CN1917464A - Distribution type task assignment and computation method based on lower bound to be raised step by step - Google Patents

Abstract

The method comprises: when allocating tasks in the expandable router, firstly giving out an initial load allocation scheme as the lower bound of optimal solution; building and gradually searching the next available allocation scheme until a currently optimal task allocation mode is found. Due to the determination of initial load allocation scheme has high influence on the search process, and in term of unknowing the load rules, the bad or good of an initial condition is hardly estimated, the invention fully uses the distributed structure to make search from different initial conditions in multi parallel paths until an optimal solution is found.

Description

Distribute and computational methods based on the distributed task scheduling that progressively promotes lower bound

Technical field

Belong to internet ip router data and task scheduling algorithm research field based on the distributed task scheduling distribution that progressively promotes lower bound with computational methods.

Background technology

In the computer network system in modern times, router is most crucial connection device.Rapid development of Internet has been brought the innovation of router architecture.Next Generation Internet has proposed new demand to ip router architecture of future generation.Single node is optimized, improves its disposal ability and no longer be the improved fundamental way of system, because this mode is subject to the development speed influence of physical device.Traditional IP router architecture based on single node can't adapt to rapid development of network.Extendible architecture is the main developing direction of ip router of future generation.

In the expandable route software architecture, how the treatment effeciency that improves different processing units by the rational management to task is the hot issue of research always.Though the Distributed Calculation theory provides of great value reference for the expandable route software architecture, but because the main target of distributed computing system is that what it stressed is the service ability that system can provide as a computing machine for a class problem provides general processing platform.Need when handling very big load, guarantee very high real-time.A complete router software architecture comprises the diverse calculation task of some patterns, for example network management, route calculating etc., its computation schema of different Routing Protocols also has very big difference, even a Routing Protocol inside, the different subtasks of a plurality of functions such as route is mutual, route calculating, routing table maintenance have also been comprised usually.In a word, the router software architecture is a very complicated multitask system, and the existing distributed computing system almost can't satisfy all properties demand of expandable route simultaneously under a structure.And in the router applications scene of reality, have a plurality of relatively independent tasks, it is also very different that each task resource takies pattern, uses the mode of the overall situation to seek that optimum allocation pattern between all tasks is nonsensical also can't to be reached.And task need satisfy two conditions in the allocation model on the node: the load on each node can not be transshipped; Inter-node communication can not transship.So how to avoid overload, how may cause single or some loads that certain association arranged of bottleneck to estimate and can bear load by the search system maximum at each is the key issue that prevents system bottleneck and improve the appearance of whole distributed system performance.

Summary of the invention

The object of the present invention is to provide a kind of distributing and computational methods of system overload of avoiding based on the distributed task scheduling that progressively promotes lower bound.

The invention is characterized in that it contains following steps successively:

Step (1.): set up a node work distribution chart, wherein, line number is represented the number of node, and columns is represented the task number, form a load matrix, the element of this load matrix be respective nodes after the Task Distribution in each element corresponding space remaining number of resources; Simultaneously take resource situation and fill in work distribution chart according to task;

Step (2.): set up a policy tag table, wherein, line number is represented interstitial content, columns is represented the task number, matrix element is meant used or does not also have used allocating task mode, wherein, with-1 when representing preliminary examination or the space when allowing to carry out Task Distribution, represent used node elements with-2;

Step (3.): from work distribution chart, select a kind of Task Distribution mode as initial mode randomly, insert in the described policy tag table, and to all nodes successively according to steps of processing:

Step (3.1.): the idling-resource number of pairing each node of branch mode that meets accident in the node work distribution chart described in the step (3) is inserted on the interior corresponding position of policy tag table, and other positions in the policy tag table are all represented with-1;

Step (3.2.): the scanning strategy label table, find in this table by that position of institute's assignment minimum in the element of assignment;

Step (3.3.): node ground is attempted pairing task switching in the described position of step (3.2.) to other nodes one by one according to the following steps.Up to find make the described minimum value of step (3.2.) obtain the lifting of maximum-norm till;

Step (3.3.1.): earlier the task in that position described in the step (3.2.) is moved to another node;

Step (3.3.2.): again task to be processed in described another node of step (3.3.1) is moved on in described another node of step (3.3.1), on the position of representing with numerical value-1, calculate its element value again;

Step (3.3.3.): the idling-resource number of element on that position that had been exchanged in the another one intranodal task described in the calculation procedure (3.3.2);

Step (3.3.4.): the scanning strategy label table, after described that of step (3.2) exchanged no longer the numerical value on the use location be revised as-2;

Step (4.): in the determination strategy label table on being labeled other positions-2 the position whether can allow to exchange next time, if allow, then forward the scanning that step (3.2.) continues the policy tag table to, otherwise search finishes.

Experimental research, these computational methods are very sensitive to the original allocation pattern, but in actual conditions, system itself is exactly distributed, therefore, the search that can on this distributed basis, walk abreast, each node is chosen different original allocation patterns, as long as there is node to obtain optimal solution, this method can stop.So just significantly improved the performance of calculating, and also reduced the dependence of this method the original allocation pattern.

We are respectively under different node scales, repeat to generate receivability load matrix at random, carry out this search procedure, Fig. 5. in provided and attempted the variation tendency of next feasible solution number of times with the node scale, each node configuration is got different maximum receivability matrixes measure 10 times, data mark with maximum, minimum value and the mean value of measurement result respectively.The mean value of number of attempt increases with more stable rule, but there be certain " shake " in maximum and minimum value, and the difference of maximum and minimum value increases with the increase of node scale.This result shows that the hunting range of its performance may be very big, but repeatedly search for if once promote the lower bound search, and its minimum value still has certain comparative, and this illustrates that we adopt parallel algorithm can improve the performance of this algorithm

Fig. 6. in further illustrate the search step number growth trend.The mean value of measuring is further from the minimum value of measuring, and the distribution of search step number under the different experiments data also more concentrates on bigger zone.This result shows the branches that increases calculating by turns by the initial matching pattern, helps significantly reducing total computation complexity.

Description of drawings

Fig. 1. node Task Distribution hoist pennants;

Fig. 2. build the policy tag hoist pennants;

Fig. 3. policy tag table task switching schematic diagram;

Fig. 4. overall process flow figure of the present invention;

Fig. 5. feasible solution growth trend figure next time;

Fig. 6. the growth trend figure of search step number;

Fig. 7. application exemplary plot of the present invention.

Embodiment

The key problem that classical distributed computing platform solves be how with load balancing to a plurality of computing nodes, improve the resource utilization of treatment effeciency and system, " overload " is not main problem category.But in router, overload is the key performance bottleneck of system, and the evaluation index that can expand software architecture is exactly the load capacity that can hold.Designing a kind of method for allocating tasks that can avoid system overload, is main contribution of the present invention.

The basic thought of this method is that given initial load allocating scheme is as the lower bound of optimal solution.Progressively search for next feasible allocative decision then, improve the lower bound of optimal solution simultaneously, up to the maximum that finds " current maximum receivability load ", promptly overall maximum receivability load.Wherein, the every wheel searched for the minimum value that " the optimal solution lower bound " that obtain is exactly receivability normalization load in the current allocative decision.Therefore, new search procedure can be skipped and comprise all allocative decisions that are not more than this minimum value, thereby all reduces the space searched for every after taking turns search.In addition, the original allocation scheme determination has a significant impact search procedure, but is difficult to judge the quality of initial condition under the situation of not knowing the load rule, therefore can utilize distributed frame, the n road is parallel to be searched for from different initial condition, up to having one the tunnel to find optimal solution.

In description of the invention, the distribution of node task is used the node work distribution chart and built the policy tag table and be described, wherein line number is represented the number of node, columns is represented number of tasks, and the list item in the node work distribution chart has represented that certain task is assigned to the surplus resources number of later this node on a certain node.

Node Task Distribution list structure is seen Fig. 1.

Build the policy tag table and be used to a kind of method of salary distribution of mark at task.Wherein, represent used task switching position with-2 with-1 position of representing to carry out task switching.

Build the policy tag list structure and see Fig. 2.

When needs carry out Task Distribution, at first to the node work distribution chart with build the policy tag table and carry out initialization, and in all nodes, handle later surplus resources number according to each task and fill in the node work distribution chart, and fill in the policy tag table, and promote operation successively at surplus resources item minimum in the policy tag table according to the initiating task methods of salary distribution different in each node.

Policy tag table task switching process is seen Fig. 3.

In Fig. 3, in the original allocation scheme, task 2 is assigned to node 2, task 5 is assigned to node 5, if the position at these two task places is intercoursed, then in the policy tag table minimum surplus resources just by exchange preceding 5 become 8 after the exchange, promptly finished the lifting of a lower bound.

Overall process flow figure of the present invention sees Fig. 4.

Application exemplary plot of the present invention is seen Fig. 7.

The present invention can clocklike promote gradually by the lower bound of system's available resources in the initiating task allocative decision that is generated at random, obtain a kind of more excellent Task Distribution pattern that in the distribution router system, is used to avoid system overload, by the bottleneck problem of optimization system, finally improved the overall performance of whole distributed system.

Claims (1)

1. distribute and computational methods based on the distributed task scheduling that progressively promotes lower bound, it is characterized in that this method is by realizing according to the following steps successively after the Task Distribution module of middle adding that can expand ip router in the Internet:

Step (1.): set up a node work distribution chart, wherein, line number is represented the number of node, and columns is represented the task number, form a load matrix, the element of this load matrix be respective nodes after the Task Distribution in each element corresponding space remaining number of resources; Simultaneously take resource situation and fill in work distribution chart according to task;

Step (2.): set up a policy tag table, wherein, line number is represented interstitial content, columns is represented the task number, matrix element is meant used or does not also have used allocating task mode, wherein, with-1 when representing preliminary examination or the space when allowing to carry out Task Distribution, represent used node elements with-2;

Step (3.): from work distribution chart, select a kind of Task Distribution mode as initial mode randomly, insert in the described policy tag table, and to all nodes successively according to steps of processing:

Step (3.1.): the idling-resource number of pairing each node of branch mode that meets accident in the node work distribution chart described in the step (3) is inserted on the interior corresponding position of policy tag table, and other positions in the policy tag table are all represented with-1;

Step (3.2.): the scanning strategy label table, find in this table by that position of institute's assignment minimum in the element of assignment;

Step (3.3.): node ground is attempted pairing task switching in the described position of step (3.2.) to other nodes one by one according to the following steps.Up to find make the described minimum value of step (3.2.) obtain the lifting of maximum-norm till;

Step (3.3.1.): earlier the task in that position described in the step (3.2.) is moved to another node;

Step (3.3.2.): again task to be processed in described another node of step (3.3.1) is moved on in described another node of step (3.3.1), on the position of representing with numerical value-1, calculate its element value again;

Step (3.3.3.): the idling-resource number of element on that position that had been exchanged in the another one intranodal task described in the calculation procedure (3.3.2);

Step (3.3.4.): the scanning strategy label table, after described that of step (3.2) exchanged no longer the numerical value on the use location be revised as-2;

Step (4.): in the determination strategy label table on being labeled other positions-2 the position whether can allow to exchange next time, if allow, then forward the scanning that step (3.2.) continues the policy tag table to, otherwise search finishes.

Images