QoS in a entire network involves capabilities in the
1. End system software running on a computer, for example the operating system.
2. The networks that carry the data being sent back and forth from one endhost to another.
We shall be mostly discussing about point 2 here.The various formal metrics to measure QoS are
1. Service Availability : The reliability of users' connection to the internet device.
2. Delay : The time taken by a packet to travel through the network from one end to another.
3. Delay Jitter : The variation in the delay encountered by similar packets following the same route through the network.
4. Throughput : The rate at which packets go through the network.
5. Packet loss rate : The rate at which packets are dropped, get lost or become corrupted (some bits are changed in the packet) while going through the network.
Any network design should try to maximize 1 and 4, reduce 2, and try to eliminate 3 & 5.
Figure 1 : 2 levels of QoS [stardust]
Therefore, the need to design networks which
1. can deliver multiple classes of service - that is they should be QoS conscious.
2. is Scalable - so that network traffic can increase without affecting network performance
3. can support emerging network intensive, mission critical applications which will be the key determinant of a companies success in the global world.
Three service models have been proposed and implemented till date.
1. Best Effort services
2. Integrated services
3. Differentiated services.
Back to Top
1. Maintaining per-flow-state
2. Traffic shaping and policing
3. Congestion Avoidance
4. Congestion Management
5. Link Efficiency Mechanisms
IntServ differentiates between the following categories of application :
1. Elastic Applications: No problem for delivery as long as packets do reach their destination. Applications over TCP fall into this category
since TCP does all the hard work of ensuring that packets are delivered. There is no demand on the delay bounds or bandwidth requirements. e.g. web browsing and email.
2. Real Time Tolerant (RTT)Applications : They demand weak bounds on the maximum delay over the network. Occasional packet loss is acceptable. e.g. video applications which use buffering, which hides the packet losses from the application.
3. Real Time Intolerant (RTI)Applications : This class demands minimal latency and jitter. e.g. 2 people in a videoconference. Delay is unacceptable and ends should be brought as close as possible. The whole application should simulate 2 persons talking face to face.
To service these classes, RSVP with the various mechanisms at the routers delivers the following classes of service :
1. Guaranteed Service [gqos] :This service is meant for RTI applications. This service guarantees
a. bandwidth for the application traffic
b. deterministic upper bound on delay.
It is important for interactive applications or real time applications. Applications can decrease delay by increasing demands for bandwidth.
2. Controlled Load Service [clqos]: This is meant to service the RTT traffic. The average delay is guaranteed, but the end-to-end delay experienced by some arbitrary packet cannot be determined deterministically. e.g. H.323 traffic.
Figure 2 : IntServ Architecture
A host wanting to send some data requiring QoS sends a special data packet - called a PATH message - to the intended receiver. This packet has the characteristics of the traffic the sender is going to send within it. The router and intermediate forwarding devices install a path state with the help of this PATH message and become aware of their adjacent RSVP aware devices. Thus a path from the source to the destination is pinned down. If the path cannot be installed then a PATH Error message is sent upstream to the sender who generated the PATH message.
After the receiver gets the PATH message, it issues a RESV message. There can be 3 types of reservations - shared reservations, wildcard filter type reservations and shared explicit type reservations. They are described in section 2.2. By this time all devices along the path have established a path state and are aware of the traffic characteristics of the potential flow. RESV contains the actual QoS characteristics expected by the receiver. Different receivers may specify different Qos features for the same multicast flow. RESV exactly traces back the path taken by the PATH message. Thus each device along the path gets to know the actual QoS characteristics of the flow requested by the receiver & each decide independently how much of the demand it should satisfy or refuse altogether. If it refuses then a RESV Error message is issued downstream to the receiver who generated it in the first place.
Note : RSVP is NOT a routing protocol. Rather it uses the path established by standard routing protocols like OSPF and RIP to determine its next hops. RSVP is a transport layer protocol if one follows the OSI 7 layer model. Therefore with changes in its routes due to link or node failures, RSVP needs to update its path states at the various links. This is done by the sender issuing PATH messages and the receiver answering with RESV messages periodically. So the path states are "soft" - since they timeout after some time interval and become invalid. Once the PATH messages are received, the path states are again created. So to have persisting path states, PATH and RESV messages should be periodically issued.
After the RESV messages is received by the source, if there are no RESV Error messages, then the source sends a RESV Confirmation message to any node who wants it. It is a oneshot message. Immediately afterwards the sender starts transmitting its messages. The intermediate network's forwarding devices service the data granting it use of the reserved resources. At the end of transmission the sender issues a PATH Tear message to which the receiver answers with a RESV Tear message. They are routed exactly like PATH and RESV messages respectively. RESV & PATH Tear messages can be issued by either a end system to explicitly break a RSVP connection or by routers , due to a timeout on a state.
1. Distinct Reservation : The receiver requests to reserve a portion of the bandwidth for each sender. In a multicast flow with multiple senders each senders flow can thus be protected from other senders' flow. This style is also called as the Fixed Filter Style.
2. Shared Reservations : Here the receiver requests the network elements to reserve common resources for all the sources in the multicast tree to share among themselves. This style is important for applications like vide conferencing, where one sender transmits at a time, since it leads to optimum usage of resources at the routers. They are of 2 types
2a Wildcard Filter Type : The receiver requests resources to be reserved for all the sources in the multicast tree. Sources may come and go but they should share the same resources to send their traffic, so that the
sink can receive from all of them.
2b Shared Explicit Reservation : This is exactly like the wildcard filter type except that the receiver chooses a fixed set of senders out of all available senders in the multicast flow to share the resources.
In many areas of the internet, the network elements might not be RSVP or IntServ capable. In order for RSVP to operate through these non RSVP clods, RSVP supports tunneling through the cloud. RSVP PATH and RESV request messages are encapsulated in the IP packets and forwarded to the next RSVP capable router downstream and upstream respectively.
Now we will be looking into the various other strategies implemented at the network elements and forwarding devices such as router, switches and gateways which work in tandem with signaling protocols like RSVP to ensure end-to-end QoS.
Policy decision point (PDP) is the logical entity which interprets the policies pertaining to the RSVP request & formulates a decision. It decides who gets what QoS,when, from where to where and so on. "PDP makes decisions based on administratively decided policies which reside on a remote database such as a directory service or a network file system"[ rsvp].
PEP's and PDP's can reside on the same machine but that would lead to scalability and consistency problems. Therefore separate policy servers exist by which a single PDP can serve multiple PEP's. PDP's are useful centers for network monitoring since they are like the central headquarters where all QoS-requesting traffic have to get approval from.
a. Admission control requests : If a packet is just received by a PEP, it asks the PDP for a admission control decision on it.
b. Resource Allocation request : The PEP requests the PDP for a decision on whether, how and when to reserve local resources for the request.
c. Forwarding request : PEP asks PDP how to modify a request and forward it to the other network devices.
COPS relies on TCP for reliable delivery. It may use IPSec for security purposes. Since PDP's and PEP's are stateful with respect to path or reservation requests, so RSVP refresh messages need not be passed to them via COPS. If a path or reservation state timeouts or a RSVP Tear message comes, then a COPS Delete message is issued to the PDP to remove the state. COPS also has provisions for changing the data headers so that it can communicate with other policy servers. COPS can accommodate RSVP in multicast flows since COPS distinguishes between 'forwarding requests' and 'admission control requests' and so differentiates between sender and receiver of RSVP flows and can control which messages are transmitted and where they are sent.
a. Creating different queues for different classes of traffic
b. A algorithm for classifying incoming packets and assigning them to different queues.
c. Scheduling packets out of the various queues and preparing them for transmission.
There are four types of queuing techniques commonly implemented
a. First in first out (FIFO) queues- Packets are transmitted in the order in which they arrive. There is just one queue for all the packets. Packets are stored in the queue when the network is congested and sent when there is no congestion. If the queue is full then packets are dropped.
b.Weighted Fair Queuing
Packets are classified into different "conversation messages" by inspection of the ToS value, destination and source port number, destination and source IP address etc. One queue is maintained for each "conversation". Each queue has some priority value or weight assigned to it (once again calculated from header data). Low volume traffic is given higher priority over high volume traffic. e.g. telnet traffic over ftp traffic. After accounting for high priority traffic the remaining bandwidth is divided fairly among multiple queues (if any) of low priority traffic. WFQ also divides packet trains into separate packets so that bandwidth is shared fairly among individual conversations. The actual scheduling during periods of congestion is illustrated through the following example [CISCO1]:
If there are 1 queue each of priority 7 to 0 respectively then the division of output bandwidth will be :
total = w0+w1+w2+w3+w4+w5+w6+w7 = Sw
priority 0 gets w0/Sw th of bandwidth, priority 1 gets w1/Sw th of bandwidth,
priority 2 gets w2/Sw th of bandwidth etc.
Figure 3 : Weighted Fair Queueing
The aim of WFQ is to ensure that low volume high priority traffic does get the service levels it expects. It also adapts itself whenever the network parameters change. WFQ cycles through the fair queues and picks up bytes proportional to the above calculation for transmission from each queue. "WFQ acts as a preparator for RSVP, setting up the packet classification and scheduling required for the reserved flows. Using WFQ, RSVP can deliver guaranteed service. RSVP uses the mean data rate, largest amount of data the router will keep in the queue and the minimum QoS to determine bandwidth reservation." During congestion periods ordinary data packets are dropped but messages which have control message data still continue to get enqueued.
c. Custom Queuing
In this method separate queues are maintained for separate classes of traffic. The algorithm requires a byte count to be set per queue. That many bytes rounded of to the nearest packet is scheduled for delivery. This ensures that the minimum bandwidth requirement by the various classes of traffic is met. CQ round robins through the queues, picking the required number of packets from each. If a queue is of length 0 then the next queue is serviced. The byte counts are calculated as illustrated in the following example :
Suppose we want to allocate 20% for protocol A, 20% for protocol B, 20% for protocol C. Packet sizes for A is 1086 bytes, B is 291 bytes, C is 831 bytes.
Step1. Calculate % / size ratio : 20/1086 , 60 / 291 , 20 / 831
Step2. Normalize (by dividing by smallest number) : 1, .20619/.01842 , .02407 / .01842
Step3. Round upto nearest integer : 1, 12 ,2
Step4. Multiply each by corresponding byte size of packet : 1086 , 3492, 1662
Step5. Add them : 1086 + 3492 + 1662 = 6240
Step6. 1086/6240 , 3492/6240 , 1662/6240 or 17.4 , 56 , 26.6 which are nearly equal to the ones at the top. CQ is a static strategy. It does not adapt to the network conditions. The system takes a longer while to switch packets since packets are classified by the processor card.
d. Priority Queuing
We can define 4 traffic priorities - high, medium, normal and low. Incoming traffic is classified and enqueued in either of the 4 queues. Classification criteria are protocol type, incoming interface, packet size, fragments and access lists. Unclassified packets are put in the normal queue. The queues are emptied in the order of - high, medium, normal and low. In each queue, packets are in the FIFO order. During congestion, when a queue gets larger than a predetermined queue limit, packets get dropped. The advantage of priority queues is the absolute preferential treatment to high priority traffic - so that mission critical traffic always get top priority treatment. The disadvantage is that it is a static scheme and does not adapt itself to network conditions and is not supported on any tunnels.
Another strategy for improving link efficiency is CPTR- Compressed Real Time Protocol header - where the header of a RTP packet is compressed from 40 bytes to 2-5 bytes before transmission. The decompressor can easily reconstruct the headers since often they do not change and even if they do, the second order difference is constant.
1. Tail drop: As usual at the output we have queues of packets waiting to be scheduled for delivery. Tail drop simply drops a incoming packet if the output queue for the packet is full. When congestion is eliminated queues have room and taildrop allows packets to be queued. The main disadvantage is the problem of TCP global synchronization where all the hosts send at the same time and stop at the same time. This can happen
because taildrop can drop packets from many hosts at the same time.
2. Random Early Dropping: REDstrategies should only be employed on top of reliable transport protocols like TCP. Only then can they act as congestion avoiders. RED starts dropping packets randomly when the average queue size is more than a threshold value. The rate of packet drop increases linearly as the average queue size increases until the average queue size reaches the maximum threshold. After that a certain fraction - designated as mark probability denominator - of packets are dropped - once again randomly. The minimum threshold should be greater than some minimum value so that packets are not dropped unnecessarily. The difference between maximum and minimum threshold should be great enough to prevent
3. Weighted Random Early Dropping (WRED) - is a RED strategy where in addition it drops low priority
packets over high priority ones when the output interface starts getting congested. For IntServ environments WRED drops non-RSVP-flow packets and for Diff Serv environments WRED looks at IP precedence bits to decide priorities and hence which ones to selectively dump. WRED is usually configured at the core routers since IP precedence is set only at the core-edge routers. WRED drops more packets from heavy users than meager users - so that sources which generate more traffic will be slowed down in times of congestion. Non IP packets have precedence 0 - that is highest probability to be dropped. The average queue size formula is :
average = (old_average * 2^ (-n)) + (current_queue_size * 2^(-n ))
where n is the exponential weight factor configured by the user. A high values of n means a slow change in the "average" which implies a slow reaction of WRED to changing network conditions - it will be slow to start and stop dropping packets. A very high n implies no WRED effect. Low n means WRED will be more in synch with current queue size and will react sharply to congestion and decongest ion. But very low n means that WRED will overreact to temporary fluctuations and may drop packets unnecessarily.
Back to Top
DS boundary nodes can be both ingressnodes and egressnodes depending on direction of traffic flow. Traffic enters the DS cloud through a ingress node and exits through a egress node. A ingress node is responsible for enforcing the TCA between the DS domain and the domain of the sender node. A egress node shapes the outgoing traffic to make it compliant with the TCA between its own DS domain the the domain of the receiver node.
Figure 4: The DS Byte
Unlike IntServ, DiffServ minimizes signaling by aggregation and per-hop behaviors. Flows are classified by predetermined rules so that they can fit into a limited set of class flows. This eases congestion from the backbone. The edge routers use the 8 bit ToS field, called the DS field in DiffServ terminology, to mark the packet for preferential treatment by the core transit routers. 6 bits of it are used and 2 are reserved for future use. Only the edge routers need to maintain per-flow states and perform the shaping and the policing. This is also desirable since the customer - service provider links are usually slow and so computational delay is not that much of a problem in these links. Therefore we can afford to do the computation intensive traffic shaping and policing strategies at the edge routers. But once inside the core of the service providers, packets need to be routed very fast and so we must incur minimum computational delay at any router/switch.
If only 1 behavior aggregate occupies a link, the observable forwarding behavior will generally depend only on the congestion of the link. Distinct behavioral patterns are only observed when multiple behavioral aggregated compete for buffer and bandwidth resources on a node. A network node allocates resources to the behavior aggregates with the help of the PHBs. PHB's can be defined either in terms of their resources (buffer and bandwidth), or in terms of their priority relative to other PHB's or in terms of their relative traffic properties (e.g. delay and loss). Multiple PHB's are lumped together to form a PHB Group to ensure consistency. PHB's are implemented at nodes through some buffer management or packet scheduling mechanisms. A particular PHB Group can by implemented in a variety of ways because PHB's are defined in terms of behavior characteristics and are not implementation dependent.
Figure 5 : The DiffServ Architecture
The standard for DiffServ describes PHB's as the building blocks for services. The focus is on enforcing service level agreements (SLA) between the user and the service provider. Customers can mark the DS byte of their traffic to indicate the desired service, or get them marked by the leaf router based on multifield classification (MF). Inside the core, traffic is shaped according to their Behavior Aggregates.. These rules are derived from the SLA. When a packet goes from one domain to another, the DS byte may be rewritten upon by the new networks edge routers. A PHB for a packet is selected at a node on the basis of its DS codepoint. The mapping from DS codepoint to PHB maybe 1 to 1 or N to 1. All codepoints must have some PHB associated with it. In absence of this condition, codepoints are mapped to a default PHB. Examples of the parameters of the forwarding behavior each traffic should receive are bandwidth partition and the drop priority. Examples of implementations of these are WFQ for bandwidth partition and RED for drop priority. 2 most popularly used PHB's are :
1. Assured Forwarding: It sets out-of-profile traffic to high drop priority. It has 2 levels of priority - four classes and 3 drop priorities inside each class. But the 4 classes are not implemented till date and only the 3 levels are implemented
2. Expedited Forwarding: It exercises strict admissions control and drops all excessive packets. Thus it prevents a queue from growing beyond a certain threshold. Forwarding is either based on priority of the packets or best effort. It guarantees a minimum service rate and has the highest data priority. So it is not affected by other PHB's.
The AF PHB is used to provide Assured Servicesto the customer, so that the customers will get reliable services even in times of network congestion. The customer gets a fixed bandwidth from the ISP - which is specified in their SLA. Then it is his responsibility to decide how his applications share the bandwidth.
The EF PHB is used to provide Premium Serviceto the customer. It is a low-delay, low-jitter service providing near constant bit rate to the customer. The SLA specifies a peak bit rate which customer applications will receive and it is the customers responsibility not to exceed the rate, in violation of which packets are dropped.
EF PHB is implemented in a variety of ways. For example, if a PQ is used ( as described in section 2.5) - then there must be a upper bound (configured by the network administrator) on the amount of EF traffic that should be allowed. EF traffic exceeding the bound is dropped.
1. Behavioral Aggregate Classifiers - These select packets on the basis of their DS codepoints.
2.Multi Field Classifiers- They select packets based on values of multiple header fields.
Classifiers send the packet to the conditioner module for further processing. Classifiers are configured by some management procedure on the basis of the relevant TCA. It is also the classifier's job to authenticate the basis on which it classifies packets.
Figure 6: Various components of DiffServ
1. Meters - The conditioner receives packets from the classifier and uses a "meter" to measure the "temporal properties" of the stream against the appropriate traffic profile from the TCA. Further processing is done by the markers, shapers and policers based on whether the packet is in or out-of-profile. The meter passes this information to the other components along with the packet.
2. Markers- This marks a packet by setting the DS value to a correct codepoint (in its header). Thus a packet is categorized into a particular behavior aggregate. When a marker changes marks a packet which has already been marked, then it is said to "remark" the packet. The marker may be configured by various policies.
3. Shapers - They buffer the traffic stream and increase the delay of a stream to make it compliant with a particular traffic profile. Packets might be discarded if there is crunch of buffer space.
4. Droppers- As their name suggests, they drop packets of a stream to make the stream profile compliant. Droppers can be implemented as special case of a shaper with buffer size = 0.
In conventional IP routing, the next hop for a packet is chosen by a router on the basis of the packet's header information and the result of running a network layer routing algorithm. Choosing the next hop at the routers is thus a composition of 2 functions :
1. Partitioning the whole set of of possible packets into Forwarding Equivalence Classes (FEC).
2. Mapping each FEC to a next hop
The mapping of packets to FEC's is done at every router where largest prefix match algorithms is used to classify packets into FECs. In MPLS the assignment is done only once - at the entry of the packet in the MPLS domain (at the ingress router). The packets are assigned a fixed length value - a "label" - depending upon the FEC category to which it belongs & this value is sent alongwith the packet. At later hops, routers and switches inside the MPLS domain donot have to use complex search algorithms, but simply use this label to index into their routing tables which gives the address of the next hop for the packet, and a new value for the label.The packets label is replaced by this new label and the packet is forwarded to the next hop. This is exactly like switching. It is multiprotocolsince its techniques can be used with any network layer protocol.
If routers R1 and R2 agree to bind a FEC F to label L for traffic going from R1 to R2, then R1 is called the "upstream" label switched router (LSR) and R2 is the "downstream" LSR. If R1 & R2 are not adjacent routers, then R1 may receive packets labeled L from R2 & R3. Now a potential error condition appears if L is bound to FEC F12 between R1 and R2, and if L is bound to FEC F13 between R1 and R3, and F12 is not equal to F13. To prevent such confusion, routers like R1 should agree to a one-to-one mapping between labels and FEC's.
In MPLS, the final decision for the binding of a label L to a FEC F is made by the downstream LSR with respect to that binding. (i.e. who will be receiving the traffic). The downstream LSR then announces the binding to the upstream LSR. Thus labels are distributed in a bottom-up fashion. This distribution is accomplished with the label distribution protocols [nortel3]. LSRs using a label distribution protocol among themselves to exchange label-to-FEC-binding information are known as "label distribution peers". The label distribution protocol also comprises of the procedures used by LSR's to learn about each others MPLS capabilities [ mplsa].
The important components of the label stack are
1. The Next Hop label Forwarding Entry(NHLFE) : This is used to forward a labeled packet. It consists of
2. The Incoming Label Map (ILM) : This maps each incoming packet's label to a set of NHLFE's. If the cardinality of the resulting NHLFE set is more than 1 then exactly one NHLFE is chosen from the set.
3. The FEC to NHLFE Map (FTN): This is used to label and forward unlabelled packets. Each FEC is mapped to a set of NHLFE's by the FTN. If the cardinality of the resulting NHLFE set is more than one then exactly one NHLFE is chosen from the set.
4. Label Swapping: This is the combination of the above mentioned procedures (1,2 & 3) to forward packets. Depending on the type of the packet, forwarding can be of the following 2 categories :
1. which begins with a LSP ingress LSR that pushes the level m label on the stack
2. "all of whose intermediate LSR's make their forwarding decision by label switching on a level m label".
3. which ends at a LSP egress LSR, where the forwarding decision is taken on a level m-k (k>0) label, or some non-MPLS forwarding procedure.
A sequence of LSR's is called the "LSP for a particular FEC F" if it is a LSP of level m for a particular packet P when P's level m label is a label corresponding to FEC F.
Figure 7: The MPLS Architecture
The level m label of the level m LSP < R1,R2,...Rn > can be popped at either the LSR Rn or the LSR R(n-1). If it is popped at LSR R(n-1) then the egress LSR Rn will have to do only 1 label lookup. Otherwise Rn will have to to a label lookup followed by either another label lookup( when Rn is a intermediate router for level m-1 LSP) or an address lookup (in case m=1). A penultimate node must pop the label if it is explicitly requested by the egress node or if the next node does not support MPLS.
1. Hop by hop routing- the route is chosen at each LSR in the same manner as in conventional IP forwarding.
2. Explicit routing - Here the boundary ingress LSR specifies the particular LSP that a packet will take through the MPLS cloud. This is done by explicitly declaring all the LSP's along the path. This is almost like TCP/IP source routing although it has several advantages over TCP/IP source routing. If the whole LSP is declared then it is "strictly explicitly routed". Else for a partial LSP stated, it is called "loosely explicitly routed".
Sometimes a router R1 (say) wishes to deliver a packet straightway to another router R2 (say) which is not one of its next hop routers for that packet and neither is R2 its final destination. So to implement the tunnel, R1 pushes a additional label at the top of the label stack. Of course this label has to be agreed upon beforehand by all the routers from R1 to R2 who will be handling this packet. In its turn, R2 pops off the label once it receives the packet. This method of tunneling is much faster than in IP where the data is encapsulated in a IP network layer packet. The tunnel in this case would be a LSP (Label Switched Path) <R1,R11,R12, ... , R2>. Tunnels can be hop by hop routed LSP tunnel or explicitly routed lsp tunnel [mplsa].The label stack mechanism allows LSP tunneling at any depth.
Back to Top
Routers have to exchange various link state information and compute routes based on the information on the fly. The most popularly used link state distribution protocol is to extend link state information contained in the advertisements of OSPF. But this congests the links even further due to frequent link state advertisements. The way to reduce this congestion is to advertise only when there has been some substantial change in the network parameters - like a sharp fall in bandwidth etc. The algorithm to calculate routing tables is based on the twin parameters of hop count and bandwidth. The order is O(N*E) where N is the hop count and E is number of links in the network. The reason for choosing these 2 particular factors is as follows. Hop count is important since the more hops a traffic traverses, the more resources it consumes. So it is a important metric to consider while making routing decisions. A certain amount of bandwidth is also desired by almost all QoS sensitive traffic. The other QoS factors like delay and jitter can be mapped to hop count and bandwidth. In constraint based routing routing tables have to be computed much more frequently than with dynamic routing since routing table computations can be triggered by a myriad of factors like bandwidth changes, congestion etc. Therefore the load on routers is very high. To reduce loadthe following are implemented :
1. A large timer value to reduce frequency of the computations.
2. Choose bandwidth and hop count as constraints.
3. Preprocessing : Prune links beforehand which are obviously out of contention to be a potential route for certain kinds of flows. E.g. A 10 Mbps traffic is not likely going to be routed on a 1 Mbps link.
The advantagesof using constraint based routing is :
1. Meeting Qos requirements of the flows better
2. Better network utilization
The disadvantagesare :
1. High computation overhead
2. Big routing table size
3. A long path may consume more resources that the shortest path.
4. Unstability in the routes : Since routing tables are being updated all too often, the routes remain in the transient state much of the time and while routes are being updated, the protocol might not be sensitive to further network changes. This may lead to race conditions.
There is a certain tradeoff involved in constraint based routing between resource conservation and load balancing. Better load balancing may lead to traffic being diverted to less congested and longer links due to which the traffic travels over more hops and consumes more resources. The compromise is to use the shortest path when there is heavy network loads and use the widest path when the load is minimum.
Back to Top
"Some Layer 2 technologies have always been QoS-enabled, such as Asynchronous Transfer Mode (ATM). However, other more common LAN technologies such as Ethernet were not originally designed to be QoS-capable. As a shared broadcast medium or even in its switched form, Ethernet provides a service analogous to standard "best effort" IP Service, in which variable delays can affect real-time applications. However, the [IEEE] has "retro-fitted" Ethernet and other Layer 2 technologies to allow for QoS support by providing protocol mechanisms for traffic differentiation.
Figure 8: A SBM managed LAN
The IEEE 802.1p, 802.1Q and 802.1D standards define how Ethernet switches can classify frames in order to expedite delivery of time-critical traffic. The Internet Engineering Task Force [IETF] Integrated Services over Specific Link Layers [ISSLL] Working Group is chartered to define the mapping between upper-layer QoS protocols and services with those of Layer 2 technologies, like Ethernet. Among other things, this has resulted in the development of the "Subnet Bandwidth Manager" (SBM) for shared or switched 802 LANs such as Ethernet (also FDDI, Token Ring, etc.). SBM is a signaling protocol [SBM] that allows communication and coordination between network nodes and switches in the" SBM framework "and enables mapping to higher-layer QoS protocols" [stardust].
The steps in the algorithm are
1. DSBM Initialization : The DSBM gathers information regarding resource constraints, such as the amount of bandwidth that can be reserved, from each of its managed segment. Usually this information is configured in the SBM's and is static.
2. DSBM Client Initialization : Each client in the managed domain searches for the existence of a DSBM on each of its interfaces. If there are no DSBMs the client itself might participate in a election to become a DSBM for that segment.
3. Admission Control : DSBM clients do the following :
3a. Whenever they receive a RSVP PATH message, they forward it to their DSBM instead of the destination address. The DSBM modifies the message, builds or adds to a PATH state, adds its own L2/L3 address to the PHOP object, and then forwards it to its destination address.
3b. When a client wishes to issue a RESV message, it looks up the DSBM's address from the PHOP object of the PATH message and sends the RESV message to the DSBM's address.
3c. The DSBM processes the RESV message - if resources are not available then a RESVError message is issued to the RESV-requester, else if resources are abundant and the reservation is made then the DSBM forwards the packet to the PHOP based on its PATH state.
3d. If the domain encompasses more than one managed segment, then PATH messages propagate through the DSBM of each segment and PATH states are maintained at each DSBM. The RESV message succeeds in reaching its destination only if admission control associated with it succeeds at each DSBM.
Figure 9 : State Diagram of a SBM
While starting up, if a SBM finds that a DSBM already exists in its domain then it keeps quite till such time as election of a DSBM becomes necessary.
1. They keep on listening to the periodic announcements from the DSBM In case the DSBM server crashes or becomes non functional for some reason, they initiate a fresh round of election and elect a new DSBM. Thus they enhance the fault tolerance of the domain.
2. The load of the election algorithm is distributed among the various SBMs. Each of the SBM's send out its own priority only if it finds its priority higher than the incoming priorities.
3. If a SBM connects two otherwise disjoint L2 domains, then it makes it feasible for each one to have a separate DSBM, otherwise both the domains might have selected a single DSBM. The two DSBMs have a simpler task to perform as each of them would be presiding over a simpler network.
IntServ Disadvantages :
1. Intserv makes routers very complicated. Intermediate routers have to have modules to support RSVP reservations and also treat flows according to the reservations. In addition they have to support RSVP messages and coordinate with policy servers.
2. It is not scalable with the number of flows. As the number of flows increases, routing becomes incredibly difficult. The backbone core routers become slow when they try to accomodate an increasing number of RSVP flows.
3. RSVP is purely receiver based. The reservations are initiated by willing receivers. But in many cases, it is the sender who has the onus of initiating a QoS based flow. Thus RSVP fails to accomodate for such flows.
4. RSVP imposes maintenance of soft states at the intermediate routers. This implies that routers have to constantly monitor and update states on a perflow basis. In addition the periodic messages sent add to the congestion in the network
5. There is no negotiation and backtracking.
To solve the problem of scalability, DiffServ was proposed. It introduced the concept of "aggregating flows" so that the number of flows in the backbone of a provider's network remain managably low. But it also has several disadvantages :
1. Providing quality of service to traffic flows on a perhop basis often cannot guarantee end-to-end QoS. Therefore only Premium service wil work in a purely DiffServ setting.
2. DiffServ cannot account for dynamic SLA's between the customer and the provider. It assumes a static SLA configuration. But in the real world network topologies change very fast.
3. DiffServ is sender-oriented. Once again, in many flows, the receiver's requests have to be accounted for.
4. Some long flows like high bandwidth videoconferencing requires per-flow guarantees. But DiffServ only provides guarantees for the aggregates.
MPLS has been proposed to be a combination of the better properties of ATM and IP. It proposes switching at the core based on labels on IP packets.
Constraint based routing and other standards
Constraint based routing is used to complement various properties of IntServ, DiffServ and MPLS. It is used to compute paths so that Qos requirements are met for DiffServ flow. MPLS uses this information to lay down its LSP's. On the other hand, MPLS's perflow statistics help constraint based routing to find out better paths. Thus they share a mutually symbiotic relationship and provide a rich set of traffic engineering capabilities. The paths found by constraint based routing can be used to route RSVP PATH messages, so that subsequent RSVP flows traverse the best possible paths. IntServ, DiffServ and RSVP are essentially transport layer protocols, constraint based routing is a network layer protocol and MPLS is a network cum link layer protocol. [xiao99]
Back to Top
Figure 10 : Network Architecture
Back to Top
Back to Top
A tutorial on Nortel's view of future networks and the prevalent network architecture.
[gqos]Specification of Guaranteed Quality of Service (RFC 2212), 19 pages
Describes the guaranteed quality of service class in intserv.
3. [mplsa]Multiprotocol Label Switching Architecture draft-ietf-mpls-arch-06.txt, August 1999, 60 pages
Describes the entire mpls concept and the basic architecture.
[nortel3]IP Traffic Engineering using MPLS Explicit Routing in Carrier Networks, Nortel NetworksWhite Paper April 1999,
Describes various MPLS features and label distribution protocols.
[clqos]Specification of the Controlled-Load Network Element Service (RFC 2211), 16 pages
The formal specification of controlled load service of intserv.
6. [xiao99]Internet QoS : the Big Picture Xipeng Xiao & Lionel M. Ni, IEEE Network, January 1999,25 pages .
Describes the major QoS issues and protocols.
[rsvp2] Resource ReSerVation Protocol (RSVP) -- Version 1 Functional Specification (RFC 2205), 110 pages
This document gives the formal specification for RSVP.
8. [ciscorsvp]Resource Reservation Protocol (RSVP) CISCO White Papers, Jun 1999 , 15 pages
A very concise and to the point summary of RSVP.
9. [nortel2] Preside Quality of Service Nortel Networks Position Paper, 11 pages
Describes the future network architecture and support of QoS.
11 [CISCO1] Congestion Management Overview, CISCO White Papers, 1999, 14 pages http://www.cisco.com/univercd/cc/td/doc/product/software/ios120/12cgcr/qos_c/qcpart2/qcconman.htm
Describes various queueing techniques as implemented in CISCO IOS.
[CISCO2] Congestion Avoidance Overview, CISCO White Papers, 1999, 16 pages
Describes various RED techniques as implemented in CISCO IOS.
[CISCO3] Link Efficiency Mechanisms, CISCO white papers, 1999, 8 pages
LFI implementation as in CISO IOS product.
[CISCO4] QoS overview, CISCO white papers 1999. 24 pages
A excellent overview of QoS capabilities in CISCO's IOS product + definitions.
[diffserva] An Architecture for Differentiated Services RFC 2475, 36 pages
This described the Diff Serv architecture and model in detail
[afphb] Assured Forwarding PHB Group RFC 2597, 10 pages
The AF PHB is described here
[efphb] Expedited Forwarding PHB Group RFC 2598, 10 pages
The EF PHB is described here
[sbm] A Protocol for RSVP-based Admission Control over IEEE 802-style networks draft-ietf-issll-is802-sbm-09.txt,
The Subnet Bandwidth Manager is proposed here
[stardust] QoS Protocols and Architectures, Stardust White Paper, 17 pages
A brief tutorial on the common QoS protocols is mentioned here.
20. [rsvp]D. Durham, R. Yavatkar, Inside the Internet's Resource Reservation Protocol, John Wiley and Sons, 1999, 351 pages
Provides excellent coverage of all aspects of RSVP and some topics of IntServ. Excellent figures.
[rj99]QoS over Data Networks, Raj Jain, CIS 788 handouts, The Ohio State University, Fall 99, 4 pages of slides (6 slides per page)
Provides excellent overview of recent advances in the area of Quality of Service issues.
Back to Top