CN111538603B

CN111538603B - Machine learning and RDMA accelerated key value storage read access method and system

Info

Publication number: CN111538603B
Application number: CN202010311855.1A
Authority: CN
Inventors: 陈榕; 魏星达; 陈海波; 臧斌宇
Original assignee: Shanghai Jiaotong University
Current assignee: Shanghai Jiaotong University
Priority date: 2020-04-20
Filing date: 2020-04-20
Publication date: 2022-03-22
Anticipated expiration: 2040-04-20
Also published as: CN111538603A

Abstract

The invention provides a machine learning and RDMA accelerated key value storage and read access method and a system, comprising the following steps: step S1: predicting the storage position of the key value of the server end by adopting a machine learning model, and enabling the client to obtain a learned model of the server end through remote process call; step S2: when a client processes a key value request of a user, the client firstly predicts an address corresponding to a key of the key value request through an obtained machine learning model, and finally reads back a value corresponding to the requested key through RDMA unidirectional operation. The invention adopts RDMA one-way (one-sided) operation to bypass the server CPU to process the key value request, thereby effectively reducing the utilization of the server processor.

Description

Machine learning and RDMA accelerated key value storage read access method and system

Technical Field

The invention relates to the technical field of networks, in particular to a key value storage and read access method and a key value storage and read access system for machine learning and RDMA acceleration. And more particularly, to a method and system for network-based memory key-value store read access using machine learning and RDMA for high throughput, low latency, low CPU utilization.

Background

Network-based memory key storage access is important for modern network applications, such as speeding up the access performance of traditional databases. Since accesses to the key-value store are all on the critical path of these applications, the performance of key-value store accesses is of great importance. At the same time, these key-value storage systems require low processor utilization; to reduce the power consumption of the system and to reduce the performance impact of other applications on the key-value storage system.

With the popularization of Remote Direct Memory Access (RDMA), an emerging network technology in data centers, people begin to utilize RDMA to implement key-value storage systems with low CPU utilization. In particular, RDMA provides a new primitive (one-sided) so that the client can read and write the server-side memory by bypassing the server-side CPU with one-way (one-sided) operation. Therefore, the one-way primitive is used for realizing the key value storage access, the client can completely bypass the processor of the server, and the low CPU utilization rate is achieved.

However, despite the high throughput and low latency characteristics of RDMA, key-value storage using RDMA does not generally fully exploit the high performance of RDMA. This is because RDMA one-sided operations provide only limited semantics, such that RDMA one-sided (one-sided) based key-value store operations require multiple RDMA operations to complete. This is because operations of the Index (Index) of a conventional key-value store, such as B + Tree, cannot be completed using one RDMA operation. Thus, even if a single RDMA operation has very high performance, there is very much overhead to complete one key request using multiple RDMA requests.

The present invention is based on the observation that: the machine learning (machine learning) based model can predict the stored location from the key (key) of the key-value pair (key-value pair). In this way, we can utilize machine learning models to simplify the indexing operations of key-value stores; this allows us to index the storage with fewer RDMA operations.

Although there is a key value storage using a machine learning model to accelerate a single machine, how to efficiently utilize RDMA one-sided operations to accelerate a network-based key value storage is still a problem to be solved.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a machine learning and RDMA accelerated key value storage read access method and a system.

The invention provides a machine learning and RDMA accelerated key value storage read access method, which comprises the following steps:

step S1: predicting the storage position of the key value of the server end by adopting a machine learning model, and enabling the client to obtain a learned model of the server end through remote process call;

step S2: when a client processes a key value request of a user, the client firstly predicts an address corresponding to a key of the key value request through an obtained machine learning model, and finally reads back a value corresponding to the requested key through RDMA unidirectional operation.

Preferably, the step S2 includes:

step S201: the client judges whether the machine learning model of the server is received: if not, go to step S206; otherwise, go to step S202;

step S202: enabling the client to predict the range of the position of the value corresponding to the key of the key value reading request at the server by using a machine learning model, and executing the step S203;

step S203: enabling the client to retrieve the key of the corresponding range and the address of the value corresponding to the key by using one RDMA unidirectional reading operation according to the range predicted in the step S202, and executing a step S204;

step S204: having the client search for the key among the retrieved keys that is required by the user request: if the required key is found, step S205 is executed; otherwise, executing step S207;

step S205: the client side uses one RDMA unidirectional reading operation to retrieve the value corresponding to the value address of the found key and returns the value to the user;

step S206: enabling the client to send a request to the server to acquire the machine learning model which is constructed by the server and used for predicting the key value position; subsequently, step S207 is performed;

step S207: and enabling the client to send a key value request for requesting to execute the user to the server, and returning the value returned by the server to the user.

Preferably, the machine learning model at the server end in step S201 is a neural network model or a linear regression model.

Preferably, in step S202, the output of the machine learning model is used as the prediction result by inputting the user' S key into the machine learning model.

Preferably, in the step S202, the server side uses an ordered array to store all key-value data.

Preferably, the key and the address of the value corresponding to the key in step S202 are stored in an ordered array.

Preferably, the range of the position in step S202 refers to a subscript range of the server-side ordered array.

Preferably, the key retrieved in step S204 refers to the key corresponding to the ordered array of servers in the subscript range in step S202.

Preferably, in step S206, the client sends the request to the server using a remote procedure call.

Preferably, in step S207, the client sends the request to the server using a remote procedure call.

The invention provides a key value storage read access system utilizing machine learning and RDMA acceleration, which comprises:

module S1: predicting the storage position of the key value of the server end by adopting a machine learning model, and enabling the client to obtain a learned model of the server end through remote process call;

module S2: when a client processes a key value request of a user, the client firstly predicts an address corresponding to a key of the key value request through an acquired machine learning model, and finally reads back a value corresponding to the requested key through RDMA unidirectional operation;

the module S2 includes:

a module S201: the client judges whether the machine learning model of the server is received: if not, executing module S206; otherwise, executing the module S202;

a module S202: enabling the client to predict the range of the position of the value corresponding to the key of the key value reading request at the server by using a machine learning model, and calling the module S203;

a module S203: the client retrieves the key of the corresponding range and the address of the value corresponding to the key by using one RDMA unidirectional read operation according to the range predicted by the module S202, and calls the module S204;

a module S204: having the client search for the key among the retrieved keys that is required by the user request: if the required key is found, calling the module S205; otherwise, calling the module S207;

the module S205: the client side uses one RDMA unidirectional reading operation to retrieve the value corresponding to the value address of the found key and returns the value to the user;

a module S206: the client sends a request to the server to acquire the machine learning model which is constructed by the server and used for predicting the key value position, and then the module S207 is called;

a module S207: the client sends a key value request for requesting to execute the user to the server, and returns a value returned by the server to the user;

the machine learning model of the server side in the module S201 is a neural network model or a linear regression model;

in the module S202, the output of the machine learning model is used as a prediction result by inputting the user' S key into the machine learning model;

the server side in the module S202 uses an ordered array to store all key value data;

the key in the module S202 and the address of the value corresponding to the key are stored in an ordered array;

the range of the position in the module S202 refers to a subscript range of the ordered array at the server end;

the key retrieved in the module S204 refers to the key to which the ordered array of servers corresponds in the range of subscripts in the module S202;

in the module S206, the client sends a request to the server by using a remote procedure call;

in the module S207, the client sends a request to the server using a remote procedure call.

Compared with the prior art, the invention has the following beneficial effects:

1. the invention adopts RDMA one-way (one-sided) operation to bypass the server CPU to process the key value request, thereby effectively reducing the utilization of the server processor.

2. The client of the invention only uses two RDMA operations to execute one key value request, and has good performance.

3. The machine learning model used by the invention occupies little memory at the client, so the system has good expansibility.

Drawings

Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:

FIG. 1 is a flow chart of exemplary steps for a client to perform a key-value operation using a machine learning model and RDMA operations as provided by the present invention.

Fig. 2 is a hardware device architecture diagram of the system provided by the present invention.

Fig. 3 is a software device architecture diagram of the system provided by the present invention.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.

Specifically, the step S2 includes:

Specifically, the machine learning model at the server end in step S201 is a neural network model or a linear regression model.

Specifically, the step S202 takes the output of the machine learning model as the prediction result by inputting the user' S key into the machine learning model.

Specifically, in the step S202, the server side uses the ordered array to store all key value data.

Specifically, the key and the address of the value corresponding to the key in step S202 are stored in the ordered array.

Specifically, the range of the position in step S202 refers to a subscript range of the server-side ordered array.

Specifically, the key retrieved in step S204 refers to the key corresponding to the ordered array of servers in the subscript range in step S202.

Specifically, in step S206, the client sends a request to the server using a remote procedure call.

Specifically, in step S207, the client sends a request to the server using a remote procedure call.

the module S2 includes:

The present invention will be described more specifically below with reference to preferred examples.

Preferred example 1:

the invention will be further explained with reference to the drawings

As shown in fig. 1, a specific process for processing a key value search request by a client according to the present invention is as follows:

in step 1, the client queries whether the machine learning model is acquired from the server, if not, step 8 is executed to acquire the machine learning model from the server, and the client uses Remote Procedure Call (RPC) to initiate a request to the server. How it has been acquired, step 2 is performed. While executing step 8, the client also executes step 9, and initiates a request to the server to execute a search operation of a key value request; step 9 is also performed using RPC. Step 8 and step 9 initiate requests concurrently at the client.

After the step 9 is executed, the client returns the result returned by the server to the user, and the execution is finished.

In step 2, if the client finds the machine learning model sent by the server, the client executes a search request by using the machine learning model, and executes step 3;

in step 3, the client inputs the key in the key value request of the user into the machine learning model, and the machine learning model returns a section of range in the ordered array of the server according to the key; for example, if the range is [0-1], then in the example of FIG. 3, the range refers to the 0 th through 1 st elements of the server-side ordered array;

in step 4, the client reads all the (key, value address) of the range memory predicted in step 3 back to the memory of the client by using an RDMA one-way (one-sided) read operation; for example, in the example of FIG. 3, if this range is [0-1], the client will read back (key 0, value address 0), (key 1, value address 1) to the client in an RDMA one-way (one-sided) read operation;

in step 5, the client searches all the read-back keys (keys, value addresses), and if the keys requested by the user exist in the keys (step 6), step 7 is executed to read back the values corresponding to the keys requested by the user to the local; otherwise, executing step 9, and returning to the mode of using RPC to execute the key value request; for example, in the example of fig. 3, if the key requested by the user is key 0, the client finds the corresponding (key 0, value address 0);

in step 7, the user reads the content of the value in the value back to the client by using one RDMA read operation according to the address of the value corresponding to the key, returns the value to the user, and then ends the key value search request. In the example of fig. 3, if the key requested by the user is key 0, the client reads the value 0 and returns to the user according to the value address 0 found in step 6.

As shown in fig. 2, the present invention is based on a conventional client/server (CS) architecture: the server side is responsible for storing the key value database, and the client side is used for serving key value requests of users through interaction with the server side. The client and the server are conventional computers. The client and the server form a local area network. Unlike the conventional CS, the client is connected to the server side through the RDMA network.

The present invention stores additional data at the client while using a machine learning model at the server to predict the location of the value corresponding to the key. As shown in fig. 3, the client needs to additionally store the machine learning model obtained from the server. The server side stores all key values by adopting an ordered array. In the ordered array, the key value is stored in a (key, value address) manner, where the value address points to an address in the memory where the value is stored.

Preferred example 2:

the purpose of the invention can be realized by the following technical scheme:

a machine learning and RDMA accelerated key value storage read access method, the method uses the machine learning model to predict the position of the key value storage of the server, the client obtains the learned model of the server through the remote process call; when a client processes a key value request of a user, the client firstly predicts an address corresponding to a key of the key value request through an acquired machine learning model, and finally reads back a value corresponding to the requested key through RDMA one-way (one-sided) operation; in a client executing a key value request, the method is characterized by comprising the following steps:

(1) the client judges whether the machine learning model of the server is received or not, if not, the step (6) is executed, otherwise, the step (2) is executed;

(2) the client predicts the range of the position of the value corresponding to the key of the key value reading request at the server by using a machine learning model, and executes the step (3); the client uses a machine learning model to predict the same as a traditional machine learning model, the machine learning model can be abstracted into a mathematical function (f), and a key (key) is input into f to obtain f (key), namely the predicted position;

(3) the client retrieves the key of the corresponding range and the address of the value corresponding to the key by using one-sided read operation according to the range predicted in the step (2), and executes the step (4);

(4) the client searches for the key required by the user request in the keys retrieved in the step (4); if so, executing the step (5), otherwise, executing the step (7);

(5) and (3) the client retrieves the value corresponding to the value address in the step (5) by using one-time RDMA unidirectional (one-sided) read operation and returns the value to the user.

(6) The client sends a request to the server to acquire the machine learning model which is constructed by the server and used for predicting the key value position; subsequently performing step (7);

(7) the client sends a key value request for requesting to execute the user to the server, and returns the value returned by the server to the user.

The machine learning model of the server side in the step (1) can be a neural network model or a linear regression model.

In the step (2), the server side uses the ordered array to store all key value data.

And (3) the key in the step (2) and the address of the value corresponding to the key are also stored in the ordered array together with the key.

The position range in the step (2) refers to a subscript range of the ordered array at the server side.

The key retrieved in step (4) refers to the key corresponding to the ordered array of the server in the subscript range in step (2).

And (6) the client sends a request to the server by using the remote procedure call.

And (7) the client sends a request to the server by using the remote procedure call.

A key value storage and read access method for machine learning and RDMA acceleration is characterized in that the key value storage adopts a client/server (CS) mode, and computers where the client and the server are located are connected by an RDMA network card to form a local area network; the client uses a machine learning model of the server.

In the description of the present application, it is to be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate orientations or positional relationships based on those shown in the drawings, and are only for convenience in describing the present application and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation, be constructed in a specific orientation, and be operated, and thus, should not be construed as limiting the present application.

Those skilled in the art will appreciate that, in addition to implementing the systems, apparatus, and various modules thereof provided by the present invention in purely computer readable program code, the same procedures can be implemented entirely by logically programming method steps such that the systems, apparatus, and various modules thereof are provided in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system, the device and the modules thereof provided by the present invention can be considered as a hardware component, and the modules included in the system, the device and the modules thereof for implementing various programs can also be considered as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.

The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims

1. A machine learning and RDMA accelerated key-value store read access method, comprising:

step S2: when a client processes a key value request of a user, the client firstly predicts an address corresponding to a key of the key value request through an acquired machine learning model, and finally reads back a value corresponding to the requested key through RDMA unidirectional operation;

the step S2 includes:

step S207: the client sends a key value request for requesting to execute the user to the server, and returns a value returned by the server to the user;

the machine learning model of the server side in the step S201 is a neural network model or a linear regression model;

in the step S202, the output of the machine learning model is used as a prediction result by inputting the user' S key into the machine learning model;

in the step S202, the server side stores all key value data by using an ordered array;

the key and the address of the value corresponding to the key in the step S202 are stored in the ordered array;

the range of the position in step S202 refers to a subscript range of the ordered array at the server;

the key retrieved in step S204 refers to the key corresponding to the ordered array of servers in the subscript range in step S202;

in step S206, the client sends a request to the server using a remote procedure call;

in step S207, the client sends a request to the server using the remote procedure call.

2. A key-value store read access system utilizing machine learning and RDMA acceleration, comprising:

the module S2 includes: