under Complete Sharing, class 3 performance is much

Lai                         Standards Track                    [Page 19]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   better off under either MAM or RDM.  In a sense, class 3 is starved
   under overload as no protection of its traffic is being provided
   under Complete Sharing.

8.  Implications on Performance Criteria

   Based on the previous results, a general theme is shown to be the
   trade-off between bandwidth sharing and class protection/isolation.
   To show this more concretely, let us compare the different BCMs in
   terms of the overall loss probability.  This quantity is defined as
   the long-term proportion of LSP requests from all classes combined
   that are lost as a result of either blocking or preemption, for a
   given level of offered load.

   As noted in the previous sections, although RDM has a higher degree
   of sharing than MAM, both ultimately converge to the Complete Sharing
   Model as the degree of sharing in each of them is increased.  Figure
   14 shows that, for a single link, the overall loss probability is the
   smallest under Complete Sharing and the largest under MAM, with that
   under RDM being intermediate.  Expressed differently, Complete
   Sharing yields the highest link efficiency and MAM the lowest.  As a
   matter of fact, the overall loss probability of Complete Sharing is
   identical to the loss probability of a single class as computed by
   the Erlang loss formula.  Yet Complete Sharing has the poorest class
   protection capability.  (Note that, in a network with many links and
   multiple-link routing paths, analysis in [6] showed that Complete
   Sharing does not necessarily lead to maximum network-wide bandwidth

   Increasing the degree of bandwidth sharing among the different
   traffic classes helps increase link efficiency.  Such increase,
   however, will lead to a tighter coupling between different classes.
   Under normal loading conditions, proper dimensioning of the link so
   that there is adequate capacity for each class can minimize the
   effect of such coupling.  Under overload conditions, when there is a
   scarcity of capacity, such coupling will be unavoidable and can cause
   severe degradation of service to the lower-priority classes.  Thus,
   the objective of maximizing link usage as stated in criterion (5) of
   Section 1 must be exercised with care, with due consideration to the
   effect of interactions among the different classes.  Otherwise, use
   of this criterion alone will lead to the selection of the Complete
   Sharing Model, as shown in Figure 14.

   The intention of criterion (2) in judging the effectiveness of
   different BCMs is to evaluate how they help the network achieve the
   expected performance.  This can be expressed in terms of the blocking
   and/or preemption behavior as seen by different classes under various
   loading conditions.  For example, the relative strength of a BCM can

Lai                         Standards Track                    [Page 20]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   be demonstrated by examining how many times the per-class blocking or
   preemption probability under overload is worse than the corresponding
   probability under normal load.

9.  Conclusions

   BCMs are used in DS-TE for path computation and admission control of
   LSPs by enforcing different BCs for different classes of traffic so
   that Diffserv QoS performance can be maximized.  Therefore, it is of
   interest to measure the performance of a BCM by the LSP
   blocking/preemption probabilities under various operational
   conditions.  Based on this, the performance of RDM and MAM for LSP
   establishment has been analyzed and compared.  In particular, three
   different scenarios have been examined: (1) all three classes have
   comparable performance objectives in terms of LSP blocking/preemption
   under normal conditions, (2) class 2 is given better performance at
   the expense of class 3, and (3) class 3 receives some minimum
   deterministic guarantee.

   A general theme is the trade-off between bandwidth sharing to achieve
   greater efficiency under normal conditions, and to achieve robust
   class protection/isolation under overload.  The general properties of
   the two BCMs are as follows:


   - allows greater sharing of bandwidth among different classes

   - performs somewhat better under normal conditions

   - works well when preemption is fully enabled; under partial
     preemption, not all preemption modes work equally well


   - does not depend on the use of preemption

   - is relatively insensitive to the different preemption modes when
     preemption is used

   - provides more robust class isolation under overload

   Generally, the use of preemption gives higher-priority traffic some
   degree of immunity to the overloading of other classes.  This results
   in a higher blocking/preemption for the overloaded class than that in
   a pure blocking environment.

Lai                         Standards Track                    [Page 21]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

10.  Security Considerations

   This document does not introduce additional security threats beyond
   those described for Diffserv [10] and MPLS Traffic Engineering [11,
   12, 13, 14], and the same security measures and procedures described
   in those documents apply here.  For example, the approach for defense
   against theft- and denial-of-service attacks discussed in [10], which
   consists of the combination of traffic conditioning at Diffserv
   boundary nodes along with security and integrity of the network
   infrastructure within a Diffserv domain, may be followed when DS-TE
   is in use.

   Also, as stated in [11], it is specifically important that
   manipulation of administratively configurable parameters (such as
   those related to DS-TE LSPs) be executed in a secure manner by
   authorized entities.  For example, as preemption is an
   administratively configurable parameter, it is critical that its
   values be set properly throughout the network.  Any misconfiguration
   in any label switch may cause new LSP setup requests either to be
   blocked or to unnecessarily preempt LSPs already established.
   Similarly, the preemption values of LSP setup requests must be
   configured properly; otherwise, they may affect the operation of
   existing LSPs.

11.  Acknowledgements

   Inputs from Jerry Ash, Jim Boyle, Anna Charny, Sanjaya Choudhury,
   Dimitry Haskin, Francois Le Faucheur, Vishal Sharma, and Jing Shen
   are much appreciated.

12.  References

12.1.  Normative References

   [1]  Le Faucheur, F. and W. Lai, "Requirements for Support of
        Differentiated Services-aware MPLS Traffic Engineering", RFC
        3564, July 2003.

12.2.  Informative References

   [2]  Le Faucheur, F., Ed., "Protocol Extensions for Support of
        Diffserv-aware MPLS Traffic Engineering", RFC 4124, June 2005.

   [3]  Boyle, J., Gill, V., Hannan, A., Cooper, D., Awduche, D.,
        Christian, B., and W. Lai, "Applicability Statement for Traffic
        Engineering with MPLS", RFC 3346, August 2002.

Lai                         Standards Track                    [Page 22]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

   [4]  Le Faucheur, F. and W. Lai, "Maximum Allocation Bandwidth
        Constraints Model for Diffserv-aware MPLS Traffic Engineering",
        RFC 4125, June 2005.

   [5]  Le Faucheur, F., Ed., "Russian Dolls Bandwidth Constraints Model
        for Diffserv-aware MPLS Traffic Engineering", RFC 4127, June

   [6]  Ash, J., "Max Allocation with Reservation Bandwidth Constraint
        Model for MPLS/DiffServ TE & Performance Comparisons", RFC 4126,
        June 2005.

   [7]  F. Le Faucheur, "Considerations on Bandwidth Constraints Models
        for DS-TE", Work in Progress.

   [8]  W.S. Lai, "Traffic Engineering for MPLS," Internet Performance
        and Control of Network Systems III Conference, SPIE Proceedings
        Vol. 4865, Boston, Massachusetts, USA, 30-31 July 2002, pp.

   [9]  W.S. Lai, "Traffic Measurement for Dimensioning and Control of
        IP Networks," Internet Performance and Control of Network
        Systems II Conference, SPIE Proceedings Vol. 4523, Denver,
        Colorado, USA, 21-22 August 2001, pp. 359-367.

   [10] Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., and W.
        Weiss, "An Architecture for Differentiated Service", RFC 2475,
        December 1998.

   [11] Awduche, D., Malcolm, J., Agogbua, J., O'Dell, M., and J.
        McManus, "Requirements for Traffic Engineering Over MPLS", RFC
        2702, September 1999.

   [12] Awduche, D., Berger, L., Gan, D., Li, T., Srinivasan, V., and G.
        Swallow, "RSVP-TE: Extensions to RSVP for LSP Tunnels", RFC
        3209, December 2001.

   [13] Katz, D., Kompella, K., and D. Yeung, "Traffic Engineering (TE)
        Extensions to OSPF Version 2", RFC 3630, September 2003.

   [14] Smit, H. and T. Li, "Intermediate System to Intermediate System
        (IS-IS) Extensions for Traffic Engineering (TE)", RFC 3784, June

Lai                         Standards Track                    [Page 23]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

Author's Address

   Wai Sum Lai
   AT&T Labs
   Room D5-3D18
   200 Laurel Avenue
   Middletown, NJ 07748

   Phone: +1 732-420-3712
   EMail: wlai@att.com

Lai                         Standards Track                    [Page 24]

RFC 4128          BC Models for Diffserv-aware MPLS TE         June 2005

Full Copyright Statement

   Copyright (C) The Internet Society (2005).

   This document is subject to the rights, licenses and restrictions
   contained in BCP 78 and at www.rfc-editor.org/copyright.html, and
   except as set forth therein, the authors retain all their rights.

   This document and the information contained herein are provided on an

Intellectual Property

   The IETF takes no position regarding the validity or scope of any
   Intellectual Property Rights or other rights that might be claimed to
   pertain to the implementation or use of the technology described in
   this document or the extent to which any license under such rights
   might or might not be available; nor does it represent that it has
   made any independent effort to identify any such rights.  Information
   on the procedures with respect to rights in RFC documents can be
   found in BCP 78 and BCP 79.

   Copies of IPR disclosures made to the IETF Secretariat and any
   assurances of licenses to be made available, or the result of an
   attempt made to obtain a general license or permission for the use of
   such proprietary rights by implementers or users of this
   specification can be obtained from the IETF on-line IPR repository at

   The IETF invites any interested party to bring to its attention any
   copyrights, patents or patent applications, or other proprietary
   rights that may cover technology that may be required to implement
   this standard.  Please address the information to the IETF at ietf-


   Funding for the RFC Editor function is currently provided by the
   Internet Society.

Lai                         Standards Track                    [Page 25]