KR100757194B1 - A Fair Packet Scheduling Method with a High Resource Utilization and Low Implementation Complexity - Google Patents

Abstract

The process packet scheduling algorithm that guarantees the quality of service through resource reservation is divided into the end time method and the start time method in terms of the reference time of the packet used when calculating the time stamp. The former is used for most schedulers because it can support various delay specifications by adjusting the scheduling rate of traffic flow. However, the former has a disadvantage of severe resource waste and high implementation complexity. On the other hand, due to the maximum delay characteristics that depend on the number of flows, it is difficult to support various delay specifications, and thus, they are not properly utilized despite the excellent resource utilization characteristics. The present invention relates to a start time-based process packet scheduling apparatus that effectively supports various delay specifications, and is expected to provide characteristics of high resource utilization and low implementation complexity.

Quality Assurance, Process Packet Scheduling, Resource Reservation, Timestamp

Description

A Fair Packet Scheduling Method with a High Resource Utilization and Low Implementation Complexity

1 is a block diagram of a high performance scheduler in accordance with the present invention;

2 is a block diagram of a burr according to the present invention;

3 is a flowchart illustrating processing upon arrival of a packet in a department server according to the present invention.

4 is a flowchart illustrating processing at the completion of packet transmission in a department server according to the present invention.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fair packet scheduling apparatus applicable to a packet switched node, which is a core device of a high-speed packet communication network, which guarantees a quality of service through resource reservation. For each traffic flow (hereinafter referred to as "flow"), the link resource allocation method guarantees the required transmission rate (hereinafter referred to as "rate") and the delay bound (hereinafter referred to as "delay specification") required by that flow. It is about.

A hypothetical theory that is the basis of a packet scheduling method commonly known in the art is described by Abhay K.Parekh and Robert G. Gallager, "A generalized Processor Sharing Approach to Flow Control in Integrated Services Networks: The Single-node case". (IEEE / ACM Transactions on Networking, Vol. 1, No. 3, June 1993). This theory, referred to as Generalized Processor Sharing (GPS), provides a conceptual performance criterion for which all process scheduling methods are oriented, although practical implementation is impossible.

As a closest implementation of the above-mentioned GPS, Abhay K.Parekh and Robert G. Gallager's paper "A generalized Processor Sharing Approach to Flow Control in Integrated Services Networks: The Single-node case" (IEEE / ACM Transactions on Networking) , Vol. 1, No. 3, June 1993). This method, called Weighted Fair Queuing (WFQ), introduces server virtual time and uses it as a reference value when estimating the expected end time of a packet. Since WFQ uses the method of setting the expected transmission end time of an arriving packet as the timestamp of the packet (hereinafter referred to as the end time method), the maximum delay of any flow is the maximum packet size of the flow divided by the scheduling rate of the flow. (The GPS service time) and the maximum packet size of all flows divided by the transmission speed of the output link (hereafter the maximum packet transmission time). In WFQ, a new packet may arrive from every flow in the worst case while one packet is transmitted, so the server virtual time calculation and update process can be repeated N times, i.e. the server virtual time calculation complexity is O (N). . Where N is the number of flows received. Therefore, it is a difficult method to be applied to a high speed network environment that requires fast transmission order determination for packets. O (N) 's server virtual time calculation complexity is described in Korean Patent No. 10-0369562 (January 30, 2003), published as "Process Packet Scheduling Method through Emulation of WFQ in High-Speed Integrated Services Network and Its Process Packet Scheduler." By O (1).

In the above-described end time based scheduler, when the maximum delay determined by the scheduling rate of the flow exceeds the delay specification of the flow, the maximum delay must be reduced to guarantee the delay quality. This is possible by increasing the scheduling speed of the flow in the scheduler. . However, the output link speed of the scheduler device is wasted by the difference between the scheduling speed and the required speed. For reference, the request rate must be guaranteed, so the scheduling rate cannot be lower than the request rate. For example: For a scheduler with an output link speed of 1 Gbps and a maximum packet size of 1000 bits for all flows, if the scheduling rate equals the required rate for flows with a request rate of 200 Kbps, a delay specification of 1 ms, and a maximum packet size of 1000 bits. Since the service time and the maximum packet transmission time are calculated as 5ms and 1uS, respectively, the maximum delay determined is about 5ms. In order to satisfy the 1 ms delay specification, the scheduling rate must be increased to 1 Mbps, which wastes the output link speed of 1 Mbps to 200 Kbps = 800 Kbps.

Meanwhile, Pawan Goyal, Harrick M. Vin and Haichen Cheng's paper "Start-Time Fair Queueing: A scheduling Algorithm for Integrated Services Packet Switching Networks" (IEEE / ACM Transactions on Networking, Vol. 5, No5, October 1997) We proposed a method of setting the expected transmission start time of an arriving packet as the time stamp of the packet (hereinafter, referred to as a start time method). In this start time method, the maximum delay of a random flow is a sum of the maximum packet sizes of each of the received flows. It is determined by dividing by the speed of the output link. Therefore, the waste of output link speed, which is a chronic problem of the aforementioned end time method, is solved, and resource utilization of the output link can be greatly improved. However, all flows have the same maximum delay value, which means that if each flow has different delay specifications, the number of flows must be strictly limited to meet the most stringent delay specifications, and as a result, despite the available speed on the output link, Due to the limitation of the number of flows, the new link cannot be accommodated, resulting in a waste of output link speed.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and proposes a process packet scheduling scheme that can effectively support flows with different delay specifications while maintaining excellent resource utilization characteristics of the start time based process packet scheduler. It is the main technical problem.

Hereinafter, with reference to the accompanying drawings will be described in detail the operation of the preferred embodiment according to the present invention.

The term "delay class" (hereinafter referred to as "class") used in this embodiment means one delay specification or a set of delay specifications. In the case of a set of delay specifications, the most stringent delay standard becomes the delay specification of the class. The maximum number of grades being M is. 1 shows a configuration of a process packet scheduling apparatus (hereinafter, referred to as a high performance scheduler) proposed in the present invention. The high performance scheduler is composed of a class discriminator 101, a front end server 102, a rear end server 103, and a resource manager 104. The grade discriminator 101 determines the grade by the delay standard value of the arriving flow so that the flow is input to the corresponding department in the front end server 102. The front end server is composed of M department servers, and each department server is a start time-based process packet scheduler that processes only traffic of the corresponding class. Traffic output from each department is treated as one virtual flow in the rear-end server 103. The latter server has M virtual flow queues and transmits packets of each queue to the output link. The latter server is a round-robin scheduler that allows M departments in the front-end server to use the output link evenly. SS Kanhere, H. Sethu and AB Parekh's paper "Fair and efficient packet scheduling using elastic round robin" IEEE Transactions on Parallel and the Elastic Round-Robin method proposed in Distributed Systems 13 (3) (2002). The resource manager 104 allocates the speed of the output link, that is, bandwidth resources to the M departments, and manages the available resources of each department to determine whether to accept or reject the arriving flow. For reference, the high performance scheduler of FIG. 1 is for one output link, and the actual router is equipped with as many high performance schedulers as the number of output links.

The packet handled by the high performance scheduler is an internal packet including all the data packets from the flow and related information such as the flow identifier, the required speed, and the delay specification.However, in order to reduce the handling burden of large data packets in actual implementation, A separate queue manager can be stored and managed, and the data packet can be transmitted to the output link after scheduling with a small control packet having only relevant information.

The grade discriminator 101 may be configured as a separate program or may be embedded in a flow processing program.

Let's look at the operation of the shear server 102 in detail. Each department runs its own server virtual time. The server virtual time of each department is initialized to a value of zero when that department starts serving traffic, i.e. at the beginning of the server active section. When the transmission of each packet is completed, it is updated with the timestamp value of the next packet to be transmitted, and is increased in real time during the transmission of each packet.

Since all the secondary servers in the front end server have the same configuration and operate the same, the operation of any department n 105 is described. The department n includes an enqueue server n 210, a queue block 220, a dequeue server n 230, and a plurality of registers 240, 250, and 260, as shown in FIG. 2. The queue block 220 includes a separate queue for each flow. When the maximum number of flows is V, V queues are provided. The register provided includes a virtual time register Vtime (250) having a server virtual time value, an update time register Tupdate (260) having a real time value when the server virtual time was most recently updated, and the most recently arrived packet for each flow. An end tag register matrix F [V] 240 having an end tag value is included.

The enqueue server n 210 calculates a timestamp based on the server virtual time whenever a new packet is input, and adds the time stamp to the header of the packet. The traffic processed by the enqueue server n is stored in the queue of the corresponding flow in the queue block 220. When the queue server 230 selects a packet to be transmitted next time, the queue server n 230 compares the timestamp of the first packet from the queue of each flow in which the packet is waiting, and selects the packet having the smallest timestamp value to provide a transmission service.

3 is a flowchart illustrating a process of processing a packet arriving at the enqueue server n 210 illustrated in FIG. 2. Upon arrival of a new packet, the server virtual time at the time of arrival is calculated in step 320. Let t be the real time at the time of arrival, and t _n at the time when the server virtual time was most recently updated. In this case, the calculation formula of the server virtual time v (t) at the time of packet arrival is shown in Equation 1 below.

In the implementation of Equation 1, the virtual time register Vtime and the update time register Tupdate are used. The values of the registers Vtime and Tupdate are called V_Vtime and V_Tupdate, respectively. These register values are all zeros for no packets waiting in department n, i.e. during idle periods. The server virtual time v (t) of Equation 1 is calculated as V_Vtime + (t-V_Tupdate).

In step 330, a start tag (Tag) representing a virtual time at which the first bit of the packet arrives, that is, an expected transmission start time, is calculated and used as a time stamp. First, a method of calculating a start tag is used. Explain. Assuming that the arrived packet is the k th packet of flow i, the start tag S _i ^k of this packet is the server virtual time v (t) calculated in step 320 and the end tag F of the (k-1) th packet of flow i. _It is calculated by the following equation (2) by _i ^{k -1} .

In Equation 2, the end tag F _i ^{k -1} of the (k-1) th packet of the flow i is calculated as shown in Equation 3 below by increasing the packet's normalized length to the start tag of the packet.

In Equation 3, l _i ^{k - 1} is the length of the k-1 th packet of the flow i, l _i ^{k -1} / r _i Denotes the normalized length of the k-1 th packet of flow i.

To implement Equations 2 and 3, the register matrix F [V] 240 having the end tag value of the most recently arrived packet for each flow is used, and the initial values of the register matrix are all set to zero. A detailed implementation of step 330 is as follows; Comparing the server virtual time calculated in step 320 with the end tag register F [i] value of the flow to which the arrived packet belongs (flow i), a large value is set as a start tag value and is time stamped. After attaching to a packet, it is queued in the flow.

In step 340, the end tag register F [i] of flow i is updated to the normalized length value of the arriving packet plus the start tag of the packet.

4 is a flowchart illustrating a packet transmission process performed by the deqserver n 230. When the transmission of the packet being transmitted is completed (step 410), it is checked whether there is a waiting packet in the department n, that is, whether there is a waiting packet in the queue block 220 (step 420). If there is a waiting packet, in step 430, the next packet to be transmitted, that is, the packet having the smallest timestamp among the first packets of the queue of the flow, is found and updated, the server virtual time is updated, and the associated register is updated. The process of calculating server virtual time is as follows; The timestamp value of the packet having the smallest timestamp selected above is TS, the real time when the transmission of the transmitting packet is completed is t, and the real time at the time when the server virtual time is most recently updated is t _n. When the server virtual time v (t) is calculated by the following equation (4).

The implementation method of Equation 4 is as follows; A large value is set as the server virtual time by comparing TS, which is a timestamp value of the packet selected for transmission, with the value of V_Vtime + (t-V_Tupdate). The real time t is stored in the register Tupdate, and the obtained server virtual time value is stored in the register Vtime to update the server virtual time.

In step 440, the packet selected as a result of step 430 is transmitted to the downstream server.

The backend server 103 uses the proposed flexible round-robin based scheduler, whose operation is as follows; The latter server sequentially services traffic of up to M virtual flows using an active list, which means a list of virtual flows with packets queued. The latter server services the packet of the virtual flow at the head of the active list, and moves the virtual flow to the end of the active list if it is still active. When a packet arrives at an inactive virtual flow, it registers the virtual flow at the end of the active list. The service that sequentially services all the virtual flows registered in the active list once is called a round, and the latter server performs the service while repeating these rounds. The transmission allowance of each virtual flow is proportional to the speed of the scheduler output link assigned to the corresponding department, and the transmission allowance of each flow is flexible in each round.

The resource manager 105 is composed of an output link speed allocation function and a department resource management function. The output link speed allocation function allocates the scheduler's output link speed, that is, bandwidth resources to M departments, according to the static allocation that assigns fixed output link speeds to each department and the status of available resources of each department. It provides two allocation functions of dynamic allocation to optimally adjust the output link speed allocated to each department. The calculation and recalculation of the output link speed assigned to each department is determined from the point of view of maximizing profit against the allocation speed. The department available resource management function records and maintains the output link speed, delay specification, the number of flows currently accepted, the sum of the required rates of flows currently accepted and the maximum packet size of the flows currently accommodated for each department From this, the available speed and available delay are calculated as follows; The available rate is calculated by subtracting the sum of the required rate of flows currently being received from the assigned output link rate, and the available delay is the maximum delay (the maximum packet size sum of the currently accepted flows) from the delay specification. Calculated by subtracting When a new flow arrives at the scheduler, it is only accepted if there are available resources for that flow, whether or not there are available resources is determined as follows; When the grade corresponding to the delay standard of the flow is obtained and the flow is accepted for the department corresponding to the obtained grade, if the available speed and the available delay value are both negative, it is determined that there is an available resource.

On the other hand, the end time based process packet scheduler requires O (1) operation when calculating the timestamp and O (logN) during packet transmission when the maximum number of flows to be accommodated in the scheduler is N. The complexity is O (logN). In the high performance scheduler, the flows are distributed across M departments, so if all departments accommodate the same number of flows, each department will accommodate up to N / M flows. Therefore, since the implementation complexity of the high performance scheduler is O (log (N / M)) = O (logN) -O (logM), it can be reduced to a maximum O (logM) compared to the implementation complexity of the conventional end time based scheduler. For reference, the implementation complexity of the scheduler is for evaluating the constraint of the algorithm performance, and thus the implementation complexity of the most complex internal device.

As described above, the present invention configures the departments composed of the start time-based process packet scheduler with good resource utilization in parallel by the number of delay classes, and the flexible round-robin based rear end servers. As the output traffic uses the output link equally, the output packet resource utilization of the first process packet scheduler can be increased, and the implementation complexity can be reduced by the log value of the number of classes, so the scheduler can be easily implemented. Can be.

Claims (1)

A process packet scheduler apparatus applied to a high speed packet node supporting resource reservation based service quality guarantee, A rating discriminator 101 for determining a delay level based on a delay standard of the traffic flow and handing the flow to a sub server that processes the flow of the corresponding level; A front end server 102 having a start time based process packet scheduling sub-servers configured in parallel as many delay classes as the packet scheduling apparatus accommodates; A flexible round-robin based rear end server 103 which transmits the output traffic of all departments equally to the output link; And A process packet scheduling apparatus, comprising: a resource manager 104 for distributing the scheduler's output link resources to department members statically and dynamically, and for determining whether each department server accepts a new flow.

Images