CN115695314A - Method for improving fast CNP congestion notification in network - Google Patents

Abstract

An improvement method for fast CNP congestion notification in a network, detecting message ECN field in the outgoing direction of a switch, sending the detected message to a control plane of the switch, analyzing a source IP by the control plane, and automatically constructing CNP message; the exchange chip adds a special mark to the marked ECN message, and the destination server can judge that the CNP message is transmitted by the intermediate switch instead of repeatedly transmitting the CNP message by identifying the mark; the beneficial effects are that: when the switch detects that the switch is congested, the switch can notify the source end server of the CNP message at the first time, the ECN mark of the original message is 11 and the CNP message is sent to the target server, the target server knows that congestion occurs in the middle of a link, and the CNP message cannot be sent repeatedly, so that the congestion event is passed quickly, the source end server is prevented from repeatedly receiving the CNP message, and the complexity of the source end server in processing the CNP message is reduced.

Description

Method for improving fast CNP congestion notification in network

Technical Field

The invention relates to a network communication technology, in particular to an improvement method for rapid CNP congestion notification in a network.

Background

In a network switch, network congestion is a phenomenon that often occurs, and a common scenario is that when traffic of multiple ingress ports is simultaneously forwarded to the same egress port, traffic is congested at the egress port. When congestion occurs for a long time, and when the flow of the switch exhausts its internal cache, packet loss occurs. Packet loss degrades the performance of the entire network, since packet loss means retransmission and the entire session link may be repeated based on the control layer protocol. In order to prevent packet loss caused by such congestion, the switch generally starts a flow control function. However, the packet loss caused by network congestion is avoided through flow control, and other problems are also brought;

as shown in FIG. 2, when the A-E ports of switch1 all forward traffic to the G port, the G port must be congested. If at this time, an F-incoming message for switch2 enters switch1 at port J through switch2 at port H and is to be forwarded to port G, while an X-incoming message for switch2 enters switch1 at port J through switch2 at port H and is to be forwarded to port Y. If flow control is enabled at the J port of switch1, because the flow forwarded from switch 1F port to switch 2G port is congested, the J port of switch1 will send a PAUSE frame to the H port of switch2, and let the H port temporarily stop sending flow. At this time, the traffic forwarded to switch1Y port by switch 2X port cannot be forwarded because H port suspends sending packets. As a result, the traffic forwarded from the X port to the Y port cannot be forwarded normally because the traffic from F to G is congested. The sum of X to Y traffic and F to G traffic may only occupy 10% or 1% of the H-port bandwidth, but cannot be forwarded properly. This is a side effect of opening flow control at the port;

turning to the above example, the root cause is congestion in the traffic forwarded from the a-F port to the G port. In the network, in order to avoid port flow control, the injection rate of the ports A-F is limited, when the injection rate of the ports A-F is limited, the flow control is not triggered by the port J, and the flow forwarded from the port X to the port Y can be forwarded normally. Network Congestion control now relies on Explicit Congestion Notification (ECN) techniques to reduce the injection rate at the a-F port;

ECN was originally defined in RFC 3168 for the TCP/IP protocol by embedding a congestion indicator in the IP header and a congestion acknowledgement in the TCP header. When the switch detects that the message is congested, the ECN field of the IP header of the message is marked as 11. The IP header is shown in the following figure. The ECN field has two bits, 00 indicates that the source end does not support ECN, and the switch normally forwards without marking ECN.01/10 indicates that the source supports ECN.11 indicates that congestion has occurred. When a source end server supporting the ECN sends a data message, the ECN is set to be 10 by default;

in fig. 3, 4, and 5, after a message sent by a source end server is congested in Device a and marked with an ECN field of 11, and reaches a target server through Device B, the target server finds that the ECN field of the message is marked as congested, and sends a CNP message to the source end server, and after receiving the CNP message, the source end server knows that the message sent before is congested in the entire communication link, and then reduces its sending rate.

However, when the Device a is congested in the transmission link, before the source end server receives the CNP packet, the source end server may still transmit the packet at the original rate, and during this time period, the packet may continue to be congested in the Device a, and as a result, the flow control may be triggered, and even packet loss may occur.

In order to enable a source end server to quickly receive a CNP message, a fast CNP technology is available at present, in which when a Device B detects that an ECN-marked congested message exists at an inlet, the Device B directly sends the CNP message to the source end server;

according to the above, the original rapid CNP technology has two problems;

1. when congestion occurs in Device a, CNP notifications are sent on Device B, which adds to the overhead of communication between the two devices. The fast CNP requires proprietary Device support, many switches do not support the fast CNP, and Device a may be separated from the first Device supporting the fast CNP by several hops, which directly increases the time overhead. Moreover, the link condition between Device a and the first fast CNP enabled Device cannot be guaranteed.

2. In the above example, device B is a Device supporting fast CNP, and when Device B detects an ECN packet and sends a CNP packet to a source end server, the source end server can slow down, but the destination server still receives the ECN packet and still sends the CNP packet to the source end server, which results in that the source end server continuously receives two CNP packets, and the source end server must ignore the second CNP packet, otherwise, the speed is continuously slowed down twice, which directly affects the utilization rate of the communication link.

Disclosure of Invention

The invention provides an improved method for rapid CNP congestion notification in a network, which can solve the problems mentioned in the background technology; when the network is congested, the method can provide the CNP congestion notification to the source end server more quickly, can prevent the source end server from repeatedly receiving the CNP message, and has good popularization and application prospects.

The technical scheme for solving the technical problems is as follows: the invention can detect the ECN field of the message in the outgoing direction of the switch, send the detected message to the control plane of the switch, analyze the source IP by the control plane and automatically construct the CNP message. Meanwhile, the exchange chip adds a special mark to the marked ECN message, and the destination server can judge that the CNP message is transmitted by the intermediate switch instead of repeatedly transmitting the CNP message by identifying the mark.

1. The control plane establishes a thread, when finding that the data plane receives a message marked congestion by ECN, the control plane analyzes the source MAC and the source IP in the message, constructs a CNP message and sends the CNP message to the source IP, and the outlet of the CNP message can be obtained by inquiring the routing table and the MAC table.

2. A loopback port is established inside the exchange chip, and if the outlet of the outgoing direction message is the oopback port, the processing flow of the chip can be repeated.

3. And (3) creating an ACL in the outgoing direction, wherein the rule of the ACL in the outgoing direction is a message with the matching ECN field of 11, the action is to mirror the message to the loopback port in the step (2), the first bit of the Flag field of the IP header is modified into 1, and the bit is used as a mark to inform a destination server that the CNP message is not required to be sent.

4. And (3) establishing an ACL rule in the incoming direction, wherein the rule matches with the message with an entry being a loopback interface and an ECN field being 11, and the action is to redirect to the CPU.

5. And the control plane receives the ECN marked message sent from the access direction ACL, analyzes the message structure and sends a CNP message to the source end server.

6. When the destination server receives a message with the first bit of the Flag field of the IP header being 1 and the ECN being 11, the destination server normally analyzes the data in the message and does not send a CNP message to the source IP.

In the scheme, the switch can inform the source end server of the CNP message in the first time when detecting that the switch is congested, and can still mark the ECN of the original message as 11 to be sent to the destination server, so that the destination server can know that the congestion occurs in the middle of the link and can not repeatedly send the CNP message.

Furthermore, the invention creates ACL in the incoming direction and the outgoing direction of the switch, the outgoing direction ACL detects the message with ECN of 11 and mirrors the message to the loopback port in the chip, and sets the first bit of the Flag field of the IP header to be 1, and the incoming direction directly redirects the message with ECN of 11 sent by the loopback port to the CPU. After receiving the message, the control plane constructs a CNP message and sends the CNP message to the source server. .

The invention has the beneficial effects that: the switch can inform the source end server of the CNP message in the first time when detecting that the switch is congested, and can still mark the ECN of the original message as 11 to be sent to the target server, and the target server can know that congestion occurs in the middle of a link and can not repeatedly send the CNP message, so that the congestion event is quickly passed, the source end server is prevented from repeatedly receiving the CNP message, and the complexity of the source end server in processing the CNP message is reduced.

Drawings

Fig. 1 is a schematic diagram of network congestion and control in the prior art.

Fig. 2 is a schematic diagram of an IP header format of a data packet sent by a source end server in the prior art.

Fig. 3 is a schematic diagram of a format of an ECN flag of an IP header of a data packet in the prior art.

Fig. 4 is a schematic diagram illustrating the adjustment of the ECN congestion flag and CNP congestion notification message rate in the prior art.

Fig. 5 is a schematic diagram illustrating ECN congestion marking and CNP congestion notification message rate adjustment in the fast CNP technique in the prior art.

Fig. 6 is a schematic diagram illustrating ECN congestion marking and CNP congestion notification message adjustment for network congestion control according to an embodiment of the present invention.

Detailed Description

Example (b):

in fig. 6, the method for improving fast CNP congestion notification in a network detects an ECN field of a packet in an outgoing direction of a switch, sends the detected packet to a control plane of the switch, and the control plane analyzes a source IP and automatically constructs a CNP packet. Meanwhile, the exchange chip adds a special mark to the marked ECN message, and the destination server can judge that the CNP message is transmitted by the intermediate switch instead of repeatedly transmitting the CNP message by identifying the mark;

the specific implementation steps are as follows:

step one, the control plane establishes a thread, when finding that the data plane receives a message with ECN marked congestion, the control plane analyzes a source MAC and a source IP in the message, constructs a CNP message and sends the CNP message to the source IP, and an outlet of the CNP message can be obtained by inquiring a routing table and an MAC table.

And step two, establishing a loopback port in the exchange chip, and if the outlet of the outgoing direction message is the oopback port, the processing flow of the chip can be repeated.

Step three, an ACL is created in the outgoing direction, the rule of the outgoing direction ACL is the message with the matching ECN field of 11, the action is to mirror the message to the loopback port in the step 2, the first bit of the Flag field of the IP head is modified into 1, and the bit is used as a mark to inform the destination server that the CNP message is not required to be sent.

And step four, establishing an ACL rule in the entering direction, wherein the rule matches with the message with an entry being a loopback interface and an ECN field being 11, and the action is to redirect to the CPU.

And step five, the control plane receives the ECN marked message sent from the access direction ACL, analyzes the message structure and sends the CNP message to the source end server.

And step six, the destination server receives a message with an IP header Flag field, wherein the first bit of the message is 1, and the ECN is 11, and then the data in the message is normally analyzed, and the CNP message is not sent to the source IP.

By adopting the scheme, the switch can inform the source end server of the CNP message in the first time when detecting that the switch is congested, and can still mark the ECN of the original message as 11 to be sent to the target server, so that the target server can know that congestion occurs in the link and can not repeatedly send the CNP message.

Claims (2)

1. An improved method for fast CNP congestion notification in a network comprises detecting a message ECN field in the outgoing direction of a switch, sending the detected message to a control plane of the switch, analyzing a source IP by the control plane, and automatically constructing a CNP message; meanwhile, the exchange chip adds a special mark to the marked ECN message, and the destination server can judge that the CNP message is transmitted by the intermediate switch instead of repeatedly transmitting the CNP message by identifying the mark;

the method is characterized in that:

step one, a control plane establishes a thread, when finding that a data plane receives a message with ECN marked congestion, a source MAC and a source IP in the message are analyzed, a CNP message is constructed and sent to the source IP, and an outlet of the CNP message can be obtained by inquiring a routing table and an MAC table;

step two, a loopback port is built in the exchange chip, and if the outlet of the message in the outgoing direction is the oopback port, the processing flow of the chip can be carried out again;

step three, an ACL is created in the outgoing direction, the rule of the outgoing direction ACL is that the message with the ECN field of 11 is matched, the action is that the message is mirrored to the loopback port in the step 2, the first bit of the Flag field of the IP head is modified into 1, and the bit is used as a mark to inform a destination server that the CNP message is not required to be sent;

step four, establishing an ACL rule in the input direction, wherein the rule matches with the message with the entry being a loopback port and the ECN field being 11, and the action is redirection to the CPU;

step five, the control plane receives ECN marked messages sent from an access direction ACL, analyzes the message structure and sends CNP messages to a source server;

and step six, the destination server receives a message with the first bit of the Flag field of the IP header being 1 and the ECN being 11, and then the data in the message is normally analyzed, and the CNP message is not sent to the source IP.

2. The improved method of fast CNP congestion notification in a network according to claim 1, wherein: when the switch detects that the switch is congested, the switch announces the CNP message to a source end server at the first time, and sends the ECN mark of the original message as 11 to a target server, so that the target server can know that congestion occurs in the middle of a link and cannot repeatedly send the CNP message.

Images