CN108306976B

CN108306976B - SDN controller architecture used in cloud computing network virtualization

Info

Publication number: CN108306976B
Application number: CN201810154017.0A
Authority: CN
Inventors: 付宏伟
Original assignee: Individual
Current assignee: Individual
Priority date: 2018-02-22
Filing date: 2018-02-22
Publication date: 2021-01-15
Anticipated expiration: 2038-02-22
Also published as: CN108306976A

Abstract

The invention relates to an SDN controller architecture used in cloud computing network virtualization. Modern cloud computing platforms often implement network virtualization by means of SDN technology, such as Neutron plug-in or Dragonflow plug-in of the Openstack platform. But the existing network virtualization schemes are not suitable for the characteristic that the cloud computing platform has elastic expansion. The invention realizes the SDN controller framework which can flexibly and flexibly stretch the number of the control nodes and the storage nodes according to various different network virtualization applications by separating the control logic processing and the state storage of the SDN controller. Compared with the traditional high-availability implementation method, the method for realizing the high-availability of the SDN controller saves more resources.

Description

SDN controller architecture used in cloud computing network virtualization

Technical Field

The invention relates to the field of cloud computing and software defined networking, in particular to an SDN controller architecture used in cloud computing network virtualization.

Background

Modern cloud computing platforms often implement network virtualization by means of SDN technology, such as the Neutron or Dragonflow plug-ins of the Openstack platform. Because SDN provides higher flexibility than conventional architectures, it is more adaptable to the need to dynamically create and allocate various network services for virtual machines in a cloud computing platform. The Neutron plug-in only uses partial SDN technology, and the Dragnflow plug-in adopts the SDN technology more completely.

The earliest Dragonflow plug-in of Openstack adopted the architecture of a single control node and multiple agents. A single control node may become a bottleneck in performance and have a problem of single point of failure in availability. Dragonflow later introduced a new distributed Dragonflow. The distributed Dragonflow requires that SDN controllers are installed on all computing nodes, the number of the control nodes is consistent with that of the computing nodes of the cloud computing platform, and data are synchronized among the SDN controllers through a distributed database. Distributed Dragonflow solves the single point problem, but has the disadvantage that the control node needs to compete with the compute nodes for resources, which may lack sufficient resources to handle SDN requests when the compute nodes are heavily loaded. The cloud computing platform has the characteristic of elastic expansion, and the increase and decrease of the computing nodes are determined according to the load of the computing nodes. The system cannot reduce the number of control nodes according to the load increase of the SDN control nodes.

Conventional SDN controllers mostly use a relational database as a data store layer. The advantage of relational databases is to provide strong data consistency and availability, i.e., they belong to CA databases from the perspective of CAP principle (any distributed system can only satisfy two of consistency, availability, and partition fault tolerance at the same time). That is, the relational database emphasizes data consistency, but it is difficult to achieve elastic scalability. However, in a cloud environment, the data storage throughput of the SDN controller tends to greatly fluctuate, and a key/value-based database, such as redis, is more suitable for data storage in the cloud environment.

Another disadvantage of relational databases is that because relational database queries use SQL statements, commonly used languages such as Python, Java that write SDN controllers are object-oriented. Resulting in impedance mismatch. The use of a key/value database also helps to solve this problem.

Disclosure of Invention

The invention aims to solve the technical problem of an SDN controller architecture applied to cloud computing network virtualization. The SDN controller architecture can flexibly and flexibly stretch the quantity of control nodes and storage nodes according to various different network virtualization applications by separating the control logic processing and the state storage of the SDN controller. Compared with the traditional high-availability implementation method, the method for realizing the high-availability of the SDN controller saves more resources.

Unlike a conventional SDN controller, in order to enable the number of control nodes to be elastically scalable, the control nodes in the present invention should only provide stateless services. Stateless means that the control node does not store the application state and all the information required to process a request is either included in the request or is available from the storage node. The advantage of stateless services is that the result of a request being processed at any one control node is the same, and an agent on a compute node may send a request to one control node and the next to another. The system can only distribute the requests according to the load condition of the control nodes, and has no influence on the processing result. When a new control node is added, the system only needs to send a new request to the new control node, because the context for processing the request is in the request or stored in the storage node, and the result of processing the request by the new control node is the same as that of the existing control node. When the system needs to delete a certain number of control nodes, the system can wait for the request being processed by the current node to be processed and then close and delete the node, and the rest of the control nodes can continue to process the subsequent requests, so that the correctness of the program is not influenced.

And the control node accesses the storage cluster in a plug-in API mode. The storage cluster may be any key/value distributed database cluster, e.g., redis. The system can change the storage cluster according to different scenes and requirements. The storage cluster is responsible for the persistence of the application state, and the specific details are configured according to the specific storage cluster.

The system may elastically scale the number of storage nodes according to the application and load requirements and balance the load on the storage cluster through a hash algorithm, such as a consistent hash algorithm, that is capable of rebalancing the storage nodes.

In order to ensure that the storage cluster can be elastically scaled, the hash algorithm used should have the following characteristics:

(1) establishing a mapping relation between the keys in the application state and the storage nodes in the storage node cluster, so that the keys are distributed to different storage nodes;

(2) the balance is ensured, namely keys are distributed to all storage nodes as far as possible, so that all the storage nodes are utilized;

(3) monotonicity is guaranteed, namely when a new storage node is added into a storage node cluster, keys distributed to the original storage node can be mapped to the new storage node;

(4) it is guaranteed that a particular storage node maps to the same content on different control nodes.

The method for elastic expansion and contraction of the storage cluster comprises the following steps:

(1) before actual scaling, recalculating the hash value of the stored application state ID according to the scaled storage node cluster;

(2) if storage nodes needing to be added exist, the storage nodes are added into the cluster, and the application state key/value mapped to the new storage node is moved from the old storage node to the new storage node;

(3) if the node needing to be removed exists, the application state key/value stored on the node is moved to the storage node mapped after the hash value is recalculated, and then the node is removed out of the storage cluster.

Each control node is assigned a unique ID when joining the cluster.

The control node assigns a unique serial number ID to each received event, and the later received event ID should be larger than the earlier received event ID.

The application state of the control node is stored on the storage node in the form of key/value. The key should be generated according to globally unique information in the request, for example, for an application that assigns an IP address to the virtual machine, the Openstack UUID of the virtual machine is unique, the UUID of the virtual machine can be used as the key, and the network information of the virtual machine, such as the mac address, the IP, and the like, can be used as the value.

Control node failure should be detected in time, and the failure detection mechanism may be implemented by Zookeeper or some gossip protocol.

When the control node fails. The transaction integrity of event processing is guaranteed. The operation of the control node to change the state stored on the storage node is divided into four types: 1. adding, adding a key-value pair; 2. deleting, deleting a key-value pair; 3. updating, namely updating one key value pair and updating the value of a certain key to another value; 4. and inquiring a key-value pair, namely inquiring the value corresponding to a given key. Other more complex operations may be broken down into combinations of these four operations.

Before processing the event, the event is stored, and the storing method comprises the following steps: generating a unique key according to the control node ID and the event ID of the received event, wherein the key is a hash value generated by using a hash algorithm according to the control node ID and the event ID; step two, the control node generates a value according to the received event, and the attribute of the value comprises the following steps: one) binary representation of the event, or some serialization of the event, the control node should be able to regenerate the event by parsing or deserializing the value, two) control node ID of the received time, three) event ID four) event transaction complete flag, which if set to true, represents that the event transaction has been processed successfully; and step three, the control node stores the key and the value on the storage node.

The system should generate and maintain a secondary index according to the control node ID and the event ID of the event, so that the control node can conveniently inquire the event according to the control node ID and the event ID.

In processing events, if the processed event requires transaction integrity, the control node saves the operation information and rollback information. The storage method comprises the following steps: the method comprises the following steps: a unique serial number is first generated before each operation to change the state of the controller is performed. The sequence number of the post-execution operation must be greater than the sequence number of the pre-execution operation, e.g., an increment operation is performed first and then an update operation is performed, the sequence numbers of the increment and update operations may be 1 and 2, respectively. Step two: generating a hash value according to the control node ID, the event ID and the serial number; and step three, generating a value for each operation needing to be saved, wherein the value should contain the following attributes, namely, the type of the operation. The operation type should be one of three operations of adding, deleting and updating. Second) operation information, for the adding operation, the content of the operation information is newly added key/value; for the delete operation, the content of the operation information is the deleted key/value; the content of the operation information following the new operation is key, the old value before the update and the new value after the update; step three, the control node saves the key and the value on the storage node; third), control node ID; four) an event ID; fifthly), operating the serial number; six) operation completion flag, if set to true, indicates that the operation has been completed.

After each operation is completed, it is marked that the operation has been completed, for example, a mark indicating whether the operation has been completed is added to the operation information. It should be atomized whether the operations handling the event and the marking operation have completed. If the distributed key/value database used by the system does not support the atomization operation, the relational database supporting the atomization operation can be used for storing the operation information, and because only the operation of the event needing transaction integrity currently being processed needs to be stored, the concurrent read-write quantity is not large, and the system performance cannot be influenced. Or a two-phase commit protocol is implemented by Zookeeper.

The system should generate and maintain a secondary index of the operation information to facilitate querying the operation information based on the control node ID, the event ID, and the operation sequence number.

After processing the event, it is to be saved whether the transaction has been completed by modifying the completion flag of the event.

When one control node fails, the other control node can continue processing. The other control node first checks whether there is an incomplete event process, and can decide whether to continue processing the event or to roll back according to the operation information stored on the storage node. The method for continuing the treatment comprises the following steps: step one, inquiring an uncompleted event of a failure control node; step two, inquiring the operation information of the uncompleted event; and step three, determining whether to continue to finish the transaction, abandon the transaction or roll back the transaction according to the operation information and the system state.

The method of the rollback transaction is as follows: step one, finding the operation which has the largest sequence number and is completed and not rolled back in the operation information of the incomplete event; step two, rolling back the operation, wherein the method for rolling back the operation is to execute a reverse operation, the reverse operation is divided into three types, 1) the reverse operation of the new adding operation is a deleting operation, and the new adding key/value is deleted; 2) the inverse operation of the deletion operation is an addition operation, and the deleted key/value is added; 3) the inverse operation of the update operation is the update operation, and the value of the key is updated to the old value stored in the operation information; step three, marking the operation as rolling back; and step four, repeating the step one to the step three until all the finished operations have been rolled back.

Claims

1. An SDN controller architecture system applied to cloud computing platform network virtualization is characterized by comprising a control node cluster and a storage node cluster;

the control node cluster stores all application states on the storage nodes and can elastically stretch and contract according to the load of the control node cluster;

the storage node cluster is a cluster of a key-value database and can elastically stretch and contract according to the load of the storage node cluster;

the SDN controller architecture system applied to the cloud computing platform network virtualization comprises a method for storing all application states on storage nodes, and the method comprises the following steps:

2.1 generating a unique control node ID for each control node;

2.2 generating a unique incremental event ID for the request received by each control node; the event ID generation method ensures that the later received event ID is greater than the first received event ID

2.3 generating a globally unique application state ID for each application state according to the content, the ID of the control node where the application state is located and the event ID;

2.4 calculating the hash value of the application state ID as a key by adopting a hash algorithm, and generating a value according to the application state;

2.5 storing the key and value generated in the step 2.4 on a storage node;

wherein the control node should only provide stateless services; the stateless state here means that the control node does not store the application state, and all information required for processing one request is either contained in the request or can be acquired from the storage node; stateless services have the benefit that the result of a request being processed at any one control node is the same, and an agent on a compute node may send a request to one control node and the next to another; the system can only distribute the requests according to the load condition of the control nodes, and has no influence on the processing result; when a new control node is added, the system only needs to send a new request to the new control node, because the context for processing the request is in the request or stored in the storage node, and the result of processing the request by the new control node is completely the same as that of the existing control node; when the system needs to delete a certain number of control nodes, the system can wait for the completion of the processing of the request being processed by the current node and then close and delete the node, and the rest of the control nodes can continue to process the subsequent requests without influencing the correctness of the program.

2. The SDN controller architecture system applied to cloud computing platform network virtualization of claim 1, wherein the method for storing all application states on storage nodes is characterized in that the hash algorithm maps keys of application states with storage nodes in a storage node cluster, so as to distribute the keys to different storage nodes;

the Hash algorithm for establishing the mapping relation between the keys and the storage nodes ensures the balance, namely, the keys are distributed on all the storage nodes as much as possible, so that all the storage nodes are utilized;

the Hash algorithm ensures monotonicity, namely when a new storage node is added into the storage node cluster, the key distributed to the original storage node can be mapped to the new storage node;

the hash algorithm ensures that a particular storage node is mapped to the same content on different control nodes.

3. The SDN controller architecture system for cloud computing platform network virtualization of claim 1, wherein the method for storing all application states on storage nodes is capable of elastic scaling of storage node clusters according to load, and wherein the method comprises the steps of:

(1) recalculating the hash value of the stored application state ID according to the storage node cluster after expansion and contraction;

(2) if storage nodes needing to be added exist, adding the storage nodes into the cluster, and moving the application states key and value mapped to the new storage node from the old storage node to the new storage node;

(3) if the node needing to be removed exists, the application states key and value stored on the node are moved to the storage node mapped after the hash value is recalculated, and then the node is removed out of the storage cluster.

4. The SDN controller architecture system applied to cloud computing platform network virtualization of claim 1, wherein the control node saves the event before processing the request event, and the saving method comprises the following steps:

(1) generating a globally unique incremental ID for each received event;

(2) calculating a hash value as a key according to the ID of the control node receiving the event and the event ID;

(3) generating a value according to the content of the received event, wherein the attribute of the value comprises the following steps: 1) binary representation of events, or some serialization of events, any one control node should be able to regenerate an event by parsing or deserializing the value; 2) receiving a control node ID of the event; 3) an event ID; 4) timestamp of time of receipt of event 5) indicia of whether the transaction was completed;

(4) saving the key and the value on a storage node;

(5) a secondary index is generated and maintained for the control node ID of the received event and the timestamp of the time of the received event, so that the system can query the event according to the control node ID and the time.

5. The SDN controller architecture system applied to cloud computing platform network virtualization, wherein the SDN controller architecture system comprises a method for achieving event processing transaction integrity, and the method comprises the following steps:

(1) after receiving the event, the control node stores the event before processing the event;

(2) the control node saves the operation before changing the operation of the application state of the control node each time in the event processing;

(3) after each operation is executed, the operation completion mark of the change saving operation is completed, and the action of actually executing the operation and the action of the change operation completion mark should be atomized, wherein the atomization refers to that the system is in a state that two actions are successful or not executed;

(4) after the event processing is finished, the completion mark for changing the saved event is finished;

(5) if a control node fails in the process of processing an event, another control node may read the saved event and operational information and decide whether to continue event processing or to rollback the system state based on the system's then-current state.

6. The SDN controller architecture system applied to cloud computing platform network virtualization, the method for achieving event processing transaction integrity, recited in claim 5, comprising a method for saving event operations, the method comprising the following steps:

7.1 before each execution of the operation of changing the state of the controller, first generating a unique serial number;

the sequence number of the later execution operation must be greater than the sequence number of the earlier execution operation;

7.2 calculating a hash value as a key according to the ID of the control node, the event ID and the serial number;

7.3 generate a value for each operation that needs to be saved, which should contain the following properties, one) operation type,

the operation type is one of three operations of adding, deleting and updating;

second) operation information, for the adding operation, the content of the operation information is newly added key/value; for the delete operation, the content of the operation information is the deleted key/value; the content of the operation information of the updating operation is key, the old value before updating and the new value after updating;

third), control node ID; four) an event ID; fifthly), operating the serial number; sixthly), if the operation completion flag is set to true, the operation is completed;

7.4 saving the key and the value on the storage node;

7.5 generating and maintaining a secondary index for the control node ID and event ID of the received event so that the system can query the operation according to the control node ID and event ID.

7. The SDN controller architecture system applied to cloud computing platform network virtualization, the method for achieving event processing transaction integrity, recited in claim 5, comprising a method for rolling back application state, the method comprising the steps of:

8.1, searching the operation which has the largest sequence number, is completed and is not rolled back in the operation information of the incomplete event;

8.2 rolling back, wherein the method for rolling back is to execute a reverse operation which is divided into three types, 1) the reverse operation of the new adding operation is deleting operation, and the new adding key/value is deleted; 2) the inverse operation of the deletion operation is an addition operation, and the deleted key/value is added; 3) the inverse operation of the update operation is the update operation, and the value of the key is updated to the old value stored in the operation information;

8.3 mark this operation as having rolled back;

8.4 repeat steps 8.1 through 8.3 until all completed operations have been rolled back.