JP2011223186A - Congestion controller and congestion control method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a congestion controller and a congestion control method capable of producing an effect stably and efficiently, and always minimizing the number of flows where a communication rate drops sharply.SOLUTION: The congestion controller 1 comprises window size addition means 7 for adding the current window size, as information, to the head packet of a packet group for the window size, window size information detection means 3 for detecting the information of the window size added to the head packet, flow selection means 4 for selecting a flow containing the largest window size from the window size information detected by the window size information detection means 3, packet discard means 6 for discarding other packet than the head packet of the flow selected by the flow selection means 4, and a control unit (control means) 5 which controls each means. The congestion controller 1 is connected with an edge node 9 by a network 8 through an interface 2.

Description

  The present invention relates to a congestion control device and a congestion control method, and more particularly to a congestion control device and a congestion control method that efficiently avoid a TCP synchronization phenomenon.

In recent years, general-purpose communication devices based on low-cost TCP / IP have been applied not only to the Internet but also to closed networks in companies. And as various applications are taken into the network, countermeasures against congestion are becoming more and more important. The basic measure to prevent congestion is to increase network equipment so that congestion does not occur. However, if it is unavoidable to delay the increase for economic reasons, or redundancy / burden is required. When a failure occurs in a distributed communication path, there is a high possibility that congestion will occur at a node where traffic is concentrated. For this reason, it is necessary to take measures to efficiently perform congestion control even when congestion occurs.
Therefore, TCP, which is used as a transfer protocol in most applications, has a congestion control function that suppresses the communication speed in order to suppress traffic within the equipment capacity. That is, when congestion occurs in a node in the communication path and a packet is discarded, the TCP of the transmission source detects it and reduces the window size that determines the number of packets that can be transmitted at one time, thereby reducing the packet to the network. It suppresses the transmission speed and eliminates congestion.

  On the other hand, many routers and switches employ a tail drop method for buffers. In the tail drop method, packets are stored in the buffer in the order of arrival, but when the buffer becomes full due to congestion, the arriving packet is discarded. At this time, since the arriving packets are randomly discarded, the packets are discarded almost continuously over a plurality of flows (here, the flow is a communication connection in which the transmission side and the reception side are connected one-to-one). ). For this reason, the TCP congestion control function works simultaneously on a plurality of flows, and the window size is reduced to half (or one packet), so that the communication efficiency is greatly reduced temporarily. Also, from this state, the window size is increased all at once. At this time, the amount of packets in the buffer has decreased, and accordingly, the RTT (Round Trip Time) has also decreased. The increase in size is abrupt. As a result, congestion occurs again and packet discard occurs. This phenomenon is called TCP synchronization and has a problem of causing a reduction in communication efficiency.

As a method for solving such TCP synchronization, Patent Document 1 discloses that when packets that reach the buffer have to be discarded due to the occurrence of congestion, instead of random discarding, for example, the flow A and B Only packets are discarded intensively (flow can be identified from packet header information). The selection criterion for the flow to be discarded is a flow of a packet that has just arrived, a flow with a high arrival rate, a random flow, or the like. As a result, packets of flows C and D enter the buffer and pass through the buffer as they are.
Since the flows A and B all decrease the window size after the packets are discarded, they are reduced in speed, but the flows C and D increase the speed so as to make up for it. The decline does not occur. Flows A and B then increase the window size together. At this time, the amount of packets in the buffer has not decreased, and the RTT has not decreased. Become. In addition, since the number of flows that increase all together is small, congestion does not occur again. For this reason, TCP synchronization does not occur and a reduction in communication efficiency can be avoided.

Japanese Patent Laid-Open No. 11-146013

However, the prior art disclosed in Patent Document 1 also has a problem. That is, the conventional method concentrates packet discards on a specific flow, so that other flows can be prevented from being discarded, and the effect of avoiding TCP synchronization is produced. It is necessary to minimize the number of flows to be performed and maximize the number of flows not to discard packets in order to produce the effect stably and efficiently. This is necessary from the viewpoint of RTO (Retransmission Time Out) operation for the packet discarding flow, and the communication speed is greatly reduced, but it does not have such a function. For this reason, there is a problem that the effect varies from time to time and is not stable, and when the effect is low, the communication speed is greatly reduced and a lot of flows are generated unnecessarily.
The present invention has been made in view of such a problem, and in the TCP on the transmission side, the current window size is added to the head packet as information in the packet group for the window size, and the node in the congestion state Then, the window size information added to the head packet is detected, and using this information, a flow including a packet as close to the head as possible is selected from the packet group corresponding to the window size. Provided is a congestion control device and a congestion control method that can produce effects stably and efficiently by continuously discarding, and can always minimize the number of flows whose communication speed is greatly reduced. For the purpose.

In order to solve such a problem, the present invention provides a congestion control apparatus for avoiding a TCP synchronization phenomenon, wherein a current window size is used as information for a leading packet in a packet group corresponding to a window size. The window size adding means to be added, the window size information detecting means for detecting the window size information added to the head packet, and the window size information detected by the window size information detecting means, Flow selection means for selecting a flow including a packet as close to the head as possible from the packet group, packet discarding means for continuously discarding packets of the flow selected by the flow selection means, and control for controlling each means Means.
In TCP, the number of packets sent at a time by the transmission side without waiting for a reply of an acknowledgment packet from the reception side is only the window size. For this reason, the number of packets discarded per flow is at most the window size. The number of discarded packets varies depending on the position in the packet group corresponding to the arrival window size from which discarding is started. That is, if it is the first in the packet group, the maximum number of packets corresponding to the window size is discarded, but if it is the last in the packet group, the number is one packet at the minimum. That is, in order to discard the maximum number of packets for one flow, it is necessary to discard the first packet as much as possible in the packet group for the window size to be reached. Therefore, in the present invention, in the TCP on the transmission side, the current window size is added to the head packet as information in the packet group corresponding to the window size, and the window size added to the head packet in the congested node. This information is used to select a flow including a packet as close to the head as possible from the packet group corresponding to the window size, and the packets of this flow are continuously discarded. This minimizes the number of flows that discard packets and maximizes the number of flows that do not discard packets.

In a second aspect of the present invention, when the window size information cannot be added by the TCP on the transmission side, the control means returns a pseudo ACK to the transmission side at the edge node, and Is characterized in that after adding a new TCP, the window size information is added to the first packet in the new TCP.
Due to various reasons, there may be a case where window size information cannot be added by TCP on the transmission side. In such a case, window size information is added to the top packet in the newly re-established TCP. As a result, even when the window size information cannot be added by the TCP on the transmission side, the window size information can be added by the new TCP.

According to a third aspect of the present invention, when the control means recognizes that the window size information detecting means has detected the window size information, the control means deletes the window size information.
In the congested node, the window size information added to the head packet is detected. However, if the window size information is transmitted while the window size information is added, the reception side does not necessarily have a problem. Therefore, in the present invention, after the detection, it is erased so as not to affect the receiving side.

4. A congestion control method for a congestion control apparatus comprising window size adding means, window size information detecting means, flow selecting means, packet discarding means, and control means, wherein the window size adding means Adding the current window size as information to the first packet in the packet group of minutes, the window size information detecting means detecting the window size information added to the first packet, A flow selecting unit selecting a flow including the maximum window size from the window size information analyzed by the window size information detecting unit; and the packet discarding unit selecting the flow selected by the flow selecting unit. Packets other than the first packet of And step, said control means, characterized by comprising the steps of: controlling said respective means.
There exists an effect similar to Claim 1.

According to the present invention, the effect can be generated stably and efficiently, and the number of flows whose communication speed is greatly reduced can always be minimized.
In addition, since there is no change in the TCP on the sending / receiving side, it can be dealt with only by the function of the node in the network. This method can also be applied to cases where the TCP cannot be changed, which is advantageous in terms of actual introduction.

It is a block diagram which shows the function of the congestion control apparatus which concerns on one Embodiment of this invention. It is a figure which shows an example of a network diagram. (A) is a figure explaining packet discard by a conventional method, (b) is a figure explaining packet discard by this invention.

  Hereinafter, the present invention will be described in detail with reference to embodiments shown in the drawings. However, the components, types, combinations, shapes, relative arrangements, and the like described in this embodiment are merely illustrative examples and not intended to limit the scope of the present invention only unless otherwise specified. .

FIG. 1 is a block diagram showing functions of a congestion control apparatus according to an embodiment of the present invention. This congestion control device 1 detects window size addition means 7 for adding the current window size as information to the first packet in the packet group for the window size, and window size information added to the first packet. A window size information detection unit 3, a flow selection unit 4 for selecting a flow including the maximum window size from the window size information detected by the window size information detection unit 3, and a flow selected by the flow selection unit 4. A packet discarding unit 6 for discarding other packets except the head packet and a control unit (control unit) 5 for controlling each unit are provided. The congestion control device 1 is connected to the edge node 9 by the network 8 via the interface 2. Connected with. In the case of this embodiment, the congestion control device 1 is provided in the congested core node 11, and the edge node 9 is also provided with the congestion control device 1 having the same configuration. Therefore, there are means that function and parts that do not function, depending on where they are provided.
That is, in TCP, the number of packets sent at one time without waiting for a reply of an acknowledgment packet (hereinafter referred to as ACK) from the receiving side is only the window size. For this reason, the number of packets discarded per flow is at most the window size. The number of discarded packets varies depending on the position in the packet group corresponding to the arrival window size from which discarding is started. That is, if it is the first in the packet group, the maximum number of packets corresponding to the window size is discarded, but if it is the last in the packet group, the number is one packet at the minimum. That is, in order to discard the maximum number of packets for one flow, it is necessary to discard the first packet as much as possible in the packet group for the window size to be reached. Therefore, in the present invention, in the TCP on the transmission side, the current window size is added to the head packet as information in the packet group corresponding to the window size, and the window size added to the head packet in the congested node. Is detected, a flow including the maximum window size is selected from the information, and all packets except the first packet of this flow are discarded. This minimizes the number of flows that discard packets and maximizes the number of flows that do not discard packets (details will be described in FIG. 3).

  FIG. 2 is a diagram illustrating an example of a network diagram. A node that aggregates the lines of the transmission end 10 is an edge node 9, and the other node is a congested node 11. There are a plurality of edge nodes 9 in the network 8. When the edge node 9 receives the packet from the transmitting end 10, it immediately returns an acknowledgment (ACK) packet to the transmitting end 10. On the other hand, the packet received by the edge node 9 is sent to the congested node 11, and when it arrives at a receiving end (not shown), an acknowledgment (ACK) packet is received from the receiving end. That is, the TCP connection is re-established at the edge node 9. At this time, when viewed from the edge node 9, the network on the transmitting end 10 side has a small RTT (Round Trip Time), so the communication speed is fast, and conversely, the network on the congested node 11 side has a large RTT, so the communication speed is slow. . Therefore, packets are naturally accumulated in the buffer of the edge node 9 due to this speed difference. This is an image in which the transmission data held by the transmission end 10 is pulled out to the buffer of the edge node 9. Based on this principle, no communication delay occurs even if packets are accumulated in the buffer. In addition, since this buffer does not accumulate packets connected by a TCP connection, even if the holding time is long, TCP is not disconnected due to a timeout.

FIG. 3 is a diagram illustrating the operation of the congestion control apparatus according to the present invention in comparison with the conventional method. FIG. 3A is a diagram for explaining packet discard according to the conventional method, and FIG. 3B is a diagram for explaining packet discard according to the present invention.
By concentrating packet discards on a specific flow, other flows can be prevented from being discarded, and the effect of avoiding TCP synchronization is produced. Therefore, it is necessary to minimize the number of flows that discard packets and maximize the number of flows that do not discard packets in order to produce the effect stably and efficiently. This is also necessary from the viewpoint of RTO (Retransmission Time Out) operation for the packet discarding flow, and the communication speed is greatly reduced (RTO is the discarding of multiple packets within the window size). When this occurs, it is determined that the degree of congestion is large, and after waiting several seconds on the transmission side, the window size is reduced to 1 packet and retransmission is started). However, the conventional method does not have such a function. For this reason, there is a problem that the flow of the communication speed greatly decreases unnecessarily when the effect is not stable because the effect varies from time to time and the effect is low.

In more detail using FIG. 3A, a case where a packet arrives with a predetermined window size for each of the flows A to D will be described. The right side of the figure is the packet arrival order to the congested node, and the left side is the later. Here, a case where the number of discarded packets is 2 will be described. In a packet group with window sizes P1-1 to P1-5, P1-1 to P1-3 are a flow A, and P1-4 to P1-5 are a flow B. Similarly, in a packet group with window sizes P2-1 to P2-12, P2-1 to P2-6 are flow B, and P2-7 to P2-12 are flow C. Similarly, P3-1 to P3-2 are a flow C in a packet group having window sizes P3-1 to P3-2. Similarly, P4-1 to P4-8 are flow D in the packet group having window sizes P4-1 to P4-8. Here, if the packets P1-1, P2-1, P3-1, and P4-1 are the first packets of the window size, there is no information in the packets P1-1, P2-1, P3-1, and P4-1. It is not added and cannot be identified by a node. As a result, when the packets P1-2 and P1-3 of the flow A and the P2-2 to P2-6 of the flow B are discarded packets, the number of discarded packets is 7, and the number of discarded flows is 2.
In contrast, the present invention addresses the discarding of the maximum number of packets per flow. As a result, the necessary number of discarded packets can be secured with the minimum number of flows. Therefore, it is possible to minimize the number of flows that discard packets and maximize the number of flows that do not discard packets.

Details will be described below.
In TCP communication, the number of packets sent at one time without waiting for a reply of an acknowledgment packet (hereinafter referred to as ACK) from the receiving side is only the window size. For this reason, the number of packets discarded per flow is at most the window size. The number of discarded packets varies depending on the position in the packet group corresponding to the arrival window size from which discarding is started. That is, if it is the first in the packet group, the maximum number of packets corresponding to the window size is discarded, but if it is the last in the packet group, the number is one packet at the minimum.
In order to return and discard the maximum number of packets per flow, it is necessary to discard from the first packet as much as possible in the packet group for the arrival window size. However, the nodes in the network do not have a function to determine where the arrived packet is in the window size packet group (although the packet order can be determined from the sequence number, the window size I ca n’t tell where the edge is).

Therefore, in the TCP on the transmission side, the current window size is added as information to the top packet in the packet group for the window size. Here, the reason why the window size is the current window size is that the window size increases and decreases with the maximum value determined by TCP as the upper limit with packet discard, and is different from time to time and from flow to flow. The window size may be represented by the number of packets or the number of bytes, but here it will be described as the former.
In order to add this information, for example, the window size field of the TCP header or the option field of the IP header may be used (the former is usually used for ACK. Is used to receive the packet in the sliding window method, and the latter is not normally used).
If information cannot be added by TCP on the transmission side for various reasons, for example, at the edge node 9, a pseudo ACK is returned to the transmission side, and a new one is sent to the reception side. After re-establishing TCP, this information may be added to the first packet by re-examining TCP.

On the other hand, the congested node 11 detects window size information added to the head packet. It should be noted that, after detection, it may be deleted so as not to affect the receiving side. A variable N whose initial value is the detected window size (represented by the number of packets in this case) is managed in the management table for each flow. The variable N is decremented by 1 each time packets of that flow arrive one after another, and becomes 1 at the last arrived packet. The congested node manages this variable N so that the current arrival packet in the packet group for the window size is located at which position and how many packets are sent for each flow. Can know.
Now assume that the buffer is full and the packet has to be discarded. In the present invention, instead of discarding packets that arrive, a flow including a packet with a larger variable N is selected from the packets currently stored in the buffer and discarded. As described above, the maximum number of packets can be discarded per one flow. As a result, it is possible to minimize the number of flows that discard packets and maximize the number of flows that do not discard packets.

  In more detail using FIG. 3B, a case where a packet arrives with a predetermined window size for each of the flows A to D will be described. The right side of the figure is the packet arrival order to the congested node, and the left side is the later. Here, a case where the number of discarded packets is 2 will be described. Since the relationship between the flow and the window size is the same as that in FIG. Here, if the packets P1-1, P2-1, P3-1 and P4-1 are the first packets having the window size, the packets P1-1, P2-1, P3-1 and P4-1 have the window size. Information is added and can be identified by the node. As a result, when the packets P4-2 to P4-8 of the flow D are discarded packets, the number of discarded packets is 7, the number of discarded flows is 1, and the number of discarded packets is the same as that in FIG. The number of discarded flows can be reduced.

DESCRIPTION OF SYMBOLS 1 Congestion control apparatus, 2 interface, 3 window size information detection means, 4 flow selection means, 5 control part, 6 packet discarding means, 7 window size addition means, 8 network, 9 edge node, 10 transmission end, 11 in congestion node

Claims (4)

A congestion control device that avoids the TCP synchronization phenomenon,
A window size adding means for adding the current window size as information to the first packet in the packet group for the window size;
Window size information detecting means for detecting window size information added to the head packet;
Using the window size information detected by the window size information detection means, a flow selection means for selecting a flow including a packet as close to the head as possible from a packet group corresponding to the window size;
Packet discarding means for continuously discarding packets of the flow selected by the flow selection means;
And a control means for controlling each means.   If the window size information cannot be added by the TCP on the transmission side, the control means returns a pseudo ACK to the transmission side at the edge node and sends a new TCP to the reception side. 2. The congestion control apparatus according to claim 1, wherein the window size information is added to the first packet in the newly re-established TCP after re-stretching.   3. The congestion control apparatus according to claim 1, wherein the control unit deletes the window size information when the window size information detection unit recognizes that the window size information is detected. A congestion control method for a congestion control device comprising window size adding means, window size information detecting means, flow selecting means, packet discarding means, and control means,
The window size adding means adding the current window size as information to the first packet in the packet group for the window size; and
The window size information detecting means detects window size information added to the head packet;
The flow selecting means selecting a flow including the maximum window size from the window size information analyzed by the window size information detecting means;
The packet discarding unit discarding a packet other than the first packet of the flow selected by the flow selecting unit;
The control means controlling each means;
A congestion control method for a congestion control apparatus, comprising:

Images