CN102073546A - Task-dynamic dispatching method under distributed computation mode in cloud computing environment - Google Patents

Abstract

A method for dynamically scheduling tasks in a distributed computing mode in a cloud computing environment. The method has four steps: Step 1, the master node receives and analyzes the heartbeat information of the child nodes; The state table pre-allocates tasks; Step 3, the child node requests tasks from the master node; Step 4, the master node assigns tasks to the child nodes. The invention firstly considers the resource requirement of the task and the performance information of the node, and dynamically controls the allocation of the task when the requirement is met, thereby improving the response speed of the job and the resource utilization rate of the node. It has extensive practical value and application prospects in the field of distributed computing technology under the cloud computing environment.

Description

A method for dynamic scheduling of tasks in a distributed computing mode in a cloud computing environment

(1) Technical field

The invention relates to a task scheduling method of a distributed computing model. Specifically, it relates to a task dynamic scheduling method in a distributed computing mode in a cloud computing environment. It is a dynamic scheduling method based on node performance for tasks in a task scheduling subsystem, and belongs to the field of computer technology.

(2) Background technology

At present, with the rapid development of network applications, the demand for computing power continues to increase, and with the development of grid computing, parallel computing, and distributed computing, cloud computing has emerged as the times require, and is listed as a key technical direction for the country's future development. And it has become a hot research topic in today's computer research and industry circles. With the popularity of cloud computing, more and more network (Web) services and business applications are deployed in the cloud computing environment. For the distributed nodes processing application layer computing requests in the cloud environment, how to use task scheduling and Efficiently processing upper-layer computing requests, improving the resource utilization of heterogeneous distributed nodes in terms of performance, and improving the response speed of jobs has become a research hotspot in the field of cloud computing.

When processing massive data in the cloud environment, task scheduling based on distributed storage and distributed parallel processing is one of the key steps. Improving the scheduling method of jobs and tasks is a research hotspot at present. Research at home and abroad mainly includes scheduling between jobs when multiple jobs are running in parallel, scheduling subtasks when a single job is running, and optimizing the number of subtasks running in parallel.

In terms of job scheduling, the current most basic scheduling method is the first-in first-out job scheduling method. Undoubtedly, this job processing method has many disadvantages, especially when the number of jobs is large, the overall response time is very long. The fair scheduler (Fair scheduler) is a better solution to this problem. When a single job is running, it will use the entire cluster. When other jobs are submitted, the system will assign task idle time slices to these new jobs, so that each job can obtain approximately the same amount of CPU time, and make small tasks respond quickly while ensuring the performance of large tasks. Service Level. The Capacity scheduler supports multiple queues. After a job is submitted, it enters a queue, resources are allocated according to the queue, and the jobs in each queue use the resources of the queue; Priority jobs use resources; but once a job starts executing, it is not preempted by higher priority jobs; to prevent one or more users from monopolizing all resources, each queue is forced to allocate a certain percentage of resources. The MR-Predict-based three-queue scheduler proposed by the Institute of Computing Technology, Chinese Academy of Sciences divides workloads into three categories according to CPU and I/O usage, and can simultaneously improve the utilization of CPU and I/O resources under different types of workload environments. usage rate.

In terms of task scheduling, the LATE (Longest Approximate Time to End) scheduling algorithm proposed by researchers at the University of California, Berkeley focuses on optimizing the scheduling of backup tasks in jobs. By speculating on how long it will take to complete, this ensures that backup tasks that are estimated to be the latest to complete are only executed on the fastest nodes. Researchers at Purdue University proposed an optimal configuration method for the number of tasks based on historical statistical data. Their research focuses on the impact of the number of concurrently running tasks on each node in the cloud environment on performance when executing jobs. According to historical statistics data, obtain optimal configurations and apply them to new jobs of the same kind.

However, in most cases, the performance of different nodes is different, and the load status of each node is also different at different times. How to determine the dynamic allocation strategy of tasks according to the performance heterogeneity and dynamic liability status of nodes is very important for efficient processing of computing tasks and It is of great significance to improve the resource utilization rate of distributed nodes and improve the response speed of jobs.

(3) Contents of the invention

1. Purpose:

The main purpose of the present invention is to provide a dynamic scheduling method for tasks in a distributed computing mode in a cloud computing environment. It first considers the resource requirements of tasks and the performance information of nodes, and dynamically controls the assignment of tasks when the requirements are met. , thereby improving job response speed and node resource utilization.

In order to achieve the above object, the present invention proposes a dynamic scheduling method for tasks based on node performance and task execution status in the distributed computing mode in the cloud computing environment. The composition structure of the distributed computing nodes in the cloud computing environment is shown in Figure 1, mainly It includes a master control node (master node) and multiple computing nodes (child nodes). The computing nodes can be physical machines or virtual machines, which are transparent to the master control node, and the nodes are interconnected through the network. The master control node interacts with the computing node through remote procedure call (RPC). The main control node is mainly responsible for receiving the heartbeat information of the computing node, and analyzing and giving feedback to control the scheduling and execution of tasks; in addition to executing tasks, the computing node is also mainly responsible for collecting the performance information and task execution information of the node and sending them to the main control node. node.

2. Technical solution:

Technical scheme of the present invention is such:

A method for dynamically scheduling tasks in a distributed computing mode in a cloud computing environment according to the present invention. The specific process is as shown in Figure 2. The method includes the following steps:

Step 201. The computing node dynamically collects performance information and task execution information of the node, and reports to the master control node in the form of heartbeat information.

Step 202. The master control node receives and analyzes the heartbeat information of each computing node, creates and continuously updates the node status table and the task status table. According to the node state table and task state table, the master control node pre-allocates tasks for computing nodes, and updates the node prefetch table and task pre-allocation table.

Step 203. If there is an empty task slot (task slot) available in the computing node, add a flag requesting a task to the master control node in the next heartbeat message.

Step 204. After receiving the task request from the computing node, the master control node allocates tasks to it according to the scheduling policy, and updates the node prefetching table and task pre-allocation table.

Wherein, the node performance information and task execution information described in step 201 are important data sources for the master control node to update the node state table and task state table. Node performance information may include CPU main frequency, memory size, CPU usage, memory usage, I/O resource usage, and the like. The task execution information includes the task execution information that has just ended and the task execution information that is in progress; the task execution information that has just ended includes the TaskID of the task, the JobID of the job, the time used for IO (copy processing data) and the time used for CPU calculation The time for copying processing data occurs when the computing node does not have input data for this task; the ongoing task execution information includes the TaskID of the task, the JobID of the job it is in, the execution progress of the task, and the execution time. Each computing node collects these two kinds of information of the node at regular intervals, and encapsulates it as heartbeat information and sends it to the master control node.

Wherein, the node state table and the task state table described in step 202 are important reference information for the master control node to formulate a task allocation scheme. The node status table describes the performance status of each computing node in the recent period, and the task status table records the processing status of each computing node in the recent period. After the master control node receives the heartbeat information of the computing node for the first time, it creates a node status table and a task status table and updates these two tables each time it receives the heartbeat information of the computing node. The node status table includes the fields NodeName, CPU_Speed, MemSize, CPU_Usage, Mem_Usage, and IO_Usage; the task status table includes the fields JobID, TaskID, NodeName, Time_IO, Time_CPU, Progress, and PastTime. The node prefetch table and task preallocation table record the preallocation information of tasks in the current cluster. The node prefetch table records the information about pre-assigned tasks for computing nodes by the master control node. The node prefetch table includes fields such as NodeName, preFetched, and preFetchedTaskID. The task pre-allocation table records the information that the master control node pre-allocates tasks to computing nodes. The task pre-allocation table includes fields such as TaskID, preScheduled, and preScheduledNodeName.

Wherein, the size of the task slot of the computing node in step 203 refers to the maximum number of tasks that the computing node can execute in parallel at the same time, and the size of the task slot is configured before the start of the distributed node cluster. Computing nodes only apply for tasks to the master control node when there are free task slots. The application of the task is transmitted through the heartbeat message. The heartbeat message contains the flag bit of the application task. If it is true, it indicates that the computing node has an empty task slot. , the master control node can assign tasks to the computing node for execution.

Wherein, the allocation of task execution by the master control node to the computing node applying for the task described in step 204 is determined by a distributed scheduling algorithm. The distributed scheduling algorithm is implemented in the scheduler of the master control node. At the same time, multiple computing nodes may apply for task execution at the same time. The scheduler reads the node status table, task status table, node prefetch table and task pre-allocation table, Combined with the remaining task queue, according to the distributed scheduling algorithm, the priority order and the number of tasks assigned to computing nodes are determined, and then the node prefetching table and task pre-allocation table are updated.

3. Advantages and effects: a method for dynamically dispatching tasks in a distributed computing mode in a cloud computing environment of the present invention, which is different from the prior art in that its main advantages are: (1) by analyzing the performance dynamic changes and history of computing nodes Task execution information makes the allocation of tasks by the master control node more reasonable, and can give full play to the performance advantages of computing nodes with better performance. However, the original task scheduling method does not consider the dynamic variability of the performance of each computing node; (2) Changed the practice in typical distributed computing models (such as MapReduce) that as long as computing nodes apply for tasks from the master control node, they can obtain task execution, and the master control node has given the right to select computing nodes to perform tasks, so that This avoids the bottleneck problem caused by computing nodes with poor performance.

(4) Description of drawings

Fig. 1 is a schematic diagram of the composition structure of distributed computing nodes in the cloud computing environment of the present invention

Figure 2 Schematic diagram of task distributed scheduling process based on distributed node performance and task execution status in cloud environment

Fig. 3 is an interactive structural diagram of three phases (initialization, information update and task scheduling) included in the present invention

The detailed flowchart of three stages that Fig. 4 present invention comprises

Fig. 5 schematic flow chart of the information update module of the present invention

Figure 6 is a schematic flow diagram of the task scheduling module of the present invention

The symbols in the figure are explained as follows:

201-204 step number; 501-505 step number; 601-604 step number;

(5) Specific implementation methods

In order to make the object, technical solution and advantages of the present invention more clearly, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

The equipment environment conditions that the present invention needs to satisfy are shown in Fig. 1, and the composition structure of the distributed computing node in the cloud environment mainly includes a main control node (master node) and a plurality of computing nodes (sub-nodes), and the computing node can be a physical machine , it can also be a virtual machine, which is transparent to the master control node, and the nodes are interconnected through the network. The master control node interacts with the computing node through remote procedure call (RPC). The main control node is mainly responsible for receiving the heartbeat information of computing nodes, and analyzing and feeding back to control the scheduling and execution of tasks; among them, the node analyzer is used to receive and analyze the performance information of computing nodes, update the node status table, and the task analyzer is used to receive And analyze the task information of computing nodes, and update the task status table. In addition to executing tasks, computing nodes are also mainly responsible for collecting the performance information and task execution information of the node and sending them to the master control node; among them, the node performance monitor is responsible for collecting the performance information of the node in the latest period, and the task monitor is responsible for collecting the latest Record information of tasks executed for a period of time.

In terms of software conditions, the present invention requires each node to adopt a Linux operating system and to be installed with Java development kit version 1.6 and above.

In terms of environmental conditions, the present invention requires each node to be able to access each other through ssh without password.

The task dynamic scheduling process based on node performance and task execution status is shown in Figure 2, which mainly includes two contents: (1) The computing node collects and encapsulates the heartbeat information of the node and sends it to the master control node, and the master control node according to the received heartbeat information Establish and update the node state table and task state table; (2) After receiving the task request from the computing node, the master control node assigns tasks to the computing node according to the scheduling algorithm and updates the node prefetching table and task pre-allocation table.

The method includes three phases: initialization, information update and task scheduling. Its interactive structure is shown in Figure 3. In the initialization phase, the master control node receives the job, and establishes the node status table and task status table; in the information update phase, the master control node receives the heartbeat information of the computing node and updates the node status table, task status table, node prefetch table and task Pre-allocate the table. If the computing node requests a task, it enters the task scheduling stage; in the task scheduling stage, the master control node assigns tasks to the computing node according to the node information and task information, and returns to the information update stage after the end to wait for the heartbeat information of the computing node.

Illustrate with an example below, as shown in Figure 4, the method of the present invention comprises the following steps:

Step 401: The node performance monitor on the computing node collects the performance information of the node, and the task monitor collects the task execution information of the node, and then encapsulates it into heartbeat information and sends it to the master control node. The cycle of information collection and heartbeat information sending is 3 seconds.

Step 402: The master control node receives and analyzes the heartbeat information of each computing node. If it is the first time to receive the heartbeat information, create a node status table and a task status table. If it has already been created, update the node every time a heartbeat information is received State table and task state table. According to the node state table and task state table, the master control node pre-allocates tasks for computing nodes, and updates the node prefetch table and task pre-allocation table. Specifically, it is shown in the information update module in FIG. 5 .

Step 403: If the computing node has an empty task slot available, add a flag requesting a task to the master control node in the next heartbeat message.

Step 404: After receiving the task request from the computing node, the master control node allocates tasks to it according to the scheduling policy. Specifically, it is shown in the task scheduling module in FIG. 6 .

The detailed process of the information update module is shown in Figure 5.

Step 501: the master control node monitors the RPC access of the computing node, and receives the heartbeat information sent by the computing node. The master control node can only receive the heartbeat information of one computing node at a time. If the master control node receives the heartbeat information of a certain computing node, other computing nodes also send heartbeat information to the master control node, the master control node will be delayed. The heartbeat computing node joins the waiting queue. The node performance monitor on the computing node monitors and collects the performance information of the node in the latest period. The task monitor monitors the information of the tasks being executed on the node and collects the records of the last three historical tasks that have been executed. The computing node Encapsulate performance information and task information as heartbeat information. If the task information has not been updated in the recent period, the heartbeat information may only contain the performance information of the node. Computing nodes send heartbeat information to the master control node through RPC at regular intervals. The heartbeat period is 3 seconds. At each heartbeat, the heartbeat information should contain the performance information of the node and the information of the currently executing task, and every time the computing node completes a task, it will add the execution record of the task just completed to the heartbeat information and send it in the next heartbeat For the master control node, the task information contains two types of task information: completed but not reported task information (may be empty) and ongoing task information.

Step 502: The master control node updates the node state table and the task state table according to the received heartbeat information. For the node status table, the computing node status information in the heartbeat information is overwritten with the information corresponding to the computing node in the node status table in the master control node. The task status table records the information of the last three historical tasks executed on each node and the task information in progress. Every time the master control node receives new task status information, it first checks whether there are tasks that have been completed but not reported. If yes, obtain the TaskID of the task and check whether the task already exists in the task status table, if it exists, update the information of the task in the task status table, otherwise delete the last Old task information and add the completed task information. For the task information in progress, obtain the TaskID of the task, if the task already exists in the task state table, update the information of the task in the task state table, otherwise, add the information of the task in the task state table.

Step 503: Update the node prefetching table and the task pre-allocation table according to the node state table and the task state table. For the mth task in the task list, predict the time required for each node to execute the task according to the node state table and task state table, the prediction algorithm is as follows:

${\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}{\mathrm{<span\; class="notranslate">\; j</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; io</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; cpu</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, T _ij is the predicted time required for the i-th computing node to execute its j _- th task, t _s is the time when an available task slot appears, t _i0 is the time required for copying data, t _cpu is the data processing time, h is the reference number of tasks that the computing nodes have successfully executed, and n is the number of computing nodes in the cluster.

After obtaining the value of T _ij , the master control node selects the smallest one, and pre-allocates tasks to its corresponding computing nodes. And mark the task m as pre-segmented, and mark the computing node that will execute the task as pre-fetched.

Next, the master control node continues to select a computing node for executing the next task not marked as pre-partitioned among computing nodes not marked as pre-fetched. Only num tasks are pre-divided each time, where num=5. After the pre-allocation is completed, there may be tasks not marked as pre-allocated in the task queue, and there may be nodes not marked as pre-fetched in the node list.

Step 504: If the computing node marks the requested task field as true in the heartbeat information, enter the task scheduling module.

The detailed process of the task scheduling module is shown in Figure 6:

Step 601: The master control node first judges whether the task has been pre-allocated for the node by looking up the node prefetch table, and if it has been pre-allocated, assign the pre-allocated task to the node, mark the task as allocated, and mark the node as Not prefetched.

Step 602: If the master control node does not pre-allocate tasks for this node, according to the prediction algorithm described in step 503, the nodes that are not marked as having been prefetched are nodes with poor performance or weak computing power, then the task queue Select a task for the node to execute, but do not mark the task as allocated, and the task will be pre-allocated to the node with faster processing speed when the next pre-allocation is performed, so this task will have a backup task in the Execute on the fast node to ensure its smooth execution.

Step 603: After allocating tasks to the computing nodes, the master control node updates the node prefetching table and task preallocation table.

Step 604: the task scheduling phase is over, and the information update phase is entered, and the master control node continues to receive and process the heartbeat information sent by the computing node.

Finally, it should be noted that the above embodiments are only used to illustrate and not limit the technical solutions of the present invention, although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the present invention can still be modified Or an equivalent replacement, any modification or partial replacement without departing from the spirit and scope of the present invention shall fall within the scope of the claims of the present invention.

Claims (5)

1. the task dynamic dispatching method under the distributed computing model in the cloud computing environment, by dynamically obtaining and the resource requirement of analysis task and the performance information and the historic task execution information of node, distribution to task under situation about satisfying the demands is dynamically controlled, thereby improve the response speed of operation and the resource utilization of node, it is characterized in that: this method may further comprise the steps:

Step 1: the performance information of this node of computing node dynamic collection and task are carried out information, report to main controlled node with the form of heartbeat message; Main controlled node receives and analyzes the heartbeat message of each computing node, generates node state table and task status table;

Step 2: main controlled node is computing node predistribution task according to node state table and task status table, and more new node is looked ahead and shown and the pre-submeter of task;

Step 3:, then in the heartbeat message of next time, add sign to main controlled node request task if free task groove is that task slot can use in the computing node;

Step 4: after main controlled node receives the task requests of computing node, press scheduling strategy and be its allocating task.

2. the task dynamic dispatching method in a kind of cloud computing environment according to claim 1 under the distributed computing model is characterized in that: described joint behavior information of step 1 and task execution information are the significant data sources that main controlled node upgrades node state table and task status table; Joint behavior information comprises CPU frequency, memory size, CPU usage, memory usage and I/O resource utilization; Task execution information comprises the task execution information of firm end and the task execution information of well afoot; Just the task execution information that finishes comprises that the JobID of TaskID, the place operation of task, the time that is used for IO are the replication processes data and are used for the time that CPU calculates, wherein, the replication processes data occur in this computing node and do not have under the input data conditions of this task and take place; The task execution information of well afoot comprises the TaskID of task, JobID, task executions progress and the executed time of place operation; Each computing node is collected these two kinds of information of this node at set intervals, and is encapsulated as heartbeat message and sends to main controlled node.

3. the task dynamic dispatching method in a kind of cloud computing environment according to claim 1 under the distributed computing model is characterized in that: described node state table of step 2 and task status table are the important references information that main controlled node is formulated the Task Distribution scheme; Node state table has been described the performance state of each computing node in a recent period of time, the task status table record situation of each computing node Processing tasks in a recent period of time; After main controlled node receives the heartbeat message of computing node for the first time, upgrade these two tables after creating node state table and task status table and receiving the heartbeat message of computing node afterwards at every turn; Node state table comprises NodeName, CPU_Speed, MemSize, CPU_Usage, these fields of Mem_Usage, IO_Usage; Task status table comprises JobID, TaskID, NodeName, Time_IO, Time_CPU, Progress, these fields of PastTime; Node look ahead table record main controlled node be the information that computing node is allocated task in advance, the node table of looking ahead comprises NodeName, preFetched, these fields of preFetchedTaskID; Task presort table record main controlled node allocate task in advance information to computing node, the pre-submeter of task comprises TaskID, preScheduled, these fields of preScheduledNodeName; The specific implementation process is as follows:

1) main controlled node is monitored the RPC visit of computing node, receives the heartbeat message that computing node sends; The main controlled node synchronization can only receive the heartbeat message of a computing node, if main controlled node is when receiving the heartbeat message of certain computing node, have other computing nodes also to send heartbeat message to main controlled node, then main controlled node adds waiting list with the computing node of later heartbeat; Joint behavior watch-dog monitoring on the computing node is also collected interior performance information of nearest a period of time of this node, the Mission Monitor device is monitored the information of carrying out on this node of task and is collected the record of executed 3 historic tasks recently, and computing node is encapsulated as heartbeat message with performance information and mission bit stream; If the mission bit stream in a period of time does not recently upgrade, also can only comprise the performance information of node in the heartbeat message; Computing node sends to main controlled node with heartbeat message by the RPC mode at set intervals; Heart beat cycle is 3 seconds, during each heartbeat, all should comprise the performance information of node and the information of current carrying out of task in the heartbeat message, and computing node whenever executes a task, the task executions that all will just finish when heartbeat next time record adds heartbeat message and sends to main controlled node, promptly comprises two generic task information in the mission bit stream: finished but the mission bit stream that do not report and the mission bit stream of well afoot;

2) main controlled node upgrades node state table and task status table according to the heartbeat message that receives; For node state table, the computing node status information in the heartbeat message is covered in the main controlled node node state table corresponding to the information of this computing node; Writing down the information of nearest 3 historic tasks of carrying out on each node and the mission bit stream of well afoot in the task status table, when main controlled node is received new task status information at every turn, at first see if there is the mission bit stream of having finished but not reported, if have, then obtain the TaskID of this task and check whether this task exists in task status table, if there is the information of this task in the updating task state table then, otherwise the oldest mission bit stream of this computing node and add this completed mission bit stream in the deletion task status table; For the mission bit stream of well afoot, obtain the TaskID of this task, if this task exists in task status table, the information of this task in the updating task state table then, otherwise, in task status table, add the information of this task;

3) according to node state table and task status table look ahead table and the pre-submeter of task of new node more.For m task in the task list, according to node state table and task status table, predict that each node carries out the time of this required by task, prediction algorithm is as follows:

$T_{i{j}_i}=\frac{{\mathrm{\&Sigma;}}_{j-h}^{j-1}(t_s+t_{\mathrm{io}}+t_{\mathrm{cpu}})}{h},i=\mathrm{1,2}......n---\left(1\right)$

Wherein, T _IjBe that i computing node carried out its j _iThe predicted time of individual required by task, t _sFor the time of available task groove, t occurring _I0Be required time of copy data, t _CpuBe data processing time, h is the computing node reference number of the task of successful execution, and n is the computing node number in the cluster;

Obtain T _IjValue after, main controlled node is selected minimum one, to its corresponding computing node predistribution task, and task m is labeled as presorts, and the computing node that is about to carry out this task is labeled as looks ahead;

Then, main controlled node continues not carry out its computing node for the next task choosing of having presorted that is not labeled as in being marked as each computing node of having looked ahead, only presorts num task, wherein num=5 at every turn; Presort finish after, may also have in the task queue not to be marked as the task of having presorted, also may have in the node listing and not be marked as the node of having looked ahead;

4) if computing node will ask the task word segment mark to be designated as very, then enter task scheduling modules in heartbeat message.

4. the task dynamic dispatching method in a kind of cloud computing environment according to claim 1 under the distributed computing model, it is characterized in that: the task groove of the described computing node of step 3, its size is meant the maximum number of tasks of computing node synchronization energy executed in parallel, and the size of task groove configures before the distributed node cluster starts; Computing node only in free task groove just to main controlled node application task, the application of task is by the heartbeat message transmission, the zone bit that comprises the application task in the heartbeat message, if for very then show the task groove that this computing node is free, main controlled node can assign the task to this computing node and carry out.

5. the task dynamic dispatching method in a kind of cloud computing environment according to claim 1 under the distributed computing model is characterized in that: the described main controlled node of step 4 is that the computing node allocating task execution of application task determines by the distributed scheduling algorithm; The distributed scheduling algorithm is realized in the scheduler of main controlled node, synchronization has a plurality of computing nodes application task simultaneously to be carried out, scheduler is by reading node state table, task status table, node look ahead table and the pre-submeter of task, and in conjunction with the residue task queue, be defined as the priority ranking and the task number of computing node allocating task according to the distributed scheduling algorithm, then look ahead table and the pre-submeter of task of new node more; The specific implementation process is as follows:

1) main controlled node at first judges whether to allocate task in advance for this node by searching the node table of looking ahead, if allocate in advance, the then task of having presorted for the distribution of this node, and this task of mark is distributed, and this node of mark is not for looking ahead;

2) if main controlled node is not allocated task in advance for this node, according to the prediction algorithm formula (1) described in the step 2, not being marked as the node of having looked ahead is the more weak node of poor-performing or computing power, then choosing a task from task queue carries out for this node, but not with this task flagging for distributing, treat pre-timesharing next time, this task will be given processing speed node faster in advance, thereby this task will have a backup tasks to carry out to guarantee its smooth execution on fast node;

3) for after the computing node allocating task finishes, main controlled node is look ahead table and the pre-submeter of task of new node more;

4) the task scheduling stage finishes, and changes the information updating stage over to, and main controlled node continues to receive and handle the heartbeat message that computing node sends over.

Images