Fair queue servicing using dynamic weights (DWFQ)

Abstract

In a method of fair queue servicing at a queuing point in a multi-service class packet switched network, incoming packets are received in buffers and outgoing packets are scheduled by a weighted fair queue scheduler. Real-time information of buffer usage along with the minimum bandwidth requirement is used to dynamically modify the weights of the weighted fair queue scheduler.

Description

[0002] This invention relates to the field of telecommunications, and more particularly to a method of fair queue servicing in asynchronous data networks, such as Asynchronous Transfer Mode (ATM) networks or more generally any packet switched network that supports more than one class of service.

[0003] The use of ATM by a continually increasing number of applications is driving a requirement to increase the number of service classes and to allow more flexibility in the service offerings. To support the application requirements, the ATM Forum is adding new service categories in new releases of ATM specifications. Furthermore, network providers are looking for the flexibility of defining multiple service classes for a given service category. The service classes are differentiated by their Quality-Of-Service requirements (QoS). The QoS requirements are configurable in accordance with a bi-dimensional matrix describing loss and delay. The delay jitter is another factor which needs to be bounded for some service classes.

[0004] Previously, three service categories were supported on an ATM network element, namely constant bit rate (CBR), variable bit rate (VBR) and unspecified bit rate (UBR). The CBR service is the only service that guarantees a bound on delay. It is used for time sensitive data, such as voice and video.

[0005] These various services can be supported by traditional exhaustive round-robin queuing among two priority queues. However, this simple technique cannot be used when the number of queues increases beyond two, because of the high potential of starvation for lower priority queues. Furthermore, the exhaustive round robin can only guarantee bounds on delay and delay variation for the highest priority queue. The support of multiple service class in an ATM switching product or multiplexer requires a minimum of one queue per class.

[0006] A queue scheduling algorithm, Weighted Fair Queuing (WFQ), has been recently proposed in the literature (see S. Golestani, A self-clocked Fair Queuing scheme for broadband applications. INFOCOM 1994. June 1994).

[0007] This scheduling scheme allows any number queues (service classes) to be serviced, while providing fair and work conserving access to bandwidth. One of the key features of WFQ is that the CDV (Cell Delay Variation) is bounded for any service class, as long as it is given a minimum weight.

[0008] This proposed scheme can be implemented in ATM products. However, it has not been determined how to set the servicing weights efficiently to take into account the dynamically changing bandwidth requirement of each service class (connection addition/removal. ABR flow control, Early packet Discard).

[0009] An object of the invention is to provide a framework that ensures that the weights are set appropriately to guarantee the desired Quality of Service and modified in real-time to ensure that the dynamic allocation of bandwidth across the classes is optimized.

[0010] According to the present invention there is provided a method of fair queue servicing at a queuing point in a multi-service class packet switched network, wherein incoming packets are received in buffers and outgoing packets are scheduled by a weighted fair queue scheduler characterized in that real-time information of buffer usage along with the minimum bandwidth requirement is used to dynamically modify the weights of the weighted fair queue scheduler.

[0011] Preferably the minimum bandwidth requirement is extracted during connection admission control.

[0012] The method is particularly suitable for use in ATM networks.

[0013] The DWFQ (Dynamic Weighted Fair Queuing) can be implemented at any queuing point which arbitrates servicing between n queues (n≧2).

[0014] The invention also provides a fair queue servicing arrangement in a multi-service class packet switched network, comprising a weighted fair queuing controller, and buffer means for receiving incoming packets in queues, characterized in that further comprises means for monitoring buffer usage for each queue, means for determining the bandwidth requirements of each class of service, and a service weights manager for dynamically modifying the weights of said weighted fair queuing controller means in response to said buffer usage and bandwidth requirements.

[0015] Preferably, the means for monitoring buffer usage a queue growth monitor which performs real-time estimation of the queue growth in said buffer means.

[0016] The invention will now be described in more detail, by way of example only, with reference to the accompanying drawings, in which:

[0017] FIG. 1 is a diagram depicting the high level queuing scheme at an ATM switch;

[0018] FIG. 2 shows the information provided by the Queue Growth Monitor;

[0019] FIG. 3 illustrates the data flow between the key components of the system and the action of the Service Weight Manager (SWM); and

[0020] FIG. 4 describes the process performed by the SWM.

[0021] Referring now to FIG. 1, ATM cells 2 arrive at buffer 1 and are placed in queues 11, 12, . . . 1n. From there the cells are passed to a weighted fair queuing unit 3. The buffer 1 is also connected to a queue growth monitor 4, which in turn is connected to congestion control unit 5, congestion analyzer 6, and connection admission controller 8, which in turn is connected to SVC & PVC (Switched Virtual Circuit and Permanent Virtual Circuit) connection handling unit 7, and service class manager 10. Queue growth monitor 4, connection admission controller 8 and service class manager 9 are connected to service weights manager 9, which is connected to weighted fair queuing scheduler 3.

[0022] The key element of the Dynamic Weighted Fair Queuing (DWFQ) scheme is the service weight manager (SWM) 9, which dynamically modifies the service weights to be used by the WFQ Scheduler 3. It uses real-time information from the service class manager 10, the connection admission controller 8, and the Queue growth monitor 4.

[0023] The service class manager 10 configures the service classes. A service class is configured with a given value of delay (CTD—Cell Transfer Delay) and loss (CLR—Cell Ratio Loss) requirements. These parameters represent the maximum nodal delay and loss allowed in order to meet the end-to-end QoS requirements of the connection. The service classes are mapped into a priority table as exemplified in Table 1. The priority table is used later by the service weight manager to allocate remaining bandwidth. The priority table is updated when a service class definition is modified. The service class manager also dictates which traffic descriptors are used to compute the minimum bandwidth required by a connection of a given class.

TABLE 1
Example of a Queue Service Priority Mapping.
	CTD
	CLR	100 μs	500 μs	None
	10−9	1	3	6
	10−7	2	5	8
	10−5	4	7	9

[0024] The connection admission controller (CAC) 8 computes the minimum bandwidth required for each service class. The minimum bandwidth is updated each time a connection of a given class is established or disconnected, based on its traffic descriptor.

[0025] Table 2 shows a typical example of which traffic descriptors that can be used to compute the minimum bandwidth for each basic service category relative to the queue service rate (SR). The CAC 8 communicates the minimum Weight table to the SWM every time the value of the minimum weights have changed by a factor of ξ.

TABLE 2
Example of minimum weight table.
	Queuej	Category	min_Wi
	Q1	CBR	(ΣPCR)/SR
	Q2	RT-VBR	(ΣSCR)/SR
	Q3	NRT-	(ΣSCR)/SR
		VBR
	Q4	UBR	0

[0026] The CTD is further taken into account in the target queue size (TQS) table, which is the maximum queue size allowed to limit the CTD. An example of TQSi computation is shown in Table 3, for typical service categories. A zero TQS indicates that the queue can grow without limitation. This table is computed by the CAC.

TABLE 2
Example of a target queue size computation.
	Queuej	Category	TQSi
	Q1	CBR	CTD/min_Wi
	Q2	RT-VBR	CTD/min_Wi
	Q3	NRT-	0
		VBR
	Q4	UBR	0

[0027] The Queue Growth Monitor (QGM) 4 performs real-time estimation of the queue growth every Ts cell slots (sampling interval). The information provided by the Queue Growth monitor 4 to the SWM 9 consists of ΔQi, the Queue Growth Rate of output queue during an interval of duration Ts, Qi, the length of output queue i at the sampling time, and Ai, the arrival rate during the same interval of time.

[0028] The corresponding parameters: queue size Qi, queue growth ΔQi, number of arrivals Ai are collected or each queue by the QGM 4 for each Ts interval. From these parameters, auxiliary parameters such as average arrival rate λi and service rate μi can be derived by the SWM 9:

[0029] λ1: average arrival rate, λi=Ai÷Ts

[0030] μ1: average service rate, μi=Si÷Ts, where Si=Ai−ΔQi is the number of cells served during Ts.

[0031] FIG. 2 shows the information provided by the Queue Growth Monitor 4. Using this information for the CAC 8, the service class manager 9 and the queue growth monitor 4, the SWM computes the service eight for each queue i (W1) to be used during the next sampling interval.

[0032] As can be seen in FIG. 3, which shows the data flow between the key components of the system and the action of the SWM 9, the queue weights, Wi, are updated using information provided by the Queue Growth monitor 4.

[0033] If {tilde over (λ)}1 denotes the arrival rate of cells in queue i in the coming Ts interval, then ideally, the target service rate {tilde over (μ)}1 can be calculated as: ({tilde over (μ)}1−{tilde over (λ)}1)·Ts=Q1−TQS1. This means at the end of next Ts interval, the queue size Qi will reach the target queue size TQSi. On the assumption that {tilde over (λ)}1 remains unchanged from λ1, the service weight W1={tilde over (μ)}1·Ts can be approximated as

W 1 ={tilde over (μ)}1·Ts≅λ1·Ts+Q1−TQS1=A1+Q1−TQS1.

[0034] However, the assumption on the stable arrival rate may not hold, and also the actual number of serviced cells Si could be less than Wi; therefore a more conservative approach is ΔQ1>0, then Q1+ΔQ1, the predicted queue size at the end of the next Ts interval, is used to calculate the target service rate and weight. That is:

({tilde over (μ)}1−{tilde over (λ)}1)·Ts=Q1+ΔQ1−TQS1 and

W 1 ≅λ1·Ts+Q1+ΔQ1−TQS1=A1+Q1+ΔQ1−TQS1

[0035] The detailed algorithm performed by the service weights manager 9 is shown in FIG. 4. The queue size Q1 at the end of each interval Ts, the number of arrivals Ai during the previous interval Ts, and the change in queue size ΔQi are input at step 20. Step 21 determines whether the queue growth is positive: if yes, the service weight Wi is conservatively adjusted to bring the queue size to TQSi at step 22; if no, the service weight Wi is adjusted to bring the queue size to TQSi at step 23. The difference ΔW is determined in step 24.

[0036] Step 25 determines whether the shared weights pool is empty: if yes, Wi is set to min_Wi in step 26; if no, step 27 determines whether Ws≧ΔWi: if yes, step 28 sets Wi=min_W1+ΔW1 and Ws=Ws+ΔW1; if no, step 29 sets Wi=min_Wi+Ws and Ws=0.

[0037] Step 30 runs through all the W1 in the ordered list L and step 31 updates the weight table used by the Weighted Fair Queuing scheduler 3.

[0038] The described technique complies with ITU and ATM Forum standards and can be applied to any switching equipment which supports more than a single service class of service.

Claims

1. A method of scheduling queues of a queuing point in a packet communications network, comprising the steps of: enqueuing incoming packets in queues associated with respective service classes; transmitting outgoing packets from said queues in accordance with a schedule determined by weights associated with the service classes used by a queue scheduler; and dynamically modifying the weights in real-time to cause queue size for each service class to move towards a target value whereby transfer delay associated with the service class is satisfied.