TWI516956B - Adaptive quick response controlling system for software defined storage system for improving performance parameter - Google Patents

Description

Adaptation for software-defined storage systems to improve performance parameters Rapid response control system

The present invention relates to a control system for software defined storage, and more particularly to a control system for software defined storage to achieve specific performance metrics required in a service level protocol.

Cloud services have grown very popular in the last decade. Cloud services are based on cloud computing and provide related services or products without increasing the burden on the client. Cloud computing involves a large number of computer hosts that are connected to each other via a communication network, such as the Internet. It relies on the sharing of resources to achieve consistency and economic size. The concept of cloud computing is an infrastructure formed by a combination of network infrastructure and resource sharing services. Among all the shared services, memory and storage devices are definitely the two most demanding projects. This is because some popular applications, such as video streaming, require huge amounts of data storage. When cloud services operate, memory and storage management is very important to ensure that customers maintain normal service quality.

For example, servers used to provide cloud services typically manage or connect to several hard disks. The client uses the server and the data is read from or written to the hard disk. The delay in response time due to the limitations of the hard disk system can cause problems in service requirements. Under the operation of a normal hard disk system, an excessive response time is generated when the access speed required by the application surface (i.e., workload) exceeds that provided by the hard disk system. The maximum amount of load that a hard disk system can provide is usually the bottleneck in the operation of the entire cloud service system. In other words, the number of input/output operations per second of the hard disk system cannot meet the external requirements. In response to this problem, it is necessary to remove or reduce the workload to achieve and improve the performance of the server. In practice, part of the workload can be shared by other servers (if any) or hard disks, and these servers or hard disks will be automatically or manually added to support existing hard disks. Regardless of which of the above methods is used to solve the problem, the added cost is a large number of hard disks that are reserved in advance due to unpredictable work conditions, and the necessary power consumption for additional hardware devices. From an economic perspective, it is really not worth doing. However, since the system's service level protocol usually specifies the shortest delay time or the minimum number of input/output operations per second, it must be achieved. For operators that maintain cloud services with limited funds, how to reduce costs is an important issue.

It is worth noting that the workload of the server (hard disk system) can more or less predict the evolution in the future based on historical records, and the demand for cloud service workload is predictable. Therefore, it can be re A hard disk that is configured in a hard disk system to meet the workload requirements with minimal cost. However, a machine cannot learn how and when to reconfigure a hard disk. In many cases, this work is done by authorized personnel, depending on the status of time or on a fixed schedule, and the effect may not be very good.

Another fast-growing demand like cloud services is software definition storage. Software definition storage refers to computer data storage technology that can separate storage hardware from the software that manages the storage infrastructure. Under Software Definition Storage, you can launch some functional options such as deduplication, replication, thin provisioning, snapshots, and backups to provide policy management. With software-defined storage technology, there are several previous cases that provide solutions to the above problems. A method for reconfiguring a storage system is disclosed, for example, in U.S. Patent No. 20,130,297,907. The method comprises two main steps: receiving user demand information of the storage device, and automatically generating a function setting of the storage device and a device setting file for the storage device by using the user's demand information; and setting the function using the function device; To automatically reconfigure the storage device as one or more logical devices with independent behavioral characteristics. The content of this application points to a new way to reconfigure storage devices by software definition storage concepts. The method and system in accordance with the application also allows the user to dynamically adjust the configuration of one or more of the logical devices to more flexibly satisfy the user's demand information. However, the application does not provide a system that automatically learns how to reconfigure the storage device in accordance with changes in application requirements (ie, workload).

Accordingly, the present invention discloses a new system for implementing automatic learning and resource reallocation for software definition storage. The system uses adaptive control and operation without manual intervention.

Since the prior art cannot provide a change in the storage system according to the application requirements, it automatically learns how to reconfigure its storage device, making it difficult for the existing storage system to meet the requirements of Service Level Agreement or Quality of Service requirements. . Therefore, the inventors have proposed a solution to the aforementioned problems by using a neural network algorithm and the operation of the software-defined storage system.

According to an aspect of the present invention, an adaptive rapid response control system for a software defined storage system to improve performance parameters includes: a flow monitoring module for obtaining an observation value of a performance parameter at a storage node; An adaptive dual neural module for learning the storage node from a historical record of the storage device configuration and related observation values between the observed value and a specific value of the performance parameter and different difference values The optimal configuration of the plurality of storage devices, and providing the optimal configuration when an existing difference value is not less than a threshold value; and a rapid response control module, if the existing difference value is not less than the threshold value, To change the existing configuration of the storage device in the storage node, the optimal configuration of the storage device provided by the adaptive dual neural module. The storage node is operated by software defined by the software, and the existing difference value will be reduced after the optimal configuration is adopted.

The adaptive dual neural module includes: a certain neural network component for providing the optimal configuration when the existing difference value is not less than a tolerance value, and the optimal configurations are preset to the adaptive fast response Before the control system is operated; and an adaptive neural network component for learning the storage device in the storage node from the historical record of the storage device configuration and the observation value associated with a long period of time under different difference values Optimal configuration, and provides the best configuration when the existing difference value is less than the tolerance but not less than the threshold.

According to the present invention, the adaptive neural network element ceases to function when the fixed neural network component operates, or the fixed neural network component ceases to function when the adaptive neural network component operates. The tolerance value is less than or equal to a preset value, and the preset value is 3 seconds. The long period range is from tens of seconds to the entire history period, and the observation value is not continuously recorded in the long period. The amount of variation between the optimal configuration provided by the fixed neural network component and the existing configuration is greater than the amount of variation between the optimal configuration provided by the adaptive neural network component and the existing configuration. The optimal configuration for learning the storage device is achieved by a neural network algorithm that is a service level protocol or a quality of service requirement. The performance parameter is the number of input/output operations per second, delay time, or throughput. The storage device is a hard disk, a solid state hard disk, a random access memory or a hybrid combination thereof, and the optimal configuration is a percentage of different types of storage devices or a fixed number of single type storage devices.

The adaptive rapid response control system further includes a calculation module for calculating the difference value and transmitting the calculated difference value to the adaptive double Neural module and rapid response control module. The flow monitoring module, the adaptive dual neural module, the rapid response control module or the computing module are hardware or software executed on at least one of the storage nodes.

Implemented by the above hardware, the system can automatically learn how to reconfigure its storage device to meet the requirements of service level agreements or service quality requirements.

10‧‧‧Adaptive rapid response control system

100‧‧‧ storage node

102‧‧‧Management Server

104‧‧‧hard disk

106‧‧‧ Solid State Drive

120‧‧‧Flow Monitoring Module

140‧‧‧Computation Module

160‧‧‧Adaptive dual nerve module

162‧‧‧Definite neural network components

164‧‧‧Adaptive Neural Network Components

180‧‧‧Quick Reaction Control Module

Figure 1 illustrates a block diagram of an adaptive fast response control system in accordance with an embodiment of the present invention.

Figure 2 shows the architecture of a storage node.

Figure 3 is a flow chart of the operation of the adaptive dual neural module.

Figure 4 shows the optimal configuration table provided by the adaptive dual neural module.

The invention will be more specifically described by reference to the following embodiments.

Please refer to FIG. 1 to FIG. 4, which are disclosed herein in accordance with an embodiment of the present invention. 1 is a block diagram of an adaptive fast response control system 10 in accordance with an embodiment of the present invention. The system can improve the performance parameters of a software-defined storage system in a network, such as the number of input/output operations per second, delay time or throughput. In an embodiment, the software definition storage system is a storage node 100, and the delay time for obtaining data from the software definition storage system is illustrated as an example. The The network can be the internet. Thus, the storage node 100 can be a database server that manages a large number of storage devices and provides client cloud services. It can also be a file server or mail server with a dedicated storage device. The network may also be used in a regional network of a laboratory, or a wide area network for a multinational enterprise, and the present invention does not limit the application of the storage node 100. However, storage node 100 must be a software definition store. In other words, the hardware (storage device) of the storage node 100 should be separable from the software managing the storage node 100. The storage node 100 is operated by a software defined by the software definition storage. Therefore, the reconfiguration of the storage device in the storage node 100 can be implemented by a separate software or hardware.

See Figure 2, which shows the architecture of the storage node 100. The storage node 100 includes a management server 102, 10 hard disks 104, and 10 solid state disks 106. The management server 102 can receive instructions to reconfigure the hard disk 104 and the solid state disk 106. The different configurations of the storage node 100, i.e., the percentage of the hard disk 104 and the solid state disk 106 used, can maintain a certain delay time at different workloads. The solid state hard disk 106 has a faster storage speed than the hard disk 104. However, at the same capacity, the price of the solid state drive 106 is much more expensive than the hard disk 104. In other words, at the same cost, the hard disk 104 has a storage capacity ten times that of the solid state disk 106. For such a storage node 100, since the life cycle of the solid state hard disk 106 will drop very fast, it is uneconomical to provide a service in which the solid state hard disk 106 is standby, and when the solid state hard disk 106 is used, the storage capacity Will become a problem. When the configuration of the storage node 100 includes some hard disks 104 and solid state disks 106, as long as the delay time can satisfy the service The storage node 100 can still operate smoothly and avoid the aforementioned problems as required by the Service Level Agreement or the Quality of Service requirements.

The adaptive rapid response control system 10 includes a flow monitoring module 120, a computing module 140, an adaptive dual neural module 160, and a rapid response control module 180. The traffic monitoring module 120 is used to obtain an observation of the delay time of the storage node 100. The calculation module 140 can calculate a difference value between an observation value and a specific value of the delay time, and transmit the calculated difference value to the adaptive dual neural module 160 and the rapid response control module 180. Here, the specific value of the delay time refers to the value required in the service level agreement or the quality of service requirement, which is the longest delay time of the storage node 100, and should be implemented in the service provided under normal use (except for the storage node) 100 when starting up or when a large workload occurs). For the present embodiment, the specific value of the delay time is 2 seconds. Any specific values are possible, and the invention is not limited thereto.

The adaptive dual neural module 160 is used to learn the hard disk 104 and solid state hard in the storage node 100 with different historical values from the historical records and associated observations of the configuration of the hard disk 104 and the solid state hard disk 106. The best configuration of the dish 106. The difference value exists between the observed value and the delay time specific value, and the adaptive dual neural module 160 can also provide an optimal configuration to the rapid response control module 180. The adaptive dual neural module 160 operates when the existing difference value is not less than a threshold. The so-called existing difference value refers to the latest difference value between the observation value from the flow monitoring module 120 and the delay time specific value (2 seconds). The threshold value is a preset timeout for a specific value of the delay time. Since the time beyond the delay time specific value is too short, it is not worth changing the configuration of the hard disk 104 and the solid state hard disk 106 to reduce the delay time, and the existing configuration can continue to operate. In this embodiment, the threshold is 0.2 seconds. Of course, it can be changed for the different services provided by the storage node 100.

In order to achieve the operation provided by the adaptive dual neural module 160, the adaptive dual neural module 160 can further comprise two main components: a certain neural network component 162 and an adaptive neural network component 164. The fixed neural network component 162 provides an optimal configuration that is preset before the adaptive fast response control system 160 operates, and when the existing difference value is not less than the tolerance value, the fixed neural network component 162 is activated. Here, the tolerance value is an extra value other than the specific value of the delay time. Once the tolerance value is perceived, some emergency processing has to be performed, quickly reducing the delay time so that the customer does not have to wait for the reply of the storage node 100 for the next few seconds. The operation of the fixed neural network component 162 can be viewed as a constraint to the extended delay time as the workload increases. In practice, the tolerance value should be less than or equal to a preset value. Preferably, it is less than or equal to 3 seconds. Therefore, the embodiment takes 3 seconds as the tolerance value.

The adaptive neural network component 164 is used to learn the hard magnetic state of the storage node 100 from the historical records of the hard disk 104 and the solid state hard disk 106 configuration and the observations associated with a long period of time with different difference values. The optimal configuration of the disc 104 and the solid state hard drive 106. It also provides the best configuration. Adaptive neural network when the existing difference value is less than the tolerance value but not less than the threshold value The network element 164 operates. The aforementioned long period can be as short as several tens of seconds up to the entire history of the storage node 100. Any record that can be applied to the adaptive neural network component 164 to learn the optimal configuration of the hard disk 104 and the solid state hard disk, any record of the storage node 100 is feasible. It is best to use the records of the entire history period. It will be appreciated that during this long period, certain observations are not continuously recorded and some data may be lost, but adaptive neural network component 164 can still use these discrete records.

Because of the complexity of the storage node 100 hardware and the different workloads generated by the customer's needs, the storage node 100 will have different delay times. The delay time and the workload have no specific relationship with time. For the adaptive rapid response control system 10, the best way to have a management method for the storage node 100 is to learn the change relationship between itself by itself. Therefore, neural network algorithms are a good way to achieve this goal. Learning the optimal configuration of hard disk 104 and solid state hard disk 106 can be achieved by neural network algorithms. Although there are many neural network algorithms available today, the invention is not limited to which one to use. In each algorithm's mode, the parameter settings of different layers can be established using the experience of other systems.

In order to know how the adaptive dual neural module 160 operates, please refer to FIG. 3, which is a flow chart of the operation of the adaptive dual neural module 160. After the observation value of the delay time is obtained by the flow monitoring module 120 (S01), and the calculation module 140 calculates the existing difference value of the delay time (S02), the adaptive dual neural module 160 determines whether the existing difference is present. The value is not less than the threshold, 0.2 seconds (S03). If not, the existing configuration of the hard disk 104 and the solid state hard disk 106 remains unchanged (S04); if so, the adaptive dual neural module 160 will determine whether the existing difference value is not less than the tolerance value, 3 seconds (S05) ). If not, the adaptive neural network component 164 operates (S06); if so, the fixed neural network component 162 operates (S07). It will be apparent that when the fixed neural network component 162 is operational, the adaptive neural network component 164 ceases to function; when the adaptive neural network component 164 is operational, the fixed neural network component 162 ceases to function.

If the existing difference value is not less than the threshold value, the rapid response control module 180 can change the existing configuration of the hard disk 104 and the solid state hard disk 106 in the storage node 100 to become the hard disk 104 provided by the adaptive dual neural module 160. The best configuration with solid state hard disk 106. Therefore, the rapid response control module 180 can always use the optimal configuration provided by the adaptive dual neural module 160 to adjust the configuration of the storage node 100. Existing difference values become less after the optimal configuration is adopted.

Please refer to FIG. 4, which is the optimal configuration table provided by the adaptive dual neural module 160 of the present embodiment. When the storage node 100 is operating with a latency of less than 2 seconds, the configuration includes 50% hard disk 104 and 50% solid state disk 106. Even if the delay time difference value is within 0.2 seconds (delay time is 2.2 seconds), since the delay time difference value is still less than the threshold value, the adaptive dual neural module 160 will not operate and the configuration remains intact. When the delay time difference value increases by more than 0.2 seconds, the adaptive neural network component 164 operates to learn the optimal configuration of the hard disk 104 and the solid state hard disk 106 with historical records and some newly accepted data, the aforementioned new acceptance. The information will be considered historical records for study. At the same time, based on the results of past learning, when the delay time difference value is not less than 0.2 seconds but less than 0.5 seconds, the adaptive neural network component 164 provides the optimal configuration of the fast response control module 180 to 40% of the hard disks 104 and 60. % solid state hard disk 106; when the delay time difference value is not less than 0.5 second but less than 1.0 second, the optimal configuration is 30% hard disk and 70% solid state hard disk 106; when the delay time difference value is not less than 1.0 second However, when less than 3.0 seconds, d is optimally configured as a 20% hard disk 104 and an 80% solid state disk 106. Of course, because the customer's behavior patterns may change in the future, the best configuration can be further changed from the existing history. After the new optimal configuration is applied at different delay time difference values, the delay time will quickly become less than the specific value, 2 seconds. It should be noted that for the optimal configuration, the full part order is not limited to the above six, and may be more than 6 or less than 6. For example, the number of steps of the delay time difference value may be five between the threshold value and the tolerance value. In other words, every 0.5 seconds is divided into a step. In this way, in the present embodiment, the number of all partial steps becomes eight instead of six. This is because the optimal configuration learned by the adaptive dual neural module 160 depends on the type of application (ie, workload) requirements and the hardware specifications of the hard disk and solid state disk in the storage node 100.

When the delay time difference value is not less than the tolerance value, the modest adjustment of the configuration is too late. In this case, a mandatory means should be implemented to quickly reduce the delay time. Thus, the fixed neural network component 162 operates and the adaptive neural network component 164 ceases to function. The fixed neural network component 162 will provide a predetermined optimal configuration for the hard disk 104 and the solid state hard disk 106. According to this implementation For example, when the delay time difference value is not less than 3.0 seconds but less than 5.0 seconds, the optimal configuration is 10% of the hard disk 104 and 90% of the solid state hard disk 106; when the delay time difference value is not less than 5.0 seconds, the best It is configured as a 0% hard disk 104 and a 100% solid state disk 106. In this extreme case, all solid state drives 106 are used.

However, while the fixed neural network component 162 and the adaptive neural network component 164 can provide an optimal configuration, as seen in Figure 4, the optimal configuration provided by the fixed neural network component 162 is compatible with the existing configuration (50% hard The amount of variation between disk 104 and 50% solid state hard disk 106) is greater than the amount of variation between the optimal configuration provided by adaptive neural network component 164 and the existing configuration.

As mentioned above, the delay time is only one of the performance parameters required by the service level protocol. Other performance parameters can be changed in the same manner to adjust the configuration of the hard disk 104 and solid state hard disk 106. For example, the number of input and output operations and throughput per second can increase as the solid state hard disk 106 increases.

It should be emphasized that the storage device is not limited to hard disks and solid state disks, and random access memory can also be used. Thus, a hard disk and a random access memory, or a mashup of a solid state hard disk and a random access memory can also be applied. The optimal configuration in the embodiment is the percentage of different types of storage devices in use. It can also be a fixed number of single-type storage devices (eg, storage nodes contain only solid-state drives, and reconfiguration is accomplished by adding new or standby solid-state drives). Most importantly, the traffic monitoring module 120, the computing module 140, the adaptive dual neural module 160, and the rapid response control module 180 can be implemented by hardware or by at least one processor in the storage node 100. .

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the invention, and those skilled in the art can make some modifications and refinements without departing from the spirit and scope of the invention. The scope of the invention is defined by the scope of the appended claims.

Claims (16)

An adaptive rapid response control system for a software defined storage system to improve performance parameters, comprising: a flow monitoring module for obtaining an observation value of a performance parameter at a storage node; an adaptive dual neural module, For determining whether to provide a preset optimal configuration or an adaptive optimal configuration, when the existing difference value is not less than a tolerance value, providing the preset optimal configuration, and when the existing difference value is less than a tolerance The adaptive optimal configuration is provided when the value is not less than a threshold, wherein the adaptive optimal configuration is stored from the difference between the observed value and a specific value of the performance parameter. Learning from the historical record of the device configuration and related observations; and a rapid response control module for changing the existing storage device in the storage node if the existing difference value is not less than the threshold value The configuration is an optimal configuration of the storage device provided by the adaptive dual neural module, wherein the storage node is operated by a software defined by the software, and is used in the optimal configuration. The existing difference value will decrease. The adaptive rapid response control system according to claim 1, wherein the adaptive dual neural module comprises: a certain neural network component, when the existing difference value is not less than the tolerance value, Providing the preset optimal configuration, and the preset optimal configuration is preset before the operation of the adaptive rapid response control system; and an adaptive neural network component for using different difference values from the The historical record of the storage device configuration and the observation value associated with a long period of time, learning the adaptive optimal configuration of the storage device in the storage node, and when the existing difference value is less than the tolerance value but not less than the threshold value Provides this adaptive optimal configuration. The adaptive rapid response control system of claim 2, wherein the adaptive neural network component stops operating when the fixed neural network component operates, or when the adaptive neural network component operates The neural network component stops working. The adaptive rapid response control system of claim 2, wherein the tolerance value is less than or equal to a predetermined value. The adaptive rapid response control system of claim 4, wherein the preset value is 3 seconds. The adaptive rapid response control system of claim 2, wherein the long period ranges from tens of seconds to the entire history period. The adaptive rapid response control system of claim 2, wherein the observation is not continuously recorded during the long period. The adaptive rapid response control system of claim 2, wherein the amount of change between the preset optimal configuration and the existing configuration provided by the fixed neural network component is greater than the adaptive neural network Component mention The amount of change between this adaptive optimal configuration and the existing configuration. The adaptive rapid response control system of claim 2, wherein the adaptive optimal configuration for learning the storage device is achieved by a neural network algorithm. The adaptive rapid response control system of claim 1, wherein the specific value is a service level agreement or a quality of service requirement. The adaptive rapid response control system of claim 1, wherein the performance parameter is an input/output operation number per second, a delay time, or a throughput. The adaptive rapid response control system of claim 1, wherein the storage device is a hard disk, a solid state hard disk, or a combination thereof. The adaptive rapid response control system of claim 1, wherein the preset optimal configuration and the adaptive optimal configuration are a percentage of different types of storage devices or a fixed number used by a single type storage device. The adaptive rapid response control system of claim 1, further comprising a calculation module for calculating the difference value and transmitting the calculated difference value to the adaptive dual neural module and the rapid response control Module. The adaptive rapid response control system according to claim 1, wherein the flow monitoring module, the adaptive dual neural module, and the rapid response The control module or the computing module is a hardware. The adaptive rapid response control system of claim 1, wherein the flow monitoring module, the adaptive dual neural module, the rapid response control module, or the computing module is at least one of the storage nodes Software executed on one processor.