There is a clear need to define attributes and behavior that together
define how traffic should be conditioned. This document defines a
set of classes and relationships that represent the QoS mechanisms
used to condition traffic; [QPIM] is used to define policies to
control the QoS mechanisms defined in this document.
However, some very basic issues need to be considered when combining
these documents. Considering these issues should help in
constructing a schema for managing the operation and configuration of
network QoS mechanisms through the use of QoS policies.
3.1. Level of Abstraction for Expressing QoS Policies
The first issue requiring consideration is the level of abstraction
at which QoS policies should be expressed. If we consider policies
as a set of rules used to react to events and manipulate attributes
or generate new events, we realize that policy represents a continuum
of specifications that relate business goals and rules to the
conditioning of traffic done by a device or a set of devices. An
example of a business level policy might be: from 1:00 pm PST to 7:00
am EST, sell off 40% of the network capacity on the open market. In
contrast, a device-specific policy might be: if the queue depth grows
at a geometric rate over a specified duration, trigger a potential
link failure event.
A general model for this continuum is shown in Figure 1 below.
| High-Level Business | Not directly related to device
| Policies | operation and configuration details
| Device-Independent | Translate high-level policies to
| Policies | generic device operational and
+---------------------+ configuration information
| Device-Dependent | Translate generic device information
| Policies | to specify how particular devices
+---------------------+ should operate and be configured
Figure 1. The Policy Continuum
High-level business policies are used to express the requirements of
the different applications, and prioritize which applications get
"better" treatment when the network is congested. The goal, then, is
to use policies to relate the operational and configuration needs of
a device directly to the business rules that the network
administrator is trying to implement in the network that the device
Device-independent policies translate business policies into a set of
generalized operational and configuration policies that are
independent of any specific device, but dependent on a particular set
of QoS mechanisms, such as random early detection (RED) dropping or
weighted round robin scheduling. Not only does this enable different
types of devices (routers, switches, hosts, etc.) to be controlled by
QoS policies, it also enables devices made by different vendors that
use the same types of QoS mechanisms to be controlled. This enables
these different devices to each supply the correct relative
conditioning to the same type of traffic.
In contrast, device-dependent policies translate device-independent
policies into ones that are specific for a given device. The reason
that a distinction is made between device-independent and device-
dependent policies is that in a given network, many different devices
having many different capabilities need to be controlled together.
Device-independent policies provide a common layer of abstraction for
managing multiple devices of different capabilities, while device-
dependent policies implement the specific conditioning that is
required. This document provides a common set of abstractions for
representing QoS mechanisms in a device-independent way.
This document is focused on the device-independent representation of
QoS mechanisms. QoS mechanisms are modeled in sufficient detail to
provide a common device-independent representation of QoS policies.
They can also be used to provide a basis for specialization, enabling
each vendor to derive a set of vendor-specific classes that represent
how traffic conditioning is done for that vendor's set of devices.
3.2. Specifying Policy Parameters
Policies are a function of parameters (attributes) and operators
(boolean, arithmetic, relational, etc.). Therefore, both need to be
defined as part of the same policy in order to correctly condition
the traffic. If the parameters of the policy are specified too
narrowly, they will reflect the individual implementations of QoS in
each device. As there is currently little consensus in the industry
on what the correct implementation model for QoS is, most defined
attributes would only be applicable to the unique characteristics of
a few individual devices. Moreover, standardizing all of these
potential implementation alternatives would be a never-ending task as
new implementations continued to appear on the market.
On the other hand, if the parameters of the policy are specified too
broadly, it is impossible to develop meaningful policies. For
example, if we concentrate on the so-called Olympic set of policies,
a business policy like "Bob gets Gold Service," is clearly
meaningless to the large majority of existing devices. This is
because the device has no way of determining who Bob is, or what QoS
mechanisms should be configured in what way to provide Gold service.
Furthermore, Gold service may represent a single service, or it may
identify a set of services that are related to each other. In the
latter case, these services may have different conditioning
This document defines a set of parameters that fit into a canonical
model for modeling the elements in the forwarding path of a device
implementing QoS traffic conditioning. By defining this model in a
device-independent way, the needed parameters can be appropriately
3.3. Specifying Policy Services
Administrators want the flexibility to be able to define traffic
conditioning without having to have a low-level understanding of the
different QoS mechanisms that implement that conditioning.
Furthermore, administrators want the flexibility to group different
services together, describing a higher-level concept such as "Gold
Service". This higher-level service could be viewed as providing the
processing to deliver "Gold" quality of service.
These two goals dictate the need for the following set of
o a flexible way to describe a service
o must be able to group different services that may use different
technologies (e.g., DiffServ and IEEE 802.1Q) together
o must be able to define a set of sub-services that together make up
a higher-level service
o must be able to associate a service and the set of QoS mechanisms
that are used to condition traffic for that service
o must be able to define policies that manage the QoS mechanisms
used to implement a service.
This document addresses this set of problems by defining a set of
classes and associations that can represent abstract concepts like
"Gold Service," and bind each of these abstract services to a
specific set of QoS mechanisms that implement the conditioning that
they require. Furthermore, this document defines the concept of
"sub-services," to enable Gold Service to be defined either as a
single service or as a set of services that together should be
treated as an atomic entity.
Given these abstractions, policies (as defined in [QPIM]) can be
written to control the QoS mechanisms and services defined in this
3.4. Level of Abstraction for Defining QoS Attributes and Classes
This document defines a set of classes and properties to support
policies that configure device QoS mechanisms. This document
concentrates on the representation of services in the datapath that
support both DiffServ (for aggregate traffic conditioning) and
IntServ (for flow-based traffic conditioning). Classes and
properties for modeling IntServ admission control services may be
defined in a future document.
The classes and properties in this document are designed to be used
in conjunction with the QoS policy classes and properties defined in
[QPIM]. For example, to preserve the delay characteristics committed
to an end-user, a network administrator may wish to create policies
that monitor the queue depths in a device, and adjust resource
allocations when delay budgets are at risk (perhaps as a result of a
network topology change). The classes and properties in this
document define the specific services and mechanisms required to
implement those services. The classes and properties defined in
[QPIM] provide the overall structure of the policy that manages and
configures this service.
This combination of low-level specification (using this document) and
high-level structuring (using [QPIM]) of network services enables
network administrators to define new services required of the
network, that are directly related to business goals, while ensuring
that such services can be managed. However, this goal (of creating
and managing service-oriented policies) can only be realized if
policies can be constructed that are capable of supporting diverse
implementations of QoS. The solution is to model the QoS
capabilities of devices at the behavioral level. This means that for
traffic conditioning services realized in the datapath, the model
must support the following characteristics:
o modeling of a generic network service that has QoS capabilities
o modeling of how the traffic conditioning itself is defined
o modeling of how statistics are gathered to monitor QoS traffic
conditioning services - this facet of the model will be added in a
This document models a network service, and associates it with one or
more QoS mechanisms that are used to implement that service. It also
models in a canonical form the various components that are used to
condition traffic, such that standard as well as custom traffic
conditioning services may be described.
3.5. Characterization of QoS Properties
The QoS properties and classes will be described in more detail in
Section 4. However, we should consider the basic characteristics of
these properties, to understand the methodology for representing
There are essentially two types of properties, state and
configuration. Configuration properties describe the desired state
of a device, and include properties and classes for representing
desired or proposed thresholds, bandwidth allocations, and how to
classify traffic. State properties describe the actual state of the
device. These include properties to represent the current
operational values of the attributes in devices configured via the
configuration properties, as well as properties that represent state
(queue depths, excess capacity consumption, loss rates, and so
In order to be correlated and used together, these two types of
properties must be modeled using a common information model. The
possibility of modeling state properties and their corresponding
configuration settings is accomplished using the same classes in this
model - although individual instances of the classes would have to be
appropriately named or placed in different containers to distinguish
current state values from desired configuration settings.
State information is addressed in a very limited fashion by QDDIM.
Currently, only CurrentQueueDepth is proposed as an attribute on
QueuingService. The majority of the model is related to
configuration. Given this fact, it is assumed that this model is a
direct memory map into a device. All manipulation of model classes
and properties directly affects the state of the device. If it is
desired to also use these classes to represent desired configuration,
that is left to the discretion of the implementor.
It is acknowledged that additional properties are needed to
completely model current state. However, many of the properties
defined in this document represent exactly the state variables that
will be configured by the configuration properties. Thus, the
definition of the configuration properties has an exact
correspondence with the state properties, and can be used in modeling
both actual (state) and desired/proposed configuration.
3.6. QoS Information Model Derivation
The question of context also leads to another question: how does the
information specified in the core and QoS policy models ([PCIM],
[PCIME], and [QPIM], respectively) integrate with the information
defined in this document? To put it another way, where should
device-independent concepts that lead to device-specific QoS
attributes be derived from?
Past thinking was that QoS was part of the policy model. This view
is not completely accurate, and it leads to confusion. QoS is a set
of services that can be controlled using policy. These services are
represented as device mechanisms. An important point here is that
QoS services, as well as other types of services (e.g., security),
are provided by the mechanisms inherent in a given device. This
means that not all devices are indeed created equal. For example,
although two devices may have the same type of mechanism (e.g., a
queue), one may be a simple implementation (i.e., a FIFO queue)
whereas one may be much more complex and robust (e.g., class-based
weighted fair queuing (CBWFQ)). However, both of these devices can
be used to deliver QoS services, and both need to be controlled by
policy. Thus, a device-independent policy can instruct the devices
to queue certain traffic, and a device-specific policy can be used to
control the queuing in each device.
Furthermore, policy is used to control these mechanisms, not to
represent them. For example, QoS services are implemented with
classifiers, meters, markers, droppers, queues, and schedulers.
Similarly, security is also a characteristic of devices, as
authentication and encryption capabilities represent services that
networked devices perform (irrespective of interactions with policy
servers). These security services may use some of the same
mechanisms that are used by QoS services, such as the concepts of
filters. However, they will mostly require different mechanisms than
the ones used by QoS, even though both sets of services are
implemented in the same devices.
Thus, the similarity between the QoS model and models for other
services is not so much that they contain a few common mechanisms.
Rather, they model how a device implements their respective services.
As such, the modeling of QoS should be part of a networking device
schema rather than a policy schema. This allows the networking
device schema to concentrate on modeling device mechanisms, and the
policy schema to focus on the semantics of representing the policy
itself (conditions, actions, operators, etc.). While this document
concentrates on defining an information model to represent QoS
services in a device datapath, the ultimate goal is to be able to
apply policies that control these services in network devices.
Furthermore, these two schemata (device and policy) must be tightly
integrated in order to enable policy to control QoS services.
3.7. Attribute Representation
The last issue to be considered is the question of how attributes are
represented. If QoS attributes are represented as absolute numbers
(e.g., Class AF2 gets 2 Mbs of bandwidth), it is more difficult to
make them uniform across multiple ports in a device or across
multiple devices, because of the broad variation in link capacities.
However, expressing attributes in relative or proportional terms
(e.g., Class AF2 gets 5% of the total link bandwidth) makes it more
difficult to express certain types of conditions and actions, such
(If ConsumedBandwidth = AssignedBandwidth Then ...)
There are really three approaches to addressing this problem:
o Multiple properties can be defined to express the same value in
various forms. This idea has been rejected because of the
difficulty in keeping these different properties synchronized
(e.g., when one property changes, the others all have to be
o Multi-modal properties can be defined to express the same value,
in different terms, based on the access or assignment mode. This
option was rejected because it significantly complicates the model
and is impossible to express in current directory access protocols
o Properties can be expressed as "absolutes", but the operators in
the policy schema would need to be more sophisticated. Thus, to
represent a percentage, division and multiplication operators are
required (e.g., Class AF2 gets .05 * the total link bandwidth).
This is the approach that has been taken in this document.
3.8. Mental Model
The mental model for constructing this schema is based on the work
done in the Differentiated Services working group. This schema is
based on information provided in the current versions of the DiffServ
Informal Management Model [DSMODEL], the DiffServ MIB [DSMIB], the
PIB [PIB], as well as on information in the set of RFCs that
constitute the basic definition of DiffServ itself ([R2475], [R2474],
[R2597], and [R3246]). In addition, a common set of terminology is
available in [POLTERM].
This model is built around two fundamental class hierarchies that are
bound together using a set of associations. The two class
hierarchies derive from the QoSService and ConditioningService base
classes. A set of associations relate lower-level QoSService
subclasses to higher-level QoS services, relate different types of
conditioning services together in processing a traffic class, and
relate a set of conditioning services to a specific QoS service.
This combination of associations enables us to view the device as
providing a set of services that can be configured, in a modular
building block fashion, to construct application-specific services.
Thus, this document can be used to model existing and future standard
as well as application-specific network QoS services.
3.8.1. The QoSService Class
The first of the classes defined here, QoSService, is used to
represent higher-level network services that require special
conditioning of their traffic. An instance of QoSService (or one of
its subclasses) is used to bring together a group of conditioning
services that, from the perspective of the system manager, are all
used to deliver a common service. Thus, the set of classifiers,
markers, and related conditioning services that provide premium
service to the "selected" set of user traffic may be grouped together
into a premium QoS service.
QoSService has a set of subclasses that represent different
approaches to delivering IP services. The currently defined set of
subclasses are a FlowService for flow-oriented QoS delivery and a
DiffServService for DiffServ aggregate-oriented QoS service delivery.
The QoS services can be related to each other as peers, or they can
be implemented as subservient services to each other. The
QoSSubService aggregation indicates that one or more QoSService
objects are subservient to a particular QoSService object. For
example, this enables us to define Gold Service as a combination of
two DiffServ services, one for high quality traffic treatment, and
one for servicing the rest of the traffic. Each of these
DiffServService objects would be associated with a set of
classifiers, markers, etc, such that the high quality traffic would
get EF marking and appropriate queuing.
The DiffServService class itself has an AFService subclass. This
subclass is used to represent the specific notion that several
related markings within the AF PHB Group work together to provide a
single service. When other DiffServ PHB Groups are defined that use
more than one code point, these will be likely candidates for
additional DiffServService subclasses.
Technology-specific mappings of these services, representing the
specific use of PHB marking or 802.1Q marking, are captured within
the ConditioningService hierarchy, rather than in the subclasses of
These concepts are depicted in Figure 2. Note that both of the
associations are aggregations: a QoSService object aggregates both
the set of QoSService objects subservient to it, and the set of
ConditioningService objects that realize it. See Section 4 for class
and association definitions.
0..1 \/ |
+--------------+ | QoSSubService +---------------+
| |0..n | | |
| QoSService |----- | Conditioning |
| | | Service |
| | | |
| |0..n 0..n| |
| | /\______________________| |
| | \/ QoSConditioning | |
+--------------+ SubService +---------------+
Figure 2. QoSService and its Aggregations3.8.2. The ConditioningService Class
The goal of the ConditioningService classes is to describe the
sequence of traffic conditioning that is applied to a given traffic
stream on the ingress interface through which it enters a device, and
then on the egress interface through which it leaves the device.
This is done using a set of classes and relationships. The routing
decision in the device core, which selects which egress interface a
particular packet will use, is not represented in this model.
A single base class, ConditioningService, is the superclass for a set
of subclasses representing the mechanisms that condition traffic.
These subclasses define device-independent conditioning primitives
(including classifiers, meters, markers, droppers, queues, and
schedulers) that together implement the conditioning of traffic on an
interface. This model abstracts these services into a common set of
modular building blocks that can be used, regardless of device
implementation, to model the traffic conditioning internal to a
The different conditioning mechanisms need to be related to each
other to describe how traffic is conditioned. Several important
variations of how these services are related together exist:
o A particular ingress or egress interface may not require all the
types of ConditioningServices.
o Multiple instances of the same mechanism may be required on an
ingress or egress interface.
o There is no set order of application for the ConditioningServices
on an ingress or egress interface.
Therefore, this model does not dictate a fixed ordering among the
subclasses of ConditioningService, or identify a subclass of
ConditioningService that must appear first or last among the
ConditioningServices on an ingress or egress interface. Instead,
this model ties together the various ConditioningService instances on
an ingress or egress interface using the NextService,
NextServiceAfterMeter, and NextServiceAfterConditioningElement
associations. There are also separate associations, called
EgressConditioningServiceOnEndpoint, which, respectively, tie an
ingress interface to its first ConditioningService, and tie an egress
interface to its last ConditioningService(s).
3.8.3. Preserving QoS Information from Ingress to Egress
There is one important way in which the QDDIM model diverges from the
[DSMODEL]. In [DSMODEL], traffic passes through a network device in
o It comes in on an ingress interface, where it may receive QoS
o It traverses the routing core, where logic outside the scope of
QoS determines which egress interface it will use to leave the
o It may receive further QoS conditioning on the selected egress
interface, and then it leaves the device.
In this model, no information about the QoS conditioning that a
packet receives on the ingress interface is communicated with the
packet across the routing core to the egress interface.
The QDDIM model relaxes this restriction, to allow information about
the treatment that a packet received on an ingress interface to be
communicated along with the packet to the egress interface. (This
relaxation adds a capability that is present in many network
devices.) QDDIM represents this information transfer in terms of a
packet preamble, which is how many devices implement it. But
implementations are free to use other mechanisms to achieve the same
| Meter-A |
a | | b d
| | c e
Figure 3: Meter Followed by Two Preamble MarkersFigure 3 shows an example in which meter results are captured in a
packet preamble. The arrows labeled with single letters represent
instances of either the NextService association (a, d, and e), or of
its peer association NextServiceAfterMeter (b and c). PreambleMarker
PM-1 adds to the packet preamble an indication that the packet exited
Meter A as conforming traffic. Similarly, PreambleMarker PM-2 adds to
the preambles of packets that come through it indications that they
exited Meter A as nonconforming traffic. A PreambleMarker appends
its information to whatever is already present in a packet preamble,
as opposed to overwriting what is already there.
To foster interoperability, the basic format of the information
captured by a PreambleMarker is specified. (Implementations, of
course, are free to represent this information in a different way
internally - this is just how it is represented in the model.) The
information is represented by an ordered, multi-valued string
property FilterItemList, where each individual value of the property
is of the form "<type>,<value>". When a PreambleMarker "appends" its
information to the information that was already present in a packet
preamble, it does so by adding additional items of the indicated
format to the end of the list.
QDDIM provides a limited set of <type>'s that a PreambleMarker may
o ConformingFromMeter: the value is the name of the meter.
o PartConformingFromMeter: the value is the name of the meter.
o NonConformingFromMeter: the value is the name of the meter.
o VlanId: the value is the virtual LAN identifier (VLAN ID).
Implementations may recognize other <type>'s in addition to these.
If collisions of implementation-specific <type>'s become a problem,
it is possible that <type>'s may become an IANA-administered range in
a future revision of this document.
To make use of the information that a PreambleMarker stores in a
packet preamble, a specific subclass PreambleFilter of
FilterEntryBase is defined, to match on the "<type>,<value>" strings.
To simplify the case where there's just a single level of metering in
a device, but different individual meters on each ingress interface,
PreambleFilter allows a wildcard "any" for the <value> part of the
three meter-related filters. With this wildcard, an administrator
can specify a Classifier to select all packets that were found to be
conforming (or partially conforming, or non-conforming) by their
respective meters, without having to name each meter individually in
a separate ClassifierElement.
Once a meter result has been stored in a packet preamble, it is
available for any subsequent Classifier to use. So while the
motivation for this capability has been described in terms of
preserving QoS conditioning information from an ingress interface to
an egress interface, a prior meter result may also be used for
classifying packets later in the datapath on the same interface where
the meter resides.
3.9. Classifiers, FilterLists, and Filter Entries
This document uses a number of classes to model the classifiers
defined in [DSMODEL]: ClassifierService, ClassifierElement,
FilterList, FilterEntryBase, and various subclasses of
FilterEntryBase. There are also two associations involved:
ClassifierElementUsesFilterList and EntriesInFilterList. The QDDIM
model makes no use of CIM's FilterEntry class.
In [DSMODEL], a single traffic stream coming into a classifier is
split into multiple traffic streams leaving it, based on which of an
ordered set of filters each packet in the incoming stream matches. A
filter matches either a field in the packet itself, or possibly other
attributes associated with the packet. In the case of a multi-field
(MF) classifier, packets are assigned to output streams based on the
contents of multiple fields in the packet header. For example, an MF
classifier might assign packets to an output stream based on their
complete IP-addressing 5-tuple.
To optimize the representation of MF classifiers, subclasses of
FilterEntryBase are introduced, which allow multiple related packet
header fields to be represented in a single object. These subclasses
are IPHeaderFilter and 8021Filter. With IPHeaderFilter, for example,
criteria for selecting packets based on all five of the IP 5-tuple
header fields and the DiffServ DSCP can be represented by a
FilterList containing one IPHeaderFilter object. Because these two
classes have applications beyond those considered in this document,
they, as well as the abstract class FilterEntryBase, are defined in
the more general document [PCIME] rather than here.
The FilterList object is always needed, even if it contains only one
filter entry (that is, one FilterEntryBase subclass) object. This is
because a ClassifierElement can only be associated with a Filter
List, as opposed to an individual FilterEntry. FilterList is also
defined in [PCIME].
The EntriesInFilterList aggregation (also defined in [PCIME]) has a
property EntrySequence, which in the past (in CIM) could be used to
specify an evaluation order on the filter entries in a FilterList.
Now, however, the EntrySequence property supports only a single
value: '0'. This value indicates that the FilterEntries are ANDed
together to determine whether a packet matches the MF selector that
the FilterList represents.
A ClassifierElement specifies the starting point for a specific
policy or data path. Each ClassifierElement uses the
NextServiceAfterClassifierElement association to determine the next
conditioning service to apply for packets to.
A ClassifierService defines a grouping of ClassifierElements. There
are certain instances where a ClassifierService actually specifies an
aggregation of ClassifierServices. One practical case would be where
each ClassifierService specifies a group of policies associated with
a particular application and another ClassifierService groups the
application-specific ClassifierService instances. In this particular
case, the application-specific ClassifierService instances are
specified once, but unique combinations of these ClassifierServices
are specified, as needed, using other ClassifierService instances.
ClassifierService instances grouping other ClassifierService
instances may not specify a FilterList using the
ClassifierElementUsesFilterList association. This special use of
ClassifierService serves just as a Classifier collecting function.
3.10. Modeling of Droppers
In [DSMODEL], a distinction is made between absolute droppers and
algorithmic droppers. In QDDIM, both of these types of droppers are
modeled with the DropperService class, or with one of its subclasses.
In both cases, the queue from which the dropper drops packets is tied
to the dropper by an instance of the NextService association. The
dropper always plays the PrecedingService role in these associations,
and the queue always plays the FollowingService role. There is
always exactly one queue from which a dropper drops packets.
Since an absolute dropper drops all packets in its queue, it needs no
configuration beyond a NextService tie to that queue. For an
algorithmic dropper, however, further configuration is needed:
o a specific drop algorithm;
o parameters for the algorithm (for example, token bucket size);
o the source(s) of input(s) to the algorithm;
o possibly per-input parameters for the algorithm.
The first two of these items are represented by properties of the
DropperService class, or properties of one of its subclasses. The
last two, however, involve additional classes and associations.
3.10.1. Configuring Head and Tail Droppers
The HeadTailDropQueueBinding is the association that identifies the
inputs for the algorithm executed by a tail dropper. This
association is not used for a head dropper, because a head dropper
always has exactly one input to its drop algorithm, and this input is
always the queue from which it drops packets. For a tail dropper,
this association is defined to have a many-to-many cardinality.
There are, however, two distinct cases:
One dropper bound to many queues: This represents the case where the
drop algorithm for the dropper involves inputs from more than one
queue. The dropper still drops from only one queue, the one to which
it is tied by a NextService association. But the drop decision may
be influenced by the state of several queues. For the classes
HeadTailDropper and HeadTailDropQueueBinding, the rule for combining
the multiple inputs is simple addition: if the sum of the lengths of
the monitored queues exceeds the dropper's QueueThreshold value, then
packets are dropped. This rule for combining inputs may, however, be
overridden by a different rule in subclasses of one or both of these
One queue bound to many droppers: This represents the case where the
state of one queue (which is typically also the queue from which
packets are dropped) provides an input to multiple droppers' drop
algorithms. A use case here is a classifier that splits a traffic
stream into, say, four parts, representing four classes of traffic.
Each of the parts goes through a separate HeadTailDropper, then
they're re-merged onto the same queue. The net is a single queue
containing packets of four traffic types, with, say, the following
o Class 1 - 90% full
o Class 2 - 80% full
o Class 3 - 70% full
o Class 4 - 50% full
Here the percentages represent the overall state of the queue. With
this configuration, when the queue in question becomes 50% full,
Class 4 packets will be dropped rather than joining the queue, when
it becomes 70% full, Class 3 and 4 packets will be dropped, etc.
The two cases described here can also occur together, if a dropper
receives inputs from multiple queues, one or more of which are also
providing inputs to other droppers.
3.10.2. Configuring RED Droppers
Like a tail dropper, a RED dropper, represented by an instance of the
REDDropperService class, may take as its inputs the states of
multiple queues. In this case, however, there is an additional step:
each of these inputs may be smoothed before the RED dropper uses it,
and the smoothing process itself must be parameterized. Consequently,
in addition to REDDropperService and QueuingService, a third class,
DropThresholdCalculationService, is introduced, to represent the
per-queue parameterization of this smoothing process.
The following instance diagram illustrates how these classes work
with each other:
| | |
+-----+ | +-----+
| | |
DTCS-1 DTCS-2 DTCS-3
| | |
Q-1 Q-2 Q-3
Figure 4. Inputs for a RED Dropper
So REDDropperService-A (RDSvc-A) is using inputs from three queues to
make its drop decision. (As always, RDSvc-A is linked to the queue
from which it drops packets via the NextService association.) For
each of these three queues, there is a
(DropThresholdCalculationService) DTCS instance that represents the
smoothing weight and time interval to use when looking at that queue.
Thus each DTCS instance is tied to exactly one queue, although a
single queue may be examined (with different weight and time values)
by multiple DTCS instances. Also, a DTCS instance and the queue
behind it can be thought of as a "unit of reusability". So a single
DTCS can be referred to by multiple RDSvc's.
Unless it is overridden by a different rule in a subclass of
REDDropperService, the rule that a RED dropper uses to combine the
smoothed inputs from the DTCS's to create a value to use in making
its drop decision is simple addition.
3.11. Modeling of Queues and Schedulers
In order to appreciate the rationale behind this rather complex model
for scheduling, we must consider the rather complex nature of
schedulers, as well as the extreme variations in algorithms and
implementations. Although these variations are broad, we have
identified four examples that serve to test the model and justify its
3.11.1. Simple Hierarchical Scheduler
A simple, hierarchical scheduler has the following properties. First,
when a scheduling opportunity is given to a set of queues, a single,
viable queue is determined based on some scheduling criteria, such as
bandwidth or priority. The output of the scheduler is the input to
another scheduler that treats the first scheduler (and its queues) as
a single logical queue. Hence, if the first scheduler determined the
appropriate packet to release based on a priority assigned to each
Figure 5 illustrates the example and how it would be instantiated
using the model. In the figure, NextService determines the first
scheduler after the queue. NextScheduler determines the
subsequent ordering of schedulers. In addition, the
ElementSchedulingService association determines the set of
scheduling parameters used by a specific scheduler. Scheduling
parameters can be bound either to queues or to schedulers. In
the case of the SchedulingElement EF1-Pri, the binding is to a
queue, so the QueueToSchedule association is used. In the case
of the SchedulingElement PriSched1-Band, the binding is to
another scheduler, so the SchedulerToSchedule association is
used. Note that due to space constraints of the document, the
SchedulingService PRISched1 is represented twice, to show how it
is connected to all the other objects.
3.11.2. Complex Hierarchical Scheduler
A complex, hierarchical scheduler has the same characteristics as
a simple scheduler, except that the criteria for the second
scheduler are determined on a per queue basis rather than on an
aggregate basis. One scenario might be a set of bounded priority
schedulers. In this case, each queue is assigned a relative
priority. However, each queue is also not allowed to exceed a
bandwidth allocation that is unique to that queue. In order to
support this scenario, the queue must be bound to two separate
schedulers. Figure 6 illustrates this situation, by describing
an EF queue and a best effort (BE) queue both pointing to a
priority scheduler via the NextService association. The
NextScheduler association between the priority scheduler and the
bandwidth scheduler in turn defines the ordering of the
scheduling hierarchy. Also note that each scheduler has a
distinct set of scheduling parameters that are bound back to each
queue. This demonstrates the need to support two or more
parameter sets on a per queue basis.
3.11.3. Excess Capacity Scheduler
An excess capacity scheduler offers a similar requirement to support
two scheduling parameter sets per queue. However, in this scenario
the reasons are a little different. Suppose a set of queues have
each been assigned bandwidth limits to ensure that no traffic class
starves out another traffic class. The result may be that one or
more queues have exceeded their allocation while the queues that
deserve scheduling opportunities are empty.
The question then is how is the excess (idle) bandwidth allocated.
Conceivably, the scheduling criteria for excess capacity are
completely different from the criteria that determine allocations
under uniform load. This could be supported with a scheduling
hierarchy. However, the problem is that the criteria for using the
subsequent scheduler are different from those in the last two cases.
Specifically, the next scheduler should only be used if a scheduling
opportunity exists that was passed over by the prior scheduler.
When a scheduler chooses to forgo a scheduling decision, it is
behaving as a non-work conserving scheduler. Work conserving
schedulers, by definition, will always take advantage of a scheduling
opportunity, irrespective of which queue is being serviced and how
much bandwidth it has consumed in the past. This point leads to an
interesting insight. The semantics of a non-work conserving
scheduler are equivalent to those of a meter, in that if a packet is
in profile it is given the scheduling opportunity, and if it is out
of profile it does not get a scheduling opportunity. However, with
meters there are semantics that determine the next action behavior
when the packet is in profile and when the packet is out of profile.
Similarly, with the non-work conserving scheduler, there needs to be
a means for determining the next scheduler when a scheduler chooses
not to utilize a scheduling opportunity.
Figure 7 illustrates this last scenario. It appears very similar to
Figure 6, except that the binding between the allocation scheduler
and the WRR scheduler is using a FailNextScheduler association. This
association is explicitly indicating the fact that the only time the
WRR scheduler would be used is when there are non-empty queues that
the allocation scheduler rejected for scheduling consideration. Note
that Figure 7 is incomplete, in that typically there would be several
more queues that are bound to an allocation scheduler and a WRR
3.11.4. Hierarchical CBQ Scheduler
A hierarchical class-based queuing (CBQ) scheduler is the fourth
scenario to be considered. In hierarchical CBQ, each queue is
allocated a specific bandwidth allocation. Queues are grouped
together into a logical scheduler. This logical scheduler in turn
has an aggregate bandwidth allocation that equals the sum of the
queues it is scheduling. In turn, logical schedulers can be
aggregated into higher-level logical schedulers. Changing
perspectives and looking top down, the top-most logical scheduler has
100% of the link capacity. This allocation is parceled out to
logical schedulers below it such that the sum of the allocations is
equal to 100%. These second tier schedulers may in turn parcel out
their allocation across a third tier of schedulers and so forth until
the lowest tier that parcels out their allocations to specific queues
representing relatively fine-grained classes of traffic. The unique
aspect of hierarchical CBQ is that when there is insufficient
bandwidth for a specific allocation, schedulers higher in the tree
are tested to see if another portion of the tree has capacity to
Figure 8 demonstrates this example with two tiers. The example is
split in half because of space constraints, resulting in the CBQTier1
scheduling service instance being represented twice. Note that the
total allocation at the top tier is 50 Mb. The voice allocation is
22 Mb. The remaining 23 Mb is split between FTP and Web. Hence, if
Web traffic is actually consuming 20 Mb (5 Mb in excess of the
allocation). If FTP is consuming 5 Mb, then it is possible for the
CBQTier1 scheduler to offer 3Mb of its allocation to Web traffic.
However, this is not enough, so the FailNextScheduler association
needs to be traversed to determine if there is any excess capacity
available from the voice class. If the voice class is only consuming
15 Mb of its 22 Mb allocation, there are sufficient resources to
allow the web traffic through. Note that FailNextScheduler is used
as the association. The reason is because the CBQTier1 scheduler in
fact failed to schedule a packet because of insufficient resources.
It is conceivable that a variant of hierarchical CBQ allows a
hierarchy for successful scheduling as well. Hence, both
associations are necessary.
Note that due to space constraints of the document, the
SchedulingService CBQTier1 is represented twice, to show how it is
connected to all the other objects.
4. The Class Hierarchy
The following sections present the class and association hierarchies
that together comprise the information model for modeling QoS
capabilities at the device level.
4.1. Associations and Aggregations
Associations and aggregations are a means of representing
relationships between two (or theoretically more) objects.
Dependency, aggregation, and other relationships are modeled as
classes containing two (or more) object references. It should be
noted that aggregations represent either "whole-part" or "collection"
relationships. For example, aggregation can be used to represent the
containment relationship between a system and the components that
constitute the system.
Since associations and aggregations are classes, they can benefit
from all of the object-oriented features that other non-relationship
classes have. For example, they can contain properties and methods,
and inheritance can be used to refine their semantics such that they
represent more specialized types of their superclasses.
Note that an association (or an aggregation) object is treated as an
atomic unit (individual instance), even though it relates/collects/is
comprised of multiple objects. This is a defining feature of an
association (or an aggregation) - although the individual elements
that are related to other objects have their own identities, the
association (or aggregation) object that is constructed using these
objects has its own identity and name as well.
It is important to note that associations and aggregations form an
inheritance hierarchy that is separate from the class inheritance
hierarchy. Although associations and aggregations are typically bi-
directional, there is nothing that prevents higher order associations
or aggregations from being defined. However, such associations and
aggregations are inherently more complex to define, understand, and
use. In practice, associations and aggregations of orders higher
than binary are rarely used, because of their greatly increased
complexity and lack of generality. All of the associations and
aggregations defined in this model are binary.
Note also that by definition, associations and aggregations cannot be
Finally, note that associations and aggregations that are defined
between two classes do not affect the classes themselves. That is,
the addition or deletion of an association or an aggregation does not
affect the interfaces of the classes that it is connecting.
4.2. The Structure of the Class Hierarchies
The structure of the class, association, and aggregation class
inheritance hierarchies for managing the datapaths of QoS devices is
shown, respectively, in Figure 9, Figure 10, and Figure 11. The
notation (CIMCORE) identifies a class defined in the CIM Core model.
Please refer to [CIM] for the definitions of these classes.
Similarly, the notation [PCIME] identifies a class defined in the
Policy Core Information Model Extensions document. This model has
been influenced by [CIM], and is compatible with the Directory
Enabled Networks (DEN) effort.
| +--LogicalElement (CIMCORE)
| +--Service (CIMCORE)
| | |
| | +--ConditioningService
| | | |
| | | +--ClassifierService
| | | | |
| | | | +--ClassifierElement
| | | |
| | | +--MeterService
| | | | |
| | | | +--AverageRateMeterService
| | | | |
| | | | +--EWMAMeterService
| | | | |
| | | | +--TokenBucketMeterService
| | | |
| | | +--MarkerService
| | | | |
| | | | +--PreambleMarkerService
| | | | |
| | | | +--TOSMarkerService
| | | | |
| | | | +--DSCPMarkerService
| | | | |