CN102681889A - Scheduling method of cloud computing open platform - Google Patents

Abstract

The invention discloses a scheduling method of a cloud computing open platform, which particularly characterized in that call requests of mass users are monitored through maintaining a central server; business service and data service of a cluster server are dynamically scheduled; and meanwhile, detachable service components are called and a multithread ability of a multi-core processor is scheduled and distributed according to use levels of the users. Through using the method, the defects that the traditional open platform is short of expansibility and flexibility can be effectively corrected, two targets of rapidly constructing and deploying an application and operation environments and dynamically adjusting the application and operation environments are established, the existing equipment is maximally utilized and the service is maximally constructed; and meanwhile, through adopting a multi-core computing manner and a multithread technology, the scheduling speed and the response ability of the system service are improved.

Description

A kind of dispatching method of cloud computing open platform

Technical field

The invention belongs to computerized information analysis and technical field of data processing, be specifically related to a kind of dispatching method of cloud computing open platform.

Background technology

Cloud computing is one of focus of current domestic and international commerce and scientific research institution's research, is the development of grid computing, parallel computation, Distributed Calculation, is a kind of emerging commercial computation schema.It has adopted ripe Intel Virtualization Technology with the user that the resource of data center is packaged as on the internet on-demand service to be provided.Job scheduling and resources allocation are two gordian techniquies of cloud computing, and the commerciality of cloud computing is paid close attention to quality of services for users it, and its Intel Virtualization Technology makes resources allocation and job scheduling parallel distributed different from the past calculate.In grid computing, in Distributed Calculation, the parallel computation, job scheduling and resources allocation are the contents of broad research.Efficient is paid attention in traditional job scheduling, resources allocation research, and the fairness of resources allocation also is the load balancing property that is related to QoS of customer, system, the importance that task is accomplished efficient.Especially cloud computing is a kind of commercial realization, and purpose is need pay attention to the satisfaction of user's request more for different users provides service, computing power, storage etc.

Also has the Web open platform with cloud computing together develops.As a kind of new network service mode; The Web open platform at first provides a basic service; Interface through open itself then; Help third party developer through utilization with assemble its interface and other third party's service interfaces produce new application, guarantee that this application can unify to operate on this open platform; When the user uses open platform, can be more more intensive completion multiple network activities, platform provides various services and relevant guarantee for the user.The basic service of open platform can be existing, and for example door, blog also can be emerging, for example customer relationship.

1. multi-threaded parallel and mutual exclusion techniques

Each program of just in system, moving all is a process.Each process comprises one or more threads.Process also possibly be the dynamic execution of whole procedure or subprogram.Thread is the set of one group of instruction; Or the particular segment of program; It can independently be carried out in program, also can be interpreted as it the context of code operation, so thread is the process of lightweight basically; It is responsible in single program, carrying out multitask, is responsible for the scheduling and the execution of a plurality of threads usually by operating system.

When a plurality of thread parallels move, adopt thread pool that multithreading is managed.Thread pool is a kind of multithreading processing form, adds task to formation in the processing procedure, after creating thread, starts these tasks automatically then, and the thread pool thread all is a background thread.Each thread all uses the storehouse size of acquiescence, with the priority operation of acquiescence, and is in the multiple thread units.If certain thread is idle (as waiting for certain incident) in Managed Code, then thread pool will insert another worker thread and make all processors maintenances busy.If it is busy that all thread pool threads all remain, but comprise the work of hang-up in the formation, then thread pool will be created another worker thread over time, but the number of thread surpasses maximal value never.Surpassing peaked thread can line up, but they just start after will waiting until other thread completion.

Under multi-thread environment, adopt semaphore to accomplish the mutual exclusion of cross-thread with synchronously.Semaphore is the total data structure of under multi-thread environment, using of a kind of overall situation, is to be used for guaranteeing that two or more critical code sections are not by concurrent invocation.Before getting into a critical code section, thread must obtain a semaphore; In case this critical code section has been accomplished, this thread must release semaphore so, and other wants that the thread that gets into this critical code section must wait for up to first thread release semaphore.In order to accomplish this process; Need to create a semaphore; To obtain semaphore (Acquire Semaphore) and release semaphore (Release Semaphore) then and be placed on the first terminal of each critical code section respectively, what confirm that these semaphores quote is the semaphore of initial creation.

2. multiprocessor distributes and load-balancing technique

Multiprocessor Task Distribution and load balancing are under the environment of multiple processor cores, and Processing tasks is carried out dynamic assignment and debugging, thereby improve processor utilization, realize the dispatching response of multi-threaded parallel.

The scheduling model of multicomputer system mainly comprises processor model, task model and dispatching algorithm three parts.The processor model is described information such as processor structure and processing power, and task model is described scheduling and waited for the needed relevant information of task for processing.If a multicomputer system is by m processor (P ₁, P ₂..., P _m) and K resource (Res=r ₁, r ₂..., r _k), processing power C _i(i=1,2 ..., m) expression processor P _iThe ability of Processing tasks in the unit interval, the processing power TPC of multicomputer system is defined as: $\mathrm{TPC}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{C}_i.$

Multicomputer system adopts centralized scheduling mechanism, by scheduler all Task Distribution is carried out to each processor; Each processor all has scheduling queue separately, and processor selection task from this scheduling queue is handled; Communicating by letter between scheduler and each processor realizes through scheduling queue.

To multicomputer system, use to be scheduled to power, average response time, processor utilization, scheduling length as the performance index of estimating dispatching algorithm.

1) be scheduled to power (Success Rate, SR): refer to counted the ratio that N ' and the task-set that is scheduled are counted N by the successful task-set of algorithmic dispatching: $\mathrm{SR}=\frac{N^{\′}}{N}.$

2) average response time (Average Response Time, ART): refer to concentrate time that all tasks begin to handle and the mean difference of time of arrival.If task S _iTime of arrival be t _i, the time that begins to handle is st _i, then its response time is rt _i=st _i-t _iIf the number of tasks in the task-set is n; Then the average response time of task is:

the average response time ART of a task-set more little; Be that the stand-by period is short more, then the performance of dispatching algorithm is good more.

3) scheduling length (Schedule Length, SL): refer to from task time of arrival the earliest poor to deadline of a last task.Deadline (finish time) is ft (i)=st (i)+pt (i), and wherein, pt (i) is the processing time of task st (i), then dispatches length and is: SL=max (ft (i))-min (t (i)).

Dispatching under the successful prerequisite, less scheduling length means that task is early disposed, and the average utilization of processor is higher.

4) processor utilization (Utilization Rate, UR): finger processor execute the task time of taking and with the scheduling length ratio, establishing among the task-set S has S _jIndividual task is processed device P _j(j=1,2 ..., m) response, then P _jUtilization factor be:

And the average utilization of all processors (Average Utilization Rate AUR) then can be expressed as:

Generally speaking, the cloud computing task scheduling system is similar to the gridding scheduling system, according to state of resources information and prediction, according to certain scheduling strategy user's set of tasks is mapped on the resource collection and carries out, and operation result is returned to the user.Though said method can utilize multiserver to carry out task and handle, and does not add the ruuning situation and the performance parameter of each server, because can not make full use of the handling property of individual server.A dispatching method can not make These parameters reach optimum simultaneously, so should be scheduled under the prerequisite of power in assurance, improves other indexs as far as possible, so that the overall performance of system is optimized.

Summary of the invention

In order to overcome the shortcoming that exists on existing dispatching algorithm and the treatment effeciency, the present invention proposes a kind of dispatching method of cloud computing open platform.

Technical scheme of the present invention is: a kind of dispatching method of cloud computing open platform comprises the steps:

The first step, central server are constructed some parallel multithreadings, and the structure process pool, put into process pool to the process of structure afterwards;

Second step, registration central server are created and are started finger daemon, and by the main thread initializes global variable, global variable should comprise semaphore at least, and sub-thread state collection and result set are used to carry out sub-thread synchronization and record queries record;

The 3rd step, operate in the connection of waiting for front end web acting server process behind the load equalizer initialize communications port in the central server; When the server connection request arrives; Load equalizer generates a thread and this server communication; Load equalizer continue to be waited for the connection request of other server, when a new client requests service, load equalizer from information table, selects load minimum be that load weights maximum service device is its service;

Request in the 4th step, the central server processing threads pond, and the service of calling application server, according to the multi-core parallel concurrent processor of user right calling application server cluster, completion process and thread are to the distribution of processor node;

The data that the 5th step, central server are collected database server are handled distributed data;

The 6th step, central server calls cluster service, and integrates mass data and return to the web acting server and call, and gives back the thread to thread pond of current request simultaneously, continues the request of monitoring users and also from thread pool, calls idle thread.

The above-mentioned first step specifically comprises as follows step by step:

1) be the concurrent threading model of model creation with the FORK-JOIN structure;

2) thread that daemon thread FORK is come out is put into thread pool, is responsible for the life cycle management of thread by thread pool;

3) under the multi-threaded parallel processing environment, adopting variable mutex is the mutex amount, is used to realize the synchronous and mutual exclusion of the resource access under the multi-thread environment; The course of work of this mutex amount is following: when the resource that mutex representes was used in one of request, whether process read the value of mutex earlier, available to judge corresponding resource; When the value of mutex greater than 0 the time, showing has resource to ask; Equal at 0 o'clock, show no available resources, process gets into sleep state when available resources are arranged; When process did not re-use the shared resource of a semaphore control, the value of mutex increased 1;

The load balancing strategy of above-mentioned the 3rd step load balancing implement body is: when the node of cloud computing open platform comes into operation for the first time, set an initial weights SW (N _i), along with the variation of node load, balanced device is adjusted weights, and the parameter dynamic calculation of each side drew when weights were moved by node, node N _iWeights describe with following formula:

SW(N _i)＝K ₀*L_CPU(N _i)+K ₁*L_Memory(N _i)+K ₂*L_Process(N _i)+K ₃*L_IO(N _i)+K ₄*L_Response(N _i)

Wherein, K ₀, K ₁, K ₂, K ₃And K ₄The expression constant coefficient, L_CPU (N _i) be node N _iCPU usage, L_Memory (N _i) be node N _iMemory usage, L_Process (N _i) be node N _iRate of people logging in, L_IO (N _i) be occupation rate, the L_Response (N of the magnetic disc i/o of node Ni _i) be node N _iThe process response time, L_CPU (N _i)=1-P_CPU (N _i), wherein, P_CPU (N _i) utilization factor of expression current C PU.

Above-mentioned the 4th step specifically comprises as follows step by step:

1) foundation of Task Assignment Model:

If application server cluster comprises N _NodeIndividual processing node D ₀, D ₁...,

Concurrent program to be allocated has N _ProcIndividual process P ₀, P ₁..., Process P _iComprise M _iIndividual thread T ₀, T ₁...,

Total thread number of concurrent program $N_{\mathrm{Thread}}=\underset{k=0}{\overset{N_{\mathrm{Proc}-1}}{\mathrm{\Σ}}}{M}_k;$

Concurrent program to be allocated is expressed as a task graph of a relation, and (V, E), wherein, V is the set { V of each application server node in the application server cluster to be specially a non-directed graph G= _i, node V _iCorresponding doublet<t _i, P _i>E is the set { E of nonoriented edge _Ij; Connected node V _iAnd V _jLimit E _Ij∈ E, expression thread T _iAnd T _jBetween communication or shared data, the weights W on limit _IjThe frequent degree of representing two thread communications or shared data;

2) carry out two wheel operations, the distribution of central server response process to cluster server accomplished in first round operation; Second wheel operation is accomplished the distribution of processing server intranodal thread to processor core; Each wheel operation comprises repeatedly iteration; Concrete processing procedure is following: begin from the initiating task graph of a relation of input; Each limit of selecting to have maximum weights, two summits of merging this edge, the number of threads that is comprised in the newly-generated node must be smaller or equal to a threshold value; Repeat this process, the number of node equals unappropriated number of threads in the central server current thread pond in the task graph of a relation; The threshold value that first and second wheel operation uses is calculated by following formula respectively:

${\mathrm{Threshold}}_{\mathrm{first}\_\mathrm{round}}=\mathrm{Max}\left(\right[\frac{N_{\mathrm{thread}}}{N_{\mathrm{proc}}}],M_{\max })\&times;(1+\mathrm{\&alpha;}),0<\mathrm{\&alpha;}<1$

${\mathrm{Threshhold}}_{\sec \mathrm{ond}\_\mathrm{round}}=\left[\frac{{\mathrm{Threashold}}_{\mathrm{first}\_\mathrm{round}}}{N_{\mathrm{core}}}\right]$

Wherein, Threshold _{First_round}The threshold value of the processing node number of expression first round operation, Threshhold _{Second_round}The threshold value of representing the composite node number of second wheel operation, rounding operation in [] expression, M _MaxBe the maximum thread that process has, N _CoreFor existing process number, Max () represent the parameter of relatively importing into and get maximal value, α is a percent value, is used for being illustrated between equally loaded and the minimizing communication weighing; Specifically be divided into as follows step by step:

2a) to first round operation, initial task graph of a relation is that unit carries out preliminary division with each server individuality, the composite node number=process number in promptly scheming, and the corresponding process of each composite node, the thread that is included in the composite node all belongs to this process; The first round termination condition of operation is composite node number≤Threshold among the figure _{First_round}, each composite node is a sub-graphs in scheming during end, corresponding processing node, and the thread that is included in the subgraph should be distributed to this processing node;

Each subgraph that 2b) first round division of operations is gone out all carries out second wheel operation, and promptly the initiating task graph of a relation of second wheel operation is the subgraph that first round operation obtains, and the termination condition of second wheel operation is the process number≤Threshhold that is scheduled in the resource pool among the figure _{Second_round}, corresponding process of each composite node among figure during end, the thread that is included in the thread pool in the central server should be distributed to this processor core;

2c) target ip address of analysis request message and load balancing in view of the above; Request scheduling with same target IP address under the load balance situation of server arrives same node; Be specially: at first find out target ip address most recently used node; If this service node is available and not overload, then user's request is sent to this service node by load-balanced server; If service node does not exist, or this service node overloads and have service node to be in its half the operating load, and then each service node of poll is selected the minimum service node of link, and request is sent to this service node.

Beneficial effect of the present invention: dispatching method of the present invention is monitored the call request of mass user through safeguarding central server; The business service of dynamic dispatching cluster server and data, services are simultaneously called the multithreading ability that serviced component removably and scheduling distribute polycaryon processor according to user's use rank.Pass through said method; Can effectively overcome the defective that the traditional open platform lacks extensibility and retractility; Set up fast construction application deployment running environment and used these two targets of runtime environment, utilized existing equipment with accomplishing maximal efficiency and set up service with dynamically adjusting; Calculate and multithreading through multinuclear simultaneously, improve the schedule speed and the responding ability of system service; Meanwhile provide modularization removably to serve, can isolate user's service of different rights and distribute corresponding calculated ability and processing power for it, support the otherness service between the different user.

Description of drawings

Fig. 1 be dispatching method of the present invention based on the platform synoptic diagram.

Embodiment

Below in conjunction with accompanying drawing and specific embodiment the present invention is done further explanation.

Dispatching method of the present invention based on the platform synoptic diagram as shown in Figure 1, specify as follows:

The web proxy server: the user collects interconnected user on the network's request, and for serving the queue scheduling buffer area is provided.

Central server: this module unified management user's registration, divide to be used in response user requested service device executive process and to be service processes distribution authority and pool service of calling and data simultaneously.

Database server cluster: comprise a plurality of database servers, be used to store the mass data that the user produces that possess high outlet bandwidth, high internal memory is equipped with the high capacity external memory storage simultaneously, supports the characteristics of big data storage.

Application server cluster: comprise a plurality of application servers, be used to deposit each system service mirror image, high outlet bandwidth supplies the central server scheduling.

The inventive method comprises the steps:

The first step, the some multi-threaded parallels of registration center's server constructs, and structure process pool are put into process pool to the process of structure afterwards; The process of structure is appreciated that and is the structure of initialization when the system start-up.

This step can be passed through to realize step by step as follows:

1) be the concurrent threading model of model creation with the FORK-JOIN structure, in the FORK-JOIN models of concurrency, a FORK statement produces a new concurrent thread path, uses the JOIN statement during end.After former thread and the new thread that produces all reached the JOIN statement, code continued to carry out in a sequential manner.More if desired concurrent thread then need be carried out more FORK statement.

2) thread that daemon thread FORK is come out is put into thread pool, is responsible for the life cycle management of thread by thread pool.

3) under the multi-threaded parallel processing environment, adopting variable mutex is the mutex amount, realizes the synchronous and mutual exclusion of the resource access under the multi-thread environment.The principle of work of this mutex amount is following: when the resource that mutex representes was used in one of request, whether process need read the value of mutex earlier, available to judge corresponding resource.When the value of mutex greater than 0 the time, showing has resource to ask.Equal at 0 o'clock, no available resources now are described, process can get into sleep state when available resources are arranged.When process did not re-use the shared resource of semaphore control, the value of mutex increased 1, the value of semaphore is increased and decreased operation be atomic operation, guaranteed the mutual exclusion of maintenance resources or the synchronization of access of multi-process.

Second step, registration center's server are created and are started finger daemon, by the main thread initializes global variable.Global variable should comprise semaphore at least, and sub-thread state collection and result set are used to carry out sub-thread synchronization and record queries record;

Wait for the connection of front-end proxy agent server Proxy Server process behind the 3rd step, the load equalizer LoadBalance initialize communications port.When the server connection request arrives, generate a thread and this server communication.LoadBalance continues to wait for the connection request of other servers.A thread receives and writes down the performance load information of a Web server.Having several Web server node LoadBalance just to generate several threads in the cluster environment communicates with it.In each computation period, the information on services tabulation can accurately reflect the load performance information of all Web servers.When a new client requests service, LoadBalance from information table, selects load minimum be that load weights maximum service device is its service.

Concrete load balancing strategy is:

When using in the first input coefficient of the node of system, set an initial weights SW (Ni), along with the variation of node load, balanced device is adjusted weights, and the parameter dynamic calculation of each side drew when weights were moved by node.In conjunction with the current weights of each node, can calculate the size of the weights that make new advances.Dynamically the purpose of weights is the situations that will correctly reflect the node load, with the in the future possible load variations of prediction node.

In system's operational process, the ratio of each parameter is suitably adjusted for ease, be constant coefficient Ki of each parameter setting, be used for representing the significance level of each load parameter, wherein ∑ Ki=1 to different application.Therefore, the weights formula of any one node Ni just can be described as:

SW(N _i)＝K ₀*L_CPU(N _i)+K ₁*L_Memory(N _i)+K ₂*L_Process(N _i)+K ₃*L_IO(N _i)+K ₄*L_Response(N _i)

Wherein, K ₀, K ₁, K ₂, K ₃And K ₄Represent constant coefficient, be used to distinguish the performance of different model server, specifically can decide K according to the configuration of server ₀The CPU of corresponding with service device, K ₁The internal memory of corresponding with service device, K ₂The network connection process number of corresponding with service device, K ₃The disk I access speed of corresponding with service device, K ₄The process number of corresponding with service device; Above-mentioned constant coefficient is a relative quantity, illustrates, if the CPU frequency of A server is 2GHz, the CPU frequency of B server is 6GHz for double-core 3GHz, then the corresponding K of A server ₀: the K that the B server is corresponding ₀=1: 3; K ₁, K ₂, K ₃, K ₄With K ₀Similar, no longer explanation.Concrete condition according to server is provided with K like this ₀, K ₁, K ₂, K ₃And K ₄Can under the prerequisite that guarantees high utilization rate, reach the maximization dispatching.

L_CPU (Ni) is the CPU usage of node Ni; L_Memory (Ni) is the memory usage of node Ni; L_Process (Ni) is the rate of people logging in of node Ni, and L_IO (Ni) is that occupation rate, the L_Response (Ni) of the magnetic disc i/o of node Ni is the process response time of node Ni, L_CPU (Ni)=1-P_CPU (Ni); Wherein, the utilization factor of P_CPU (Ni) expression current C PU.

Request in the 4th step, the central server processing threads pond, and the service interface of calling application server.Call the multi-core parallel concurrent processor of cluster server according to user right, be used for the distribution of completion process and thread to processor node:

1) foundation of Task Assignment Model:

If server cluster comprises N _NodeIndividual processing node D ₀, D ₁...,

Concurrent program to be allocated has N _ProcIndividual process P ₀, P ₁...,

Process P _iComprise M _iIndividual thread T ₀, T ₁...,

Total thread number of concurrent program

Concurrent program to be allocated can be expressed as a task graph of a relation, and it is that (V, E), wherein V is the set { V of each application server node in the application server cluster to a non-directed graph G= _i, node V _iCorresponding doublet<t _i, P _i>, T wherein _iBe the corresponding thread number of node, the P of front is exactly a process ID number, promptly is used for the pid of identification process in the operating system, uses pid to replace the process of expression instantiationization here; E is the set { E of nonoriented edge _Ij; Connected node V _iAnd V _jLimit E _Ij∈ E, expression thread T _iAnd T _jBetween communication or shared data, the weights W on limit _IjThe frequent degree of representing two thread communications or shared data.

2) carry out two wheel operations, the distribution of central server response process to cluster server accomplished in first round operation; Second wheel operation is accomplished the distribution that the interior thread of child servers processing node is managed device nuclear everywhere.Each wheel operation comprises repeatedly iteration, and its processing procedure is: begin from the initiating task graph of a relation of input, and each limit of selecting to have maximum weights, two summits of merging this edge that is to say, merge two composite nodes and generate a new composite node.In order to satisfy the load balancing condition, the number of threads that is comprised in the newly-generated composite node must be smaller or equal to a threshold value.Repeat this " selection-merging " process, the number of composite node equals the number of threads of the distribution of central server thread pool in the task graph of a relation.The threshold value that first and second wheel operation uses is calculated respectively as follows:

${\mathrm{Threshold}}_{\mathrm{first}\_\mathrm{round}}=\mathrm{Max}\left(\right[\frac{N_{\mathrm{thread}}}{N_{\mathrm{proc}}}],M_{\max })\&times;(1+\mathrm{\&alpha;}),0<\mathrm{\&alpha;}<1$

${\mathrm{Threshhold}}_{\sec \mathrm{ond}\_\mathrm{round}}=\left[\frac{{\mathrm{Threashold}}_{\mathrm{first}\_\mathrm{round}}}{N_{\mathrm{core}}}\right]$

Wherein, Threshold _{First_round}The threshold value of the processing node number of expression first round operation, Threshhold _{Second_round}The threshold value of representing the composite node number of second wheel operation, rounding operation in [] expression, M _MaxBe the maximum thread that process has, N _CoreFor existing process number, Max () represent the parameter of relatively importing into and get maximal value, α is a percent value, is used for being illustrated between equally loaded and the minimizing communication weighing; Specifically be divided into as follows step by step:

2a) to first round operation, initial task graph of a relation is that unit carries out preliminary division with each server individuality, the composite node number=process number in promptly scheming, and the corresponding process of each composite node, the thread that is included in the composite node all belongs to this process; The first round termination condition of operation is composite node number≤Threshold among the figure _{First_round}, each composite node is a sub-graphs in scheming during end, corresponding processing node, and the thread that is included in the subgraph should be distributed to this processing node;

Each subgraph that 2b) first round division of operations is gone out all carries out second wheel operation, and promptly the initiating task graph of a relation of second wheel operation is the subgraph that first round operation obtains, and the termination condition of second wheel operation is the process number≤Threshhold that is scheduled in the resource pool among the figure _{Second_round}, corresponding process of each composite node among figure during end, the thread that is included in the thread pool in the central server should be distributed to this processor core.Here, the thread of the thread pool in central server the inside includes the process of establishment, and process is according to the node threshold value Threshhold that calculates _{Second_round}Distribute.

2c) target ip address of analysis request message and load balancing in view of the above.Request scheduling with same target IP address under the load balance situation of server arrives same node, improves the locality of reference and the main memory Cache hit rate of each service node, thereby improves the processing power of whole group system.The practical implementation step is: at first find out target ip address most recently used node, if this service node is available and not overload, then by load-balanced server user's request is sent to this service node; If service node does not exist, or this service node overloads and have service node to be in its half the operating load, and then each service node of poll is selected the minimum service node of link, and request is sent to this service node.

The data that the 5th step, central server are collected database server are handled distributed data.

Concrete step is following: establish database server cluster and comprise N _NodeSub-processes node D ₀, D ₁...,

The D0 server is the corresponding numbering of each server-assignment D also as each server of central server poll ₀Arrive

To the previous server that is numbered even number, two data in server of merger then are such as merger data transmission for the server that is numbered odd number<d ₀, D ₁>,<d ₂, D ₃>Or the like; The server that obtains data is poll numbering again again, the server that is numbered odd number equally data transmission to the previous server that is numbered even number, two data in server of merger then; By that analogy, move in circles up to all being integrated into D to total data ₀On.

The 6th step, central server call cluster service and service handle, and integrate mass data and return to the web acting server and call.Give back the thread to thread pond of current request simultaneously, continue the request of monitoring users and from thread pool, call idle thread.

Dispatching method of the present invention is through safeguarding that a customer center is the call request that central server is monitored mass user; In the face of the large-scale consumer request time, can dynamically adjust the load strategy, the effect that reaches the dynamic task scheduling is to improve the response time of entire system.The simultaneously business service and the data, services of dynamic dispatching cluster server are called the multithreading ability that serviced component removably and scheduling distribute polycaryon processor according to user's use rank.Can reach the defective that effectively overcomes traditional open platform shortage extensibility and retractility in this way sets up fast construction application deployment running environment simultaneously and dynamically adjusts these two targets of application runtime environment; Calculate and multithreading through multinuclear simultaneously, improve the schedule speed and the responding ability of system service; Meanwhile provide modularization removably to serve, can isolate user's service of different rights and distribute corresponding calculated ability and processing power for it, support the otherness service between the different user.

Those of ordinary skill in the art will appreciate that embodiment described here is in order to help reader understanding's principle of the present invention, should to be understood that protection scope of the present invention is not limited to such special statement and embodiment.Those of ordinary skill in the art can make various other various concrete distortion and combinations that do not break away from essence of the present invention according to these teachings disclosed by the invention, and these distortion and combination are still in protection scope of the present invention.

Claims (4)

1. the dispatching method of a cloud computing open platform is characterized in that, comprises the steps:

The first step, central server are constructed some parallel multithreadings, and the structure process pool, put into process pool to the process of structure afterwards;

Second step, registration central server are created and are started finger daemon, and by the main thread initializes global variable, global variable should comprise semaphore at least, and sub-thread state collection and result set are used to carry out sub-thread synchronization and record queries record;

The 3rd step, operate in the connection of waiting for front end web acting server process behind the load equalizer initialize communications port in the central server; When the server connection request arrives; Load equalizer generates a thread and this server communication; Load equalizer continue to be waited for the connection request of other server, when a new client requests service, load equalizer from information table, selects load minimum be that load weights maximum service device is its service;

Request in the 4th step, the central server processing threads pond, and the service of calling application server, according to the multi-core parallel concurrent processor of user right calling application server cluster, completion process and thread are to the distribution of processor node;

The data that the 5th step, central server are collected database server are handled distributed data;

The 6th step, central server calls cluster service, and integrates mass data and return to the web acting server and call, and gives back the thread to thread pond of current request simultaneously, continues the request of monitoring users and also from thread pool, calls idle thread.

2. the dispatching method of cloud computing open platform according to claim 1 is characterized in that, the first step specifically comprises as follows step by step:

1) be the concurrent threading model of model creation with the FORK-JOIN structure;

2) thread that daemon thread FORK is come out is put into thread pool, is responsible for the life cycle management of thread by thread pool;

3) under the multi-threaded parallel processing environment, adopting variable mutex is the mutex amount, is used to realize the synchronous and mutual exclusion of the resource access under the multi-thread environment; The course of work of this mutex amount is following: when the resource that mutex representes was used in one of request, whether process read the value of mutex earlier, available to judge corresponding resource; When the value of mutex greater than 0 the time, showing has resource to ask; Equal at 0 o'clock, show no available resources, process gets into sleep state when available resources are arranged; When process did not re-use the shared resource of a semaphore control, the value of mutex increased 1.

3. the dispatching method of cloud computing open platform according to claim 1; It is characterized in that; The 3rd step load balancing strategy of described load balancing implement body is: when the node of cloud computing open platform comes into operation for the first time, set an initial weights SW (N _i), along with the variation of node load, balanced device is adjusted weights, and the parameter dynamic calculation of each side drew when weights were moved by node, node N _iWeights describe with following formula:

SW(N _i)＝K ₀*L_CPU(N _i)+K ₁*L_Memory(N _i)+K ₂*L_Process(N _i)+K ₃*L_IO(N _i)+K ₄*L_Response(N _i)

Wherein, K ₀, K ₁, K ₂, K ₃And K ₄The expression constant coefficient, L_CPU (N _i) be node N _iCPU usage, L_Memory (N _i) be node N _iMemory usage, L_Process (N _i) be node N _iRate of people logging in, L_IO (N _i) be occupation rate, the L_Response (N of the magnetic disc i/o of node Ni _i) be node N _iThe process response time, L_CPU (N _i)=1-P_CPU (N _i), wherein, P_CPU (N _i) utilization factor of expression current C PU.

4. the dispatching method of cloud computing open platform according to claim 1 is characterized in that, the 4th step specifically comprised as follows step by step:

1) foundation of Task Assignment Model:

If application server cluster comprises N _NodeIndividual processing node D ₀, D ₁..., Concurrent program to be allocated has N _ProcIndividual process P ₀, P ₁..., Process P _iComprise M _iIndividual thread T ₀, T ₁...,

Total thread number of concurrent program

Concurrent program to be allocated is expressed as a task graph of a relation, and (V, E), wherein V is the set { V of each application server node in the application server cluster to be specially a non-directed graph G= _i, node V _iCorresponding doublet<t _i, P _i>E is the set { E of nonoriented edge _Ij; Connected node V _iAnd V _jLimit E _Ij∈ E, expression thread T _iAnd T _jBetween communication or shared data, the weights W on limit _IjThe frequent degree of representing two thread communications or shared data;

2) carry out two wheel operations, the distribution of central server response process to cluster server accomplished in first round operation; Second wheel operation is accomplished the distribution of processing server intranodal thread to processor core; Each wheel operation comprises repeatedly iteration; Concrete processing procedure is following: begin from the initiating task graph of a relation of input; Each limit of selecting to have maximum weights, two summits of merging this edge, the number of threads that is comprised in the newly-generated node must be smaller or equal to a threshold value; Repeat this process, the number of node equals unappropriated number of threads in the central server current thread pond in the task graph of a relation; The threshold value that first and second wheel operation uses is calculated by following formula respectively:

Wherein, Threshold _{First_round}The threshold value of the processing node number of expression first round operation, Threshhold _{Second_round}The threshold value of representing the composite node number of second wheel operation, rounding operation in [] expression, M _MaxBe the maximum thread that process has, N _CoreFor existing process number, Max () represent the parameter of relatively importing into and get maximal value, α is a percent value, is illustrated between equally loaded and the minimizing communication to weigh; Specifically be divided into as follows step by step:

2a) to first round operation, initial task graph of a relation is that unit carries out preliminary division with each server individuality, the composite node number=process number in promptly scheming, and the corresponding process of each composite node, the thread that is included in the composite node all belongs to this process; The first round termination condition of operation is composite node number≤Threshold among the figure _{First_round}, each composite node is a sub-graphs in scheming during end, corresponding processing node, and the thread that is included in the subgraph should be distributed to this processing node;

Each subgraph that 2b) first round division of operations is gone out all carries out second wheel operation, and promptly the initiating task graph of a relation of second wheel operation is the subgraph that first round operation obtains, and the termination condition of second wheel operation is the process number≤Threshhold that is scheduled in the resource pool among the figure _{Second_round}, corresponding process of each composite node among figure during end, the thread that is included in the thread pool in the central server should be distributed to this processor core;

2c) target ip address of analysis request message and load balancing in view of the above; Request scheduling with same target IP address under the load balance situation of server arrives same node; Be specially: at first find out target ip address most recently used node; If this service node is available and not overload, then user's request is sent to this service node by load-balanced server; If service node does not exist, or this service node overloads and have service node to be in its half the operating load, and then each service node of poll is selected the minimum service node of link, and request is sent to this service node.

Images