Traffic Engineering Working Group Wai Sum Lai Internet Draft AT&T Labs Document: September 2003 Category: Informational Bandwidth Constraints Models for Differentiated Services-aware MPLS Traffic Engineering: Performance Evaluation Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC2026. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet- Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. This document is available in both .txt and .pdf formats. Abstract The 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 describing in more details some of the selection criteria and their implications. Results of a performance evaluation of the two models are also included. Conventions used in this document The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC-2119. Table of Contents Lai Category - Expiration [Page 1] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 Status of this Memo................................................1 Abstract...........................................................1 1. Introduction....................................................2 2. Bandwidth Constraints Models....................................3 3. Performance Model...............................................4 3.1 LSP Blocking and Preemption....................................5 3.2 Example Link Traffic Model.....................................6 3.3 Performance Under Normal Load..................................7 4. Performance Under Overload......................................9 4.1 Bandwidth Sharing Versus Isolation.............................9 4.2 Improving Class 2 Performance at the Expense of Class 3.......10 4.3 Comparing Bandwidth Constraints of Different Models...........11 5. Performance Under Partial Preemption...........................13 5.1 Russian Dolls.................................................14 5.2 Maximum Allocation............................................14 6. Performance Under Pure Blocking................................15 6.1 Russian Dolls.................................................15 6.2 Maximum Allocation............................................16 7. Performance Under Complete Sharing.............................17 8. Implications on Selection Criteria.............................17 9. Conclusions....................................................18 10. Security Considerations.......................................19 11. References....................................................19 12. Acknowledgments...............................................20 13. Author's Address..............................................20 Full Copyright Statement..........................................20 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, 2], respectively. To achieve per-class traffic engineering, rather than on an aggregate basis across all classes, DS-TE enforces different bandwidth constraints on different classes. Reference [1] specifies the requirements and selection criteria for bandwidth constraints models for the purpose of allocating bandwidth to individual classes. Two bandwidth constraints models are 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 selection criteria are also listed in [1]: Lai Category - Expiration [Page 2] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 (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 bandwidth constraints model 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]. Therefore, the criteria used to select a model must enable us to evaluate how a particular model delivers its performance, relative to other models. This document complements [1] by describing in more details the performance-oriented selection criteria and their implications in a network implementation. Thus, our focus is only on criteria (2), (3), and (5); we will not address criteria (1), (4), and (6). Also included are the results of a performance evaluation of the above two models under various operational conditions: normal load, overload, preemption fully or partially enabled, pure blocking, or complete sharing. Related documents in this area include [4, 5, 6, 7, 8]. 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 (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 getting 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. Suppose that the three classes of traffic are denoted as 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 bandwidth constraints models for the purpose of our analysis in accordance with the above assumptions, let Lai Category - Expiration [Page 3] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 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 the maximum allocation model, 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 the Russian Dolls model, the bandwidth constraints 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 In [8], a 3-class Markov-chain performance model is presented to analyze a general class of bandwidth constraints models. The models that can be analyzed include, besides the maximum allocation and the Russian Dolls, also models with privately reserved bandwidth that cannot be preempted by other classes. 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 packet within an LSP. Also, LSP lifetime may exhibit heavy-tail characteristics. This effect should be accounted for when the performance of a particular bandwidth constraints model by itself is evaluated. As the effect would be common for all bandwidth constraints models, we ignore it for simplicity in the comparative analysis of the relative performance of different models. 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]. Lai Category - Expiration [Page 4] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 The model in [8] assumes the use of admission control to allocate link bandwidth to LSPs of different classes in accordance with their respective bandwidth constraints. 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 bandwidth constraints 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 bandwidth constraints. To understand the implications of using criteria (2), (3), and (5) in the Introduction Section to select a bandwidth constraints model, we present some numerical results of the analysis in [8]. This is to gain some insight to facilitate the discussion of the issues that can arise. The major performance objective is to achieve a balance between the need for bandwidth sharing so as to gain bandwidth efficiency, and the need for bandwidth isolation so as to protect bandwidth access by different classes. 3.1 LSP Blocking and Preemption As described in Section 2, the three classes of traffic 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 with bandwidth allocation and usage at the LSP level. Hence, the erlang as a measure of resource utilization in a link-speed independent manner 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 to *minimize* QoS degradation by enforcing the bandwidth constraints established for the different classes, according to the rules of the Lai Category - Expiration [Page 5] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 bandwidth constraints model adopted. That is, the bandwidth constraints 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 bandwidth constraints model can therefore be measured by how well the given model handles the offered traffic, under normal or overload conditions, while maintaining packet-level service objectives. Thus, assuming the enforcement of Diffserv QoS objectives by admission control as a given, the performance of a bandwidth constraints model can be expressed in terms of *LSP blocking and preemption probabilities*. Different models have different strengths and weaknesses. Depending on the bandwidth constraint values chosen for a given load, a model may perform well in one operating region and poorly in another region. Service providers are mainly concerned with the utility of a model to meet their operational needs. Regardless of which model is deployed, the foremost consideration is that the model 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 model handles overload "reasonably" well. Thus, for comparison, the common operating point we choose for each model 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. 3.2 Example Link Traffic Model As an 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 whereby 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 bandwidth constraints are selected as follows: For the explicit maximum allocation model: 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 the Russian Dolls model: Lai Category - Expiration [Page 6] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 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 bandwidth constraints model. The above bandwidth constraints are not arbitrarily picked; 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 bandwidth constraint values may be as follows. Class 1 bandwidth constraint is the same (6) for both models, as class 1 is treated the same way under either model with preemption. However, the maximum allocation and the Russian Dolls models operate in fundamentally different ways and give different treatments to classes with lower preemption priorities. In can be seen from Section 2 that while the Russian Dolls model imposes a strict ordering of the different bandwidth constraint values (B1 < B2 < B3) and a hard boundary (B3 = Nmax), the maximum allocation model uses a soft boundary (B1+B2+B3 >= Nmax) with no specific ordering. As to be explained in Section 4.3, this allows the Russian Dolls model to have a higher degree of sharing among different classes. Such a higher degree of coupling means that the numerical values of the bandwidth constraints can be relatively smaller when compared with those for the maximum allocation model, to meet given performance requirements under normal load. Thus, in the above example, the bandwidth constraints of (6, 11, 15) in the Russian Dolls model may be thought of as roughly corresponding to the bandwidth constraints of (6, 6+7, 6+7+15) for the maximum allocation model. (The intent here is just to point out that the design parameters for the two models need to be different as they operate differently - strictly speaking, the numerical correspondence is incorrect.) Of course, both models are bounded by the same aggregate constraint of the link capacity (15). The values 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 bandwidth constraint values chosen, the blocking and preemption probabilities for LSP setup requests under normal conditions for the two models are given in Table 1. Remember that the bandwidth constraint values 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 Model PB1 PB2 PB3 PP2 PP3 PB2+PP2 PB3+PP3 MaxAll 0.03692 0.03961 0.02384 0 0.02275 0.03961 0.04659 RussDoll 0.03692 0.02296 0.02402 0.01578 0.01611 0.03874 0.04013 Lai Category - Expiration [Page 7] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 In the above table, 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 models. Then, we observe that the (PB1, PB2+PP2, PB3+PP3) values for the maximum allocation model are very similar to the (PB1, PB2+PP2, PB3+PP3) values for the Russian Dolls model. This indicates that, in this scenario, both models offer very similar performance under normal load. From column 2 of the above table, it can be seen that class 1 sees exactly the same blocking under both models. This should be obvious since both allocate up to 6 simultaneous LSPs for use by class 1 only. Slightly better results are obtained from the Russian Dolls model, as shown by the last two columns in Table 1. This comes about because the cascaded bandwidth separation in the Russian Dolls design 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 bandwidth constraints for the maximum allocation model 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 that 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 model 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 details later in the section on Complete Sharing. Lai Category - Expiration [Page 8] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 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 for each case are not shown separately: 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, 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 correspondingly increased. The following observations apply to both models: 1. The performance of any class generally degrades as its load increases. 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 models: being negligible in the maximum allocation and significant in the Russian Dolls. 1. While class 2 sees little improvement (no improvement in this particular example) in performance when class 1 traffic is below normal when the explicit maximum allocation algorithm is used, it sees better than normal performance under the Russian Dolls algorithm. 2. Class 2 sees no degradation in performance when class 1 traffic is above normal when the explicit maximum allocation algorithm is used. In this example, with bandwidth constraints 6 + 7 < 15, class 1 and class 2 traffic are effectively being served by separate pools. Therefore, class 2 sees no preemption, and only class 3 is being preempted whenever necessary. This fact is Lai Category - Expiration [Page 9] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 confirmed by the Erlang loss formula: a load of 2.7 erlangs offered to 6 servers sees a 0.03692 blocking, 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 MaxAll. 3. This is not the case in the Russian Dolls algorithm. 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 somewhat from preemption. (2) Class 2 traffic is sharing a bandwidth constraint with class 1. Consequently, class 2 suffers when class 1 traffic increases. Thus, it appears that while the cascaded bandwidth arrangement and the resulting bandwidth sharing makes the Russian Dolls algorithm works 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, the bandwidth constraint for class 2 is appropriately increased. As an example, bandwidth constraints (6, 9, 15) are now used for the maximum allocation, and (6, 13, 15) for the Russian Dolls. For both models, as shown in Figures 1bis and 2bis, while 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 models 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 Model PB1 PB2 PB3 PP2 PP3 PB2+PP2 PB3+PP3 MaxAll 0.03692 0.00658 0.02733 0 0.02709 0.00658 0.05441 RussDoll 0.03692 0.00449 0.02759 0.00272 0.02436 0.00721 0.05195 6 0.00721 0.05195 Under overload, the observations in Section 4.1 regarding the difference in the general behavior between the two models still apply, as shown in Figures 1bis and 2bis. Some frequently asked questions about the operation of bandwidth constraints models are as follows. For a link capacity of 15, would a bandwidth constraint of 6 for class 1 and a bandwidth constraint of 9 for class 2 in the maximum allocation model 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 simultaneously established. Such an offered load (with 6 class 1 Lai Category - Expiration [Page 10] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 and 9 class 2 LSP requests) will not cause a lockout of class 3 with the Russian Dolls model having a bandwidth constraint of 13 for classes 1 and 2 combined, but will result in class 2 LSPs being rejected. If class 2 traffic were considered relatively more important then class 3 traffic, then the Russian Dolls would perform very poorly when compared with the maximum allocation model with bandwidth constraints of (6, 9, 15). Should the maximum allocation model with bandwidth constraints of (6, 7, 15) be used instead so as to make the performance of the Russian Dolls 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 normal operating condition is incompatible with the engineering of bandwidth constraints 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 bandwidth constraints models should be designed to operate under realistic network conditions. For any bandwidth constraints model to work properly, the selection of values for different bandwidth constraints must therefore be based on the projected bandwidth needs of each class, as well as the bandwidth allocation rules of the model itself. This is to ensure that the model works as expected under the intended design conditions. In operation, the actual load may well turn out to be different from the design. Thus, an assessment of the performance of a bandwidth constraints model under overload is essential to see how well the model can cope with traffic surges or network failures. Reflecting this view, the basis for comparison of two bandwidth constraints model 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 bandwidth constraints 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 pointed out in Section 3.2, the higher degree of sharing among the different classes in the Russian Dolls model means that the numerical values of the bandwidth constraints could be relatively smaller, when compared with those for the maximum allocation model. We now examine this aspect in more details by considering the following scenario. We set the bandwidth constraints so that, (1) for both models, the same value is used for class 1, (2) the same Lai Category - Expiration [Page 11] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 minimum *deterministic* guarantee of bandwidth for class 3 is offered by both models, and (3) the blocking/preemption probability is minimized for class 2. We want to emphasize that this may not be the way service providers select bandwidth constraint values. It is done here to investigate the *statistical* behavior of such a deterministic mechanism. For illustration, we use bandwidth constraints (6, 7, 15) for the maximum allocation, and (6, 13, 15) for the Russian Dolls. In this case, both models 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 models under normal conditions is shown in Table 3. It is clear that the maximum allocation model with (6, 7, 15) gives fairly comparable performance objectives across the three classes, while the Russian Dolls model 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 Model PB1 PB2 PB3 PP2 PP3 PB2+PP2 PB3+PP3 MaxAll 0.03692 0.03961 0.02384 0 0.02275 0.03961 0.04659 RussDoll 0.03692 0.00449 0.02759 0.00272 0.02436 0.00721 0.05195 6 0.00721 0.05195 By comparing Figures 1 and 2bis, it can be seen that, when being subjected to the same set of bandwidth constraints, the Russian Dolls model gives class 2 much better performance than the maximum allocation model, 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 bandwidth constraints for the Russian Dolls can be relatively smaller than those for the maximum allocation model. This should not be surprising in view of the hard boundary (B3 = Nmax) in the Russian Dolls versus the soft boundary (B1+B2+B3 >= Nmax) in maximum allocation. The strict ordering of bandwidth constraints (B1 < B2 < B3) gives the Russian Dolls model 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 bandwidth constraints 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 bandwidth constraints in the maximum allocation model together with its soft boundary implies 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 bandwidth constraints (6, 9, 15) for the maximum allocation model. Lai Category - Expiration [Page 12] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 For illustration, Table 4 shows the performance under normal conditions of a Russian Dolls model with bandwidth constraints (6, 15, 15). Table 4. Blocking and preemption probabilities Model PB1 PB2 PB3 PP2 PP3 PB2+PP2 PB3+PP3 RussDoll 0.03692 0.00060 0.02800 0.00032 0.02740 0.00092 0.05540 Regardless of whether deterministic guarantees are used or not, both models are bounded by the same aggregate constraint of the link capacity. Also, in both models, 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 models (ordering, hard versus soft boundary), the dimensions of one model 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 bandwidth constraints of a model 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, while 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 priorities. We now consider the three cases shown in Table 5 when 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 model. Thus, we are comparing how a given model performs when preemption is fully enabled versus how the same model performs when preemption is partially enabled. The performance of these preemption modes is shown in Figures 3 to 5 for the Russian Dolls, and Figures 6 to 8 for the maximum allocation model, respectively. In all of these figures, the bandwidth constraints of Section 3.2 are used for illustration, i.e., (6, 7, 15) for maximum allocation model and (6, 11, 15) for Russian Dolls Lai Category - Expiration [Page 13] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 model. However, the general behavior is similar when the bandwidth constraints 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 Let us first examine the performance under the Russian Dolls model. There are two sets of results, depending on whether class 2 is preemptable or not: (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, and when compared with the case of class 2 being preemptable, then the behavior of each class is: 1. Class 1 generally sees a higher blocking probability when class 2 is non-preemptable. As the class 1 space allocated by the class 1 bandwidth constraint 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 - being 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 or not. In this case, when class 2 is not preemptor-enabled, class 2 blocking/preemption is increased when class 3 load is increased (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 Turning now to the maximum allocation model, 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. 1. When class 1 load is increased (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 (the last set of four points in the series for class 2 shown in Figures 7 and 8, when comparing with Lai Category - Expiration [Page 14] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 Figures 1 and 6). This is similar to the Russian Dolls model, i.e., class 2 and class 3 are now competing with each other. When comparing Figure 1 (for the case of fully enabled preemption) with Figures 6 to 8 (for partially enabled preemption), it can be seen that the performance of the maximum allocation model 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 the Russian Dolls model, whose behavior is more dependent on the preemption mode in use. 6. Performance Under Pure Blocking This section covers the case when 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 For the Russian Dolls model, we consider two different settings: "Russian Dolls (1)" bandwidth constraints: 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)" bandwidth constraints: 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 bandwidth constraints is the same as previously with preemption enabled, while the "Russian Dolls (2)" has the cascade of bandwidth arranged in *reverse* order of the classes. 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 they 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 Lai Category - Expiration [Page 15] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 eight data points for each of the three series. However, the bandwidth constraints can be tuned to get a bigger differentiation. 6.2 Maximum Allocation For the maximum allocation model, we also consider two different settings: "Exp. Max. Alloc. (1)" bandwidth constraints: 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)" bandwidth constraints: 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 bandwidth constraints 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 is of course at the expense of a slight increase in the blocking of classes 2 and 3 traffic. When comparing the "Russian Dolls (2)" in Figure 10 with the explicit maximum allocation algorithm 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 the Russian Dolls algorithm leads to a tighter coupling between the different classes of traffic when under overload. Their performance therefore 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 the explicit maximum allocation algorithm is somewhat higher than the Russian Dolls algorithm, because of reduced sharing. The difference in the behavior of the Russian Dolls algorithm with or without preemption has already been discussed at the beginning of this section. For the explicit maximum allocation algorithm, some notable difference 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 their 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 the case when preemption is disabled. Lai Category - Expiration [Page 16] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 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 complete sharing of the total bandwidth among all the classes. For the explicit maximum allocation model, this means that the constraints are such that up to 15 simultaneous LSPs are allowed for any class. Similarly, for the Russian Dolls model, the constraints 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 the two models. Figure 13 shows the performance when all classes have equal access to link bandwidth under the complete sharing scheme. With preemption being fully enabled, it can be seen that class 1 virtually sees 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, while the performance of both classes 1 and 2 is far superior under complete sharing, class 3 performance is much better off under either the explicit maximum allocation or Russian Dolls models. In a sense, class 3 is starved under overload as no protection of its traffic is being provided under complete sharing. 8. Implications on Selection 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 models 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 from the previous sections, while the Russian Dolls model has a higher degree of sharing then explicit maximum allocation, both converge ultimately to the complete sharing model as the degree of sharing in each of them is increased. Figure 14 shows that, for Lai Category - Expiration [Page 17] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 a single link, the overall loss probability is the smallest under complete sharing and the largest under explicit maximum allocation, with Russian Dolls being intermediate. Expressed differently, complete sharing yields the highest link efficiency and explicit maximum allocation the lowest. As a matter of fact, the overall loss probability of complete sharing is identical to loss probability of a single class as computed by the Erlang loss formula. Yet complete sharing has the poorest class protection capability. (We want to point out 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 efficiency.) Increasing the degree of bandwidth sharing among the different traffic classes helps to 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 selection criterion (5) 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 scheme, as shown in Figure 14. The intention of criterion (2) in judging the effectiveness of different models is to evaluate how they help the network to 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 model can be demonstrated by examining how many times the per-class blocking or preemption probability under overload is worse off than the corresponding probability under normal load. 9. Conclusions Bandwidth constraints models are used in DS-TE for path computation and admission control of LSPs by enforcing different bandwidth constraints for different classes of traffic so that Diffserv QoS performance can be maximized. Therefore, it is of interest to measure the performance of a bandwidth constraints model by the LSP blocking/preemption probabilities under various operational conditions. Based on this, the performance of the Russian Dolls and the maximum allocation models 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. Lai Category - Expiration [Page 18] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 A general theme is shown to be the trade-off between bandwidth sharing to achieve greater efficiency under normal conditions, and robust class protection/isolation under overload. The general properties of the two models are: Russian Dolls model . 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 Maximum allocation model . 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 against the overloading of other classes. This results in a higher blocking/preemption for the overloaded class, when compared with a pure blocking environment. 10. Security Considerations No new security considerations are raised by the bandwidth constraints models presented in this document; they are the same as in the DS-TE Requirements document [1]. 11. References Normative References 1 F. Le Faucheur and W.S. Lai, "Requirements for Support of Differentiated Services-aware MPLS Traffic Engineering," RFC 3564, July 2003. Informative References 2 F. Le Faucheur (Editor), "Protocol extensions for support of Diff-Serv-aware MPLS Traffic Engineering," Internet-Draft, Work in Progress. 3 J. Boyle, V. Gill, A. Hannan, D. Cooper, D. Awduche, B. Christian, and W.S. Lai, "Applicability Statement for Traffic Engineering with MPLS," RFC 3346, July 2002. 4 F. Le Faucheur and W.S. Lai, "Maximum Allocation Bandwidth Constraints Model for Diff-Serv-aware MPLS Traffic Engineering," Internet-Draft, Work in Progress. Lai Category - Expiration [Page 19] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 5 F. Le Faucheur (Editor), "Russian Dolls Bandwidth Constraints Model for Diff-Serv-aware MPLS Traffic Engineering," Internet- Draft, Work in Progress. 6 J. Ash, "Max Allocation with Reservation Bandwidth Constraint Model for MPLS/DiffServ TE & Performance Comparisons," Internet- Draft, Work in Progress. 7 F. Le Faucheur, "Considerations on Bandwidth Constraints Models for DS-TE," Internet-Draft, 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. 256- 267. (URL: http://www.columbia.edu/~ffl5/waisum/bcmodel.pdf) 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. 12. Acknowledgments Inputs from Jerry Ash, Jim Boyle, Anna Charny, Sanjaya Choudhury, Dimitry Haskin, Francois Le Faucheur, Vishal Sharma, and Jing Shen are much appreciated. 13. Author's Address Wai Sum Lai AT&T Labs Room D5-3D18 200 Laurel Avenue Middletown, NJ 07748, USA Phone: +1 732-420-3712 Email: wlai@att.com Full Copyright Statement "Copyright (C) The Internet Society (date). All Rights Reserved. This document and translations of it may be copied and furnished to others, and derivative works that comment on or otherwise explain it or assist in its implementation may be prepared, copied, published and distributed, in whole or in part, without restriction of any kind, provided that the above copyright notice and this paragraph are included on all such copies and derivative works. However, this document itself may not be modified in any way, such as by removing the copyright notice or references to the Internet Society or other Internet organizations, except as needed for the purpose of developing Internet standards in which case the procedures for copyrights defined in the Internet Standards process must be followed, or as required to translate it into languages other than English. Lai Category - Expiration [Page 20] Internet-Draft BC Models for Diffserv-aware MPLS TE Sep 2003 The limited permissions granted above are perpetual and will not be revoked by the Internet Society or its successors or assigns. This document and the information contained herein is provided on an "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Lai Category - Expiration [Page 21]