JP4265377B2 - Load balancing method, apparatus, system, and program - Google Patents

Description

  The present invention relates to a load distribution method and apparatus, and more particularly to a load distribution method, apparatus, system, and program suitable for use in a cluster composed of a plurality of nodes connected by a network.

  A load distribution method for processing a plurality of types of tasks on a system composed of a plurality of computing resources is described in, for example, Non-Patent Document 1 described later. The load distribution method described in Non-Patent Document 1 is static. In other words, control based on the assumption that the performance of each computing resource, task load content, task arrival characteristics, etc. are determined in advance by some method and that the characteristics determined in advance do not change. Is the method. Static load balancing is effective for balancing and balancing steady loads with little fluctuation, but the actual load is constantly fluctuating, and in a short time interval, the load is always constant. There is an equilibrium. Dynamic load balancing solves this problem.

  As an example of a dynamic load distribution method for multiple types of transactions, dynamic load distribution can be achieved by estimating the load characteristics (job mix) that combines multiple loads and the expansion rate of program processing time. There has been proposed a method (see, for example, Patent Document 1 described later). Using a job mix is one solution, but even if the job mix is the same, any type of transactions that make up the job mix will affect the overall system performance. The factor of whether it is given is not taken into account. Therefore, a problem has not been solved for a case where a certain type of transaction processing is impeding the execution of other types of transactions more than necessary.

  An example of a load distribution method in which each node determines a load transfer destination based on load information of its own node is described in, for example, Non-Patent Document 2 described later. In Non-Patent Document 2, each node periodically collects load information of all other nodes, and the difference Δ between the load of the own node and the load of the lightest node among all nodes is larger than a certain value. In such a case, a load distribution method has been proposed in which the arrived task is transferred to the node with the lightest load.

  In the method described in Non-Patent Document 2, in order to collect information from all other nodes and select a task transfer destination, in a system composed of a large number of nodes, computation and communication necessary for sharing load information The cost increases and the performance improvement due to load balancing is negated by the information sharing cost. However, if the information update interval (information collection cycle) is increased in order to reduce the information sharing cost, load distribution is performed based on old information. Although the load is light when updating the load information, a task may be transferred to a node with a heavy load at present, and it becomes difficult to balance the load.

  In addition, a method of sharing load information of a large number of nodes at low cost by sharing information with different accuracy is described in the prior application 1, but the method described in Non-Patent Document 2 is accurate. Cannot handle different information.

  When there is load information of a low load node with a large number of nodes, a large number of nodes transfer the load to this low load node almost simultaneously, and the low load node becomes an overloaded node. The “crowd effect” occurs. The crowd effect is one of the reasons that hinders load balancing.

  Since the node that collects load information is a node that can be a task transfer destination, if a fixed number of nodes that collect load information are randomly selected, the cost of sharing information can be reduced, and the occurrence of crowd effects can be suppressed. Can do. For such an algorithm, for example, the description of Non-Patent Document 3 described later is referred to. However, the method described in Non-Patent Document 3 cannot know the load information of the entire system, and thus the load is not always balanced across the entire system.

  The load changes not only by the arrival of the task but also by the departure of the task. What the two algorithms do is to distribute task arrivals. For this reason, the load imbalance caused by the variation in task execution time cannot be directly resolved. In other words, tasks that have already arrived cannot be transferred.

  An algorithm for transferring a task that has already arrived to a node with the lowest load is described in Non-Patent Document 4, for example. However, in the method described in Non-Patent Document 4, the maximum number of tasks to be transferred at one time is two, and the transfer destination node is limited to a single node. Therefore, when the degree of load imbalance is large Takes time to balance the load. A crowd effect also occurs.

[Prior Application 1]
Japanese Patent Application No. 2002-333741 (FIGS. 1 and 2)

JP 2001-109638 A (page 6) “Load Balancing Problems for Multi-class Jobs in Distributed / Parallel Computer Systems” by Jie Li and Hisao Kameda IEEE Transactions on Computers, USA, 1998, 47, 3, 322-332 Songnian Zhou, “A Trace-Driven Simulation Study of Dynamic Load Balancing”, I.E.E.Transactions on Software Engineering ( IEEE Transactions on Software Engineering), USA, 1988, Vol. 14, No. 9, pp. 1327-1341 By Michael Mitzenmacher, “How Useful Is Old Information?”, IEE Transactions on Parallel and Distributed Systems ( IEEE Transactions on Parallel and Distributed Systems), USA, 2000, Vol. 11, No. 1, pp. 6-20 Simulations of three adaptive, decentralized controlled, job scheduling algorithms by John Stankovic, Computing Networks, USA, 1984 Year, Vol. 8, 199-217 Kazuo Kameda, Makoto Kiichi and Wang Li, “Basics and Applications of Performance Evaluation”, Kyoritsu Shuppan, 1998, Chapter 4.2 Masao Iri and Kazuken Fujino, “Common sense of numerical computation”, Chapter 15 Interpolation (interpolation), Chapter 16 Cloud ruler and spline-continuous interpolation (pages 109-122), Kyoritsu Shuppan, 1985,

  The conventional load balancing described above has the following problems.

  The first problem is that, in a situation where a plurality of task processes are performed on a cluster system composed of a plurality of nodes, the occurrence / arrival of a certain type of task is extremely increased. This means that the execution of a task that is not present will be affected, resulting in a situation such as extreme deterioration in response time.

  The reason is that a plurality of types of tasks are executed in the node, so that execution of a certain task type prevents execution of other task types.

  The second problem is that a crowd effect occurs.

  The reason is that each node does not consider the load transfer of other nodes.

  The third problem is that if the degree of load imbalance is large, it takes a long time to balance the load.

  The reason is that only the task that arrived is transferred, or even if the task that has already arrived is transferred, only a small number of tasks are transferred to a single node.

  The fourth problem is that when the information sharing cost is lowered, appropriate load distribution cannot be performed.

  The reason is that if the information update interval is increased, the task is transferred to a node that is lightly loaded at the time of information update but is currently heavily loaded. Further, when only the load information of some nodes is shared, there is a problem that even if the load can be balanced locally, it becomes difficult to achieve the overall load balance.

  Therefore, the present invention was created in view of the above problems, and its purpose is to assign nodes to task types so that loads due to different types of tasks do not affect each other's task execution. It is an object to provide a method, an apparatus, a system, and a program.

  Another object of the present invention is to provide a load distribution method, apparatus, system, and program that make it difficult to generate crowd effects.

  Still another object of the present invention is to provide a load distribution method, apparatus, system, and program capable of balancing loads at high speed.

  Still another object of the present invention is to provide a load distribution method, apparatus, system, and program capable of balancing loads even when load distribution is performed based on load information with different accuracy, which is provided at low cost. There is.

  The present invention that achieves the above-described object enables a plurality of types of tasks to be processed so as to satisfy the response time desired by the user, and each node can determine the load transfer destination based on the load information of its own node. It is what I did.

  More specifically, a load balancer according to one aspect of the present invention (also referred to as a “global load balancer”) inputs system configuration information, system operating conditions, and task-node characteristics measured in advance. System configuration input means; task receiving means for determining the type of task arriving at the system and measuring task arrival rate for each task; execution performance measuring means for measuring runtime performance information of a node during execution; and the system A node group is configured using configuration information, system operating conditions, task-node characteristics measured in advance, the task arrival rate measured at the time of execution, and node performance information at the time of execution, Select the node group construction means to determine the task type to be executed on the node group and the node to execute the task from the node group, Has a task transfer means for transferring tasks on-option node, the.

  The load distribution apparatus according to the present invention may further include global queuing means for queuing tasks so that the number of staying tasks on the task transfer destination node does not exceed a predetermined number.

  In the load distribution apparatus according to the present invention, the node group construction means corresponds to a task type that satisfies the operating condition for one node group and corresponds to a task type that does not satisfy the operating condition for another node group. A configuration may be adopted.

  In the load distribution apparatus according to the present invention, the node group construction means configures a node group so that one type of task processing is performed on one node group, and associates a task type with the node group. It is good also as a structure to perform.

A node according to another aspect of the present invention is a node constituting a computer system,
Task transmission / reception means for transmitting / receiving tasks or task execution results, local queuing means for queuing tasks received by the task transmission / reception means, load information sharing means for sharing load information, and the load information sharing means Based on the load information of the other nodes to be provided, it is determined which task is to be transferred to which node, load distribution means for requesting the task transmission / reception means to transfer the task, and processing the task requested to be executed, Task execution means for returning the execution result of the task to the local queuing means.

  In the node according to the present invention, the load distribution unit preferably includes a load information management unit that manages load information, a balance degree calculation unit that dynamically calculates the amount of tasks to be executed by the node, and a node Node priority assigning means for assigning priorities, and transfer task determining means for selecting a load transfer target node in order of priority and determining the number of tasks to be transferred to the node.

  In the node according to the present invention, the priority assigned to the other nodes by the node priority assignment means is cyclic monotonically increasing priority assignment.

  In the node according to the present invention, the load information sharing unit shares load information having different accuracy, the load information management unit of the load distribution unit performs simple space accuracy complement, and the node of the load distribution unit The priority assignment means may be configured to assign a higher priority to a node with high load information accuracy.

  The node according to the present invention has a response time table for calculating throughput according to the number of executions of a task, and the local queuing unit of the load distribution unit refers to the response table and executes tasks simultaneously. It is good also as a structure which changes a number dynamically.

  According to yet another aspect of the present invention, a method for configuring a node group based on at least one of system configuration information, pre-measured task-node characteristics, task arrival rate, and node runtime performance information. And a step of setting the size of each node group to an appropriate value using the load information obtained at the time of execution.

  The method according to another aspect of the present invention includes a step in which a load balancer receives a task processing request, determines a task type, and measures a task arrival rate, and the load balancer constitutes a computer system. The node is divided into a plurality of node groups, and which task type is executed in which node group is determined, using the system configuration information obtained from the input device and the measured performance measurement information at the time of execution. Adjusting the number of nodes included in each node group to control so as to satisfy the operating conditions specified by the user, and the load balancer is configured to select one of the node groups that execute a certain task type. Selecting one node and transferring the task to the selected node.

  In the method according to the present invention, the load distribution apparatus includes a step of queuing tasks so that the number of tasks staying on the node does not exceed a predetermined number.

  In a method according to another aspect of the present invention, in a load distribution method for a system in which a plurality of nodes constitute a computer system, a load transfer target node is determined based on load information of another node, and the node is assigned to the node. It includes a step of performing load distribution processing by transferring a task.

In the method according to the present invention, the step of performing the load balancing process includes:
Calculating the amount of tasks to be executed on the own node;
Assigning a priority to a node;
Selecting a node to which a load is to be transferred in order of priority, and determining the number of tasks to be transferred to the node.

Furthermore, a program according to another aspect of the present invention is provided in a computer constituting a global load distribution apparatus for a computer system including a plurality of nodes.
System configuration input processing for inputting system configuration information, system operating conditions, and task-node characteristics measured in advance;
Task reception processing that determines the type of task that arrives at the system and measures the task arrival rate for each task;
An execution performance measurement process that measures the execution performance information of a node at the time of execution;
At least one of the system configuration information, the operating conditions of the system, the task-node characteristics measured in advance, the task arrival rate measured at the time of execution, and the runtime performance information of the node A node group is constructed, and a node group construction process for determining a task type to be executed on the node group;
A task transfer process for selecting a node to execute a task from the node group and transferring the task to the node;
It consists of a program to execute.

Furthermore, a program according to another aspect of the present invention is provided in a computer constituting a node constituting a computer system.
Task transmission / reception processing for transmitting / receiving tasks or task execution results;
A queuing process for queuing a task received by the task transmitting / receiving means;
Load distribution processing for determining which task is transferred to which node based on load information of other nodes provided by a load information sharing unit that shares load information, and requesting task transfer to the task transmission / reception means;
A task execution process for processing a task requested to be executed and returning an execution result of the task to the local queuing means;
It consists of a program to execute.

In the program according to the present invention, the program for executing the load balancing process is:
Load information management processing for managing load information;
Balance calculation processing that dynamically calculates the amount of tasks to be executed on the own node;
Node priority assignment processing for assigning priorities to nodes;
It comprises a program for causing the computer constituting the node to execute a transfer task determination process for selecting a load transfer target node in order of priority and determining the number of tasks to be transferred to the node.

  According to the present invention, in order to configure a node group, to control the number of nodes constituting the node group to an optimal value at the time of execution, and to determine the type of task to be executed on the node group, Even if the task type executed on a certain node group is overloaded, the execution performance of the task type executed on another node group is not adversely affected.

  According to the present invention, since the node group for executing the task type is configured, the task execution is not affected beyond the node group. The execution of is not affected and can satisfy a predetermined response time.

  According to the present invention, when the load of the own node exceeds the threshold, the load transfer is determined from the load of the node that has made the transfer determination in the past and the load of the node that is currently determining the transfer. The crowd effect can be mitigated by considering the task transfer of other nodes.

  According to the present invention, it is possible to balance the load at a high speed by deciding to transfer one or more tasks to one node and to transfer the task to one or more nodes. is there. According to the present invention, a plurality of surplus loads are moved with respect to a plurality of nodes.

  According to the present invention, it is possible to share at low cost by determining load transfer based on load information having different accuracy. Furthermore, by reconstructing the load information from the accuracy and load of the load information having different accuracy, appropriate load distribution can be realized.

  Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 1 shows the configuration of a global load distribution apparatus 1 according to the first embodiment of this invention. Referring to FIG. 1, the global load distribution apparatus 1 according to the first embodiment of the present invention includes information on a system configuration (for example, system configuration information, system operation conditions, and task-node characteristics measured in advance). ), A task receiving unit 11 that determines the type of task arriving at the system and measures the task arrival rate for each task, and an execution performance measuring unit 14 that measures the runtime performance information of the node. A node group using system configuration information, system operating conditions, pre-measured task-node characteristics, task arrival rate measured at execution time, and node runtime performance information, The node group construction means 12 for determining the task type to be executed on the node group and the node for executing the task are selected from the node group. For the selected node, and a task transmitting unit 15 for transferring the task. The processing and functions of each of the above means are realized by a program executed by a computer constituting the global load distribution apparatus 1.

In the cluster system targeted by the present invention, K types of nodes having different processing capacities are respectively composed of N (k) (k = 1 to K), and the types of tasks executed on the system are all It is assumed that there are J types. The total number of nodes is denoted as N. That is,
N = N (1) +... + N (K) (1)

  FIG. 2 is a flowchart for explaining the processing procedure of the global load distribution apparatus 1 according to the first embodiment of this invention. With reference to FIG.1 and FIG.2, operation | movement of the global load distribution apparatus 1 of the 1st Embodiment of this invention is demonstrated.

  The task receiving unit 11 receives a task processing request to the global load balancer 1 (step 101 in FIG. 2), determines the type of task (step 102), and measures the arrival rate of the task (step 103).

  The node group constructing means 12 manages a total of N nodes divided into several groups (referred to as “node groups”), and determines which task type is executed in which node group. At this time, using the information obtained from the system configuration input means 13 and the information obtained from the execution performance measurement means 14, the number of nodes included in each group is controlled at the time of execution (step 104) and specified by the user. Operates to satisfy the operating conditions.

  The task transmission means 15 selects one appropriate node from the node group that executes a certain task type (step 105), and transfers the task to the selected node (step 106).

  In order to perform such an operation, the global load distribution apparatus 1 has a node group configuration table 193 (see FIG. 19) that manages the nodes that construct the node group, and a task type that manages a node group that executes a certain task type. A node group correspondence table 194 (see FIG. 14) and a measurement data management table for managing performance data measured at the time of execution are stored in the storage device. This measurement data management table includes a task arrival rate management table 191 shown in FIG. 17 and a node performance management table 192 shown in FIG.

  Furthermore, when the global queuing means 16 is added in the global load balancer 1, tasks are retained on the global queuing means 16 so that the number of tasks staying on each node does not exceed a certain number. The behavior can also be extended to queue

  Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 3 is a diagram illustrating a configuration of a system and a node according to the second embodiment of this invention. Referring to FIG. 3, the second embodiment of the present invention includes a switching device 2 and nodes 3-1 to 3-n. However, n is a certain natural number. The nodes 3-1 to 3-n can communicate via the switching device 2, and the task arrives at the nodes 3-1 to 3-n via the switching device 2. Each of the nodes 3-1 to 3-n has the same configuration, and FIG. 3 shows the configuration of the node 3-1. Each node includes a load distribution unit 31, a load information sharing unit 32, a task transmission / reception unit 33, a task execution unit 34, and a local queuing unit 35. The functions of these means are realized by a program executed by a computer constituting the node, and each operates roughly as follows.

  The task transmission / reception means 33 transmits / receives a task or a task execution result.

  The task execution means 34 processes the requested task and returns the execution result to the local queuing means 35.

  The local queuing means 35 queues the task received by the task transmission / reception means 33 and passes a certain number of tasks to the task execution means 34 from the head of the queue. If the execution result is received from the task execution means 34 or if the execution result has arrived at the task transmission / reception means 33, the corresponding task is removed from the queue. Then, the task transmission / reception means 33 is requested to transmit the execution result.

  The load information sharing means 32 performs an operation equivalent to the information sharing method and apparatus described in the prior application 1. However, a function is added to push information to the following load distribution means 31 when the load information of other nodes is updated.

  The load distribution unit 31 determines which task is transferred to which node based on the load information of the other nodes provided by the load information sharing unit 32 and requests the task transmission / reception unit 33 to transfer the task.

  FIG. 4 is a diagram showing the configuration of the load distribution means 31 of FIG. Referring to FIG. 4, the load distribution means 31 includes a load information management means 36, a node priority assignment means 37, a balance calculation means 38, and a transfer task determination means for determining which task is transferred to which node. 39. Each of these means generally operates as follows.

  The load information management unit 36 receives the load information pushed from the load information sharing unit 32, and reconfigures the information according to a predetermined method. Further, the load information management means 36 answers the inquiry of load information from the following transfer task determination means 39. Further, the load information management unit 36 holds the update of the load information of the node from the transfer task determination unit 39 until the information is pushed from the load information sharing unit 32 regarding the node.

  The node priority assigning means 37 assigns a priority for determining a node from which a load transfer determination is performed according to a predetermined method to a node. However, in the following, it is assumed that the lower the priority value, the higher the priority.

  The balance degree calculation means 38 dynamically calculates the amount of tasks to be executed in the own node according to a predetermined method.

  The transfer task determination unit 39 determines whether to transfer a task from a node having a high priority.

  FIG. 5 is a flowchart for explaining the operation of the second exemplary embodiment of the present invention. Next, the overall operation of the second exemplary embodiment of the present invention will be described in detail with reference to FIGS.

  First, the task is received by the task transmitting / receiving means 33 (step L1 in FIG. 5).

  Next, the local queuing means 35 queues the received task. However, it is possible to determine the return destination of the task execution result at the time of queuing. For example, if the task is transmitted by TCP communication, the correspondence between the task and the connection is recorded at the time of queuing (step L2).

  The load distribution means 31 determines which node should be executed with respect to the tasks in the local queuing means 35, excluding a certain number Q of tasks from the head of the queue (step L3).

  Based on this determination, the task transmitting / receiving unit 33 transfers the task to another node (step L4).

  The Q tasks from the head of the queue that have not been transferred to other nodes are executed by the task execution means 34 (step L5).

  If the execution result of the task transferred to the other node is returned, the task transmission / reception means 33 receives this (step L6).

  The local queuing means 35 accepts the execution result from the task execution means 34 or, if the execution result is received by the task transmission / reception means 33, removes the corresponding task from the queue and uses the task transmission / reception means 33 to obtain the execution result. Transmit (step L7). However, the load information sharing unit 32 pushes load information to the load distribution unit 31 as appropriate.

  FIG. 6 is a flowchart for explaining the operation of the load distribution means 31 of the present embodiment. With reference to FIG. 4 and FIG. 6, the operation of the load distribution means 31 of this embodiment will be described in detail below.

  First, the threshold value (threshold) which is the upper limit of the load of the own node which can be permitted is calculated using the balance degree calculation means 38 (step A1 in FIG. 6).

  Next, the node priority assigning means 37 is used to assign priorities to other nodes (step A2). Hereinafter, a node having a high priority is referred to as a “close node”, and a node having a low priority is referred to as a “far node”.

  Then, the transfer task determination unit 39 calculates the amount “over” (over) of the load to be transferred. “Over” is the difference between the load of the own node and the smallest integer equal to or greater than the threshold “threshold”. Further, the load amount “under” (under) that can be accepted by the load transfer target node is initialized to “0” (step A3).

  The transfer task determination means 39 determines whether or not the load amount “over” to be transferred is a positive value, and if it is not a positive value, the transfer task determination means 39 ends without performing load distribution (step A4).

  If the amount “over” of the load to be transferred is a positive value, the transfer task determining means 39 selects the node that performs load transfer determination in order from the closest node (step A5). Hereinafter, the selected node is referred to as a “load transfer target node”.

  If load transfer determination has already been made for all nodes, there is no node that can be selected, and load distribution is terminated (step A6).

  The transfer task determination unit 39 adds the difference between the threshold “threshold” and the load information (obtained by inquiring of the load information management unit 36) of the load transfer target node to the acceptable load amount “under” ( Step A7).

  If the added acceptable load amount “under” is not “1” or more, the process returns to step A5 to change the load transfer target node (step A8).

  If the added acceptable load amount “under” is “1” or more, it is determined to transfer the load by “1” to the load transfer target node (step A9).

  Then, the load information management means 36 is used to decrease the load information of the own node by “1”, increase the load information of the transfer target node of the own node by “1” (step A10), and terminate the load distribution.

  Next, the effect of the 2nd Embodiment of this invention is demonstrated.

In the second embodiment of the present invention,
“Over” means the amount of load to be transferred in order to eliminate the overload state of the own node.
“Under” means the amount of load that the node from the nearest node to the load transfer target node can accept without being overloaded.

  Therefore, to transfer the load to the closest node (node α) in which “under” is positive, the node between the own node and the node α is transferred by load transfer between those nodes. This means taking into account load balancing. In this way, the load transfer of the own node is determined in consideration of the load transfer of the other nodes, so that the crowd effect can be prevented.

  Next, a third embodiment of the present invention will be described in detail with reference to the drawings. The basic configuration of the third embodiment of the present invention is the same as the configuration of the second embodiment described with reference to FIGS.

  The third embodiment of the present invention is an operation for determining a task transfer destination in FIG. 6 showing the processing procedure of the load distribution means 31 (see FIG. 4) with respect to the second embodiment. 7 in that steps B1 to B5 in FIG. 7 are executed instead of A10 which is an operation of updating the load information.

  When “under” is “1” or more and “over” is “under” or less, the transfer task determination unit 39 determines to transfer the load of “over” to the load transfer target node. (Step B2 in FIG. 7).

  Then, using the load information management means 36 (see FIG. 4), the load information of the own node is reduced by “over”, and the load information of the transfer target node possessed by the own node is increased by “over” (step B3). Load balancing ends.

  If “under” is equal to or greater than “1” and the value of “over” is larger than the value of “under”, the transfer task determination means 39 determines the amount of floor (under) (= the maximum integer equal to or less than under). Is determined to be transferred to the load transfer target node (step B4).

  Then, the load information management means 36 is used to reduce the load information of the own node and “over” by “floor (under)”, and increase the load information of the load transfer target node of the own node by “floor (under)”. . Further, “floor (under)” is subtracted from “under” (step B5). Then, the process proceeds to Step A5.

  Next, the effect of the 3rd Embodiment of this invention is demonstrated.

  In the third embodiment of the present invention, the load of “under” is transferred to the node (node α) in which “under” becomes positive. This means that the node α and a node closer to the node α are not overloaded, and therefore load transfer of the own node is performed in consideration of load transfer of other nodes. Conventionally, performing a plurality of task transfers does not improve performance because the crowd effect increases, but in this embodiment, the crowd effect can be mitigated by considering the task transfer of other nodes. The load can be balanced at high speed even if the node is transferred to a plurality of nodes.

Next, a fourth embodiment of the present invention will be described in detail. The fourth embodiment of the present invention has the same configuration as that of the third embodiment, but the load information sharing means 32 (see FIG. 3) and the load information management means 36 of the load distribution means 31 (see FIG. 4) is different. That is, the fourth embodiment of the present invention
(I) The load information sharing means 32 (see FIG. 3) provides information with different accuracy to the load distribution means 31,
(II) The load information management means 36 (see FIG. 4) of the load distribution means 31 is configured to reconstruct the load information from the value and accuracy of the load information. In (I) and (II) above, This is different from the third embodiment.

  Next, the effect of the 4th Embodiment of this invention is demonstrated.

  In the fourth embodiment of the present invention, the load information is reconstructed from the value and accuracy of the load information sent from another node, so that the load information can be shared at low cost. Even if it is based on load information with different accuracy for each node, considering the task transfer of other nodes, it is configured to transfer multiple surplus loads to multiple nodes at low cost. Nevertheless, the load can be balanced at high speed.

  Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings. The fifth embodiment of the present invention is different from the third and fourth embodiments of the present invention in that the load distribution means 31 (see FIGS. 3 and 4) has a response time table (FIG. 8). See).

  In the local queuing means 35 (see FIG. 3), the response time of the task is measured, the result is reflected in the response time table, and the number of tasks executed simultaneously at the time of execution is changed with reference to the response time table. Control is performed.

  Referring to FIG. 8, the response time table 310 records the total of response times and the number of measurements (the number of response times obtained by summing up). These two values are recorded according to the number of executions (the number of tasks executed simultaneously). For example, when three tasks are being executed and one of them is completed, the response time is treated as when the number of executions is “3”.

  Next, the operation of the local queuing means 35 in the present embodiment will be described in detail with reference to FIG. 5 and FIG.

  The local queuing unit 35 holds, for each task, the time when the task is transferred to the task execution unit 34 in step L5 in FIG. Further, in step L6 in FIG. 5, every time the execution result is received, the operation shown in FIG. 9 is added.

  First, when the local queuing means 35 receives the execution result from the task execution means 34, the local queuing means 35 obtains the difference between the current time and the task execution start time held in step L5 (step C1 in FIG. 9).

  Next, the response time obtained in step C1 is added to the sum of the response times of the number of executions obtained by adding “1” to the current queue length, the corresponding measurement number is increased by “1”, and the response time table 310 is updated. (Step C2).

  Further, referring to the response time table 310, an average response time is obtained by dividing the sum of s or more response times by the corresponding number of measurements, and a desired number of executions R is obtained from the value of these average response times ( Step C3). However, s is a predetermined value and is referred to as an “effective sampling number”.

  Then, the maximum value Q of the number of tasks executed simultaneously is changed to (R + v) (step C4). However, v is a predetermined non-negative integer and is referred to as “additional execution number”.

  Next, the effect of the 5th Embodiment of this invention is demonstrated.

  In the fifth embodiment of the present invention, the response time is measured according to the number of tasks to be executed simultaneously, and the maximum number of tasks to be executed simultaneously is determined dynamically. Therefore, the throughput of the node can be improved by multiple executions, and the load imbalance can be traded off by limiting task transfer (tasks that have been executed once are not transferred), thereby improving the overall system throughput.

  Next, a sixth embodiment of the present invention will be described in detail with reference to the drawings.

  Referring to FIG. 10, unlike the fifth embodiment, the sixth embodiment of the present invention includes a global load balancer 1 instead of the switching device 2 (see FIG. 3). . The global load balancer 1 is configured as shown in FIG. 1, but the processing of the node group construction unit 12 is partially different. The node is configured as shown in FIG. 3, but the load information sharing means 32 is different.

  Unlike the first embodiment, the global load balancer 1 is configured such that, in step 104 of FIG. 2, the node group construction unit 12 changes the node group to each node included in the group to which the change is applied. An operation of notifying the load information sharing means 32 is performed.

  Further, unlike the fifth embodiment, the load information sharing means 32 in the present embodiment receives a notification of a node group change so as not to share information with nodes belonging to other node groups. Perform the action. Therefore, in this embodiment, the load is not transferred across the node groups.

  Next, the effect of the 6th Embodiment of this invention is demonstrated.

  According to the sixth embodiment of the present invention, a task is executed in a different node group for each task type, and is configured so that task transfer between node groups does not occur. Even if a large number of sudden occurrences occur, the execution of other task types is not affected, and a predetermined response time can be satisfied.

  Moreover, since load balancing within the node group is achieved at high speed, a predetermined response time can be satisfied even if the node group is configured with a smaller number of nodes.

  Furthermore, immediately after adding a node to an overloaded node group, it is possible to shift to a load balanced state, so that the timing of adding a node can be delayed. As a result, it is now overloaded, but it is possible to reduce inappropriate node addition judgment that has become a low load when adding a node, and improve the throughput of the system.

  Next, in order to describe the above-described embodiment in more detail, a description will be given according to specific examples.

  FIG. 1 is a diagram showing the configuration of the first exemplary embodiment of the present invention. The task arriving at the global load balancer 1 is received by the task receiving unit 11, and the task receiving unit 11 determines the type of task. Here is an example of how to determine the type of task. When HTTP (HyperText Transport Protocol) is used as the protocol, the task type is clearly indicated in the first line (request line), request header, and request body of the HTTP request. The type can be determined. In addition, it is possible to distinguish different tasks from different protocols. Alternatively, it can be determined that the task is different when the port number to be used is different. In this way, the task receiving means 11 determines the type of task.

  Further, the task receiving means 11 counts the number of tasks that arrive within a predetermined measurement time interval for each task type, calculates a task arrival rate (the number of tasks that arrive per unit time), and obtains a task arrival rate. The measurement data is managed in the management table 191 (see FIG. 17). Each time the measurement interval elapses, the task receiving unit 11 calculates a task arrival rate and updates the task arrival rate management table 191.

  The task arrival rate management table 191 stores not only the latest task arrival rate but also the previous measurement value and the previous measurement value, and stores a predetermined number (= L) of measurement data. To do.

  The system configuration input means 13 is measured in advance, the type K of nodes constituting the cluster system, the node number N (k) of the k-th node type, the number J of types of tasks executed on the system. It accepts input of task characteristics and system operating conditions that determine under what conditions the system is operated.

In addition,
N = N (1) + N (2) +... + N (K) (2)
And

  FIG. 11A and FIG. 11B are diagrams showing one specific example of task characteristics. FIG. 11A shows the relationship between the task arrival rate (horizontal axis) and the throughput (vertical axis) when one type of task processing is performed on one node, and FIG. Is a diagram showing the relationship between the task arrival rate (horizontal axis) and the response (vertical axis) (the characteristics of throughput, response, and task arrival rate shown in FIG. Called "TN characteristics").

  One specific example of the system operation condition is an allowable value (allowable response) as the task processing time and a limit arrival rate that defines a limit range in which the system operates so as to achieve such an allowable response.

  If the task arrival rate is less than a predetermined limit arrival rate, the node group construction unit 12 in FIG. 1 performs group configuration so that the system achieves an acceptable response. However, when the task arrival rate exceeds the limit arrival rate, it is regarded as being out of the guaranteed range, and the system does not perform a group configuration that achieves an acceptable response.

  FIG. 12 shows an example of input. Referring to FIG. 12, for the jth task type and the kth node type, an allowable response time R (“0.11” in the figure) and a marginal arrival rate Λ_Lim (“1” in the figure). .99 ") is specified, and the task arrival rate to one node in a table format and the relationship between the throughput and the response time (referred to as" task-node characteristics "or" TN characteristics ") are given. It has been.

  Such a table is acquired by conducting an experiment for measuring task processing performance using one node at a stage before the system is actually operated. Alternatively, by using the single profile method, it can be calculated from the CPU usage time and disk access time required for executing one task (refer to the description of Non-Patent Document 5 for details).

  In FIG. 12, one row of the TN characteristic is an average value of throughput and response time obtained when a large number of task processes are performed on one node with a task arrival rate given to one node being constant. is there.

  Since there are J types of tasks to be processed in the system and K types of nodes constituting the system, input data as shown in FIG. 12 is prepared for combinations of j = 1 to J and k = 1 to K. Is done.

  The execution performance measuring means 14 in FIG. 1 measures the processing performance of individual nodes during operation. Specifically, the number of tasks transferred to the node during the measurement interval (number of transfers), the number of tasks completed on the node (number of processes), and the time required for task processing (response time) Measure. The measurement interval is usually several hundred milliseconds to several seconds.

  The throughput of the node can be obtained by dividing the number of tasks that have been processed by the measurement interval.

  In addition, the performance measurement value at the immediately preceding measurement interval, the previous value, etc. are also managed, and a predetermined number (= L) of measurement data is stored. In order to manage such values, a node performance management table 192 as shown in FIG. 18 is used. The node performance management table 192 has columns of response, throughput, number of transfers, and number of processes for the node ID, the average value of each value for the entire node is recorded, and the past (L times before) measurement is recorded. From the current measurement.

  The node performance management table 192 is created for each task type and each node type. Here, the number of remaining tasks can be calculated by using the number of transfer tasks and the number of completed tasks. That is, the relationship of the following formula (3) is established.

(Number of staying tasks) = (Number of staying tasks in the previous measurement interval)
+ (Number of transfer tasks)
-(Number of completed tasks) (3)

  Therefore, in this embodiment, it is not necessary to directly measure the number of staying tasks.

  The node group construction unit 12 groups N nodes in total, and manages which task type is executed in which node group. In order to manage information necessary for this, the node group configuration table 193 (see FIG. 19) and the task type / node group correspondence table 194 (see FIG. 14) are updated.

  As illustrated in FIG. 19, the node group configuration table 193 is a table in which a node group number and a node number configuring the node group are associated with each node type pair. That is, the node group number, node number, and node number information of the nodes constituting the node group is stored for the node types 1, 2,.

  As shown in FIG. 14, a task type / node group correspondence table 194 is a table for managing a node group that executes the j-th task type, and includes a correspondence table of task IDs and node group numbers. In the example shown in FIG. 14, tasks with task IDs 1 and 3 are executed with node group number 1 and constitute tasks of task group A, and tasks with task ID 2 are executed with node group number 2. Tasks of task group B are configured.

  When updating these two tables (node group configuration table 193, task type-node group correspondence table 194), the node group construction means 12 receives the input operating conditions, TN characteristics, and task arrival rate. The management table 191 and the node performance management table 192 are referred to.

  In the following, there is one node type (K = 1, that is, when all nodes have the same processing capability), and one task type is associated with one node group (that is, J Node group configuration table 193 and task type-node group correspondence table 194 are updated with respect to the case where there are n node groups and the jth task type is executed on the jth node group) The specific method to do is demonstrated in detail. For convenience of explanation, the following symbols are defined.

  Let R (j) be an allowable response given as input data of the j-th task type, and Λ_Lim (j) the limit arrival rate.

  The maximum throughput achievable at one node is T_max (j), and the allowable arrival rate at one node is λ (j).

  Here, the allowable arrival rate λ (j) is a task arrival rate to one node that gives an allowable response R (j).

  The maximum throughput T_max (j) that can be achieved by one node and the allowable arrival rate λ (j) of one node are expressed as follows using the tabular TN characteristics as shown in FIG. Ask for it.

  As the maximum throughput T_max (j), the maximum throughput value is selected from the input TN characteristics and set as the value.

  The allowable arrival rate λ (j) of one node is obtained using the relationship between the arrival rate and the response and the allowable response time.

  As for the arrival rate and the response, a data set at the measurement point is given as input data.

  By interpolating such a plurality of measurement points by an appropriate method, the response time can be expressed as a function of the arrival rate, or conversely, the arrival rate can be expressed as a function of the response time.

  For details of the interpolation, see Non-Patent Document 6 ("Common sense of numerical computation" (by Masao Iri and Kazunori Fujino), Kyoritsu Shuppan (1985), Chapter 15 Interpolation (interpolation) and Chapter 16 Clouds. The ruler and spline-continuous interpolation) are described in detail. In this embodiment, details of interpolation are omitted.

  As described above, the relationship between the arrival rate and the response time is approximately expressed as a function, whereby the arrival rate λ (j) that gives the allowable response time R (j) can be obtained.

  As described above, FIGS. 11A and 11B are graphs showing the TN characteristics. The graph of task arrival rate and response shows that if the arrival rate can be calculated as a function of response time, the allowable arrival rate λ (j) of one node can be easily obtained from the allowable response R (j). Yes.

  When the task type executed on the j-th node group is only the j-th task type, the allowable response R (j) is given when the arrival rate Λ (j) of the j-th task type is given. In order to satisfy the above, a function N_h (j) (Λ (j)) for calculating the number of nodes included in the j-th node group (referred to as “allowable node number”) is as follows.

  The arrival rate Λ (j) of the task type j is a value managed in the task arrival rate management table 191. In this table, the latest L measurement results are managed, and a weighted average value of the latest L measurement results can be used as Λ (j).

  If the number of allowable nodes is n (j), it is considered that the allowable response is satisfied if the arrival rate for one node is equal to or less than the allowable arrival rate λ (j).

That is,
λ (j) ≧ Λ (j) / n (j) (4)
N (j) may be determined so that.

Therefore,
n (j) ≧ Λ (j) / λ (j) (5)

Considering that the number of nodes is an integer value and the minimum required value,
N_h (j) (Λ (j)) = Ceil (Λ (j) / λ (j)) (6)
It is. Here, Ceil (x) represents the smallest integer not smaller than x.

  When the task arrival rate Λ (j) is larger than the limit arrival rate λ_Lim (j), it is necessary not to allocate more nodes than necessary to the jth node group.

  Considering such conditions, N_h finally becomes a function as follows.

    N_h (j) (Λ (j)) = Ceil (min (Λ (j), Λ_Lim (j)) / λ (j)), (7)

  Here, min (a, b) is a function that returns the lesser of a and b.

  The permissible arrival rate λ (j) of one node and the function N_h (j) (Λ (j)) that gives the number of permissible nodes obtained above are obtained at the stage of analyzing the input data and are executed. There is no particular need to use values that are sometimes measured.

  Next, a procedure (procedure P) for calculating the number of nodes allocated to J tasks according to the task arrival rate Λ (j) that changes during execution will be described with reference to the flowchart of FIG. . This procedure is periodically executed on the node group construction means 12 of FIG. 1 and includes the following steps.

From J tasks where j = 1 to J,
Λ_Lim (j)> Λ (j) (8)
The task type that satisfies (that is, the arrival rate is less than the allowable arrival rate) is defined as task group A,
The task type that is not satisfied is defined as task group B (step 122).

  For the task type j of task group B, the number of assigned nodes derived above is calculated (step 123).

  A value obtained by the function N_h (j) (Λ (j)) is assigned as the number of nodes. Here, the total number of nodes assigned to the task type j of the task group B is Nb.

The remaining Na = N-Nb nodes are appropriately assigned to task type j of task group A (step 124). An example of such an allocation method is an allocation method in which the virtual usage rates of the node groups are balanced. Here, the virtual usage rate of the node group is as follows. The throughput of task type j is Φ (j), and the number of nodes in the jth node group is n (j). At this time, the maximum throughput that can be achieved by n (j) nodes is the maximum throughput T_max (j) of one node.
n (j) × T_max (j) (9)
It can be expressed as.

Ratio of the current throughput to the maximum value of this throughput Φ (j) / (n (j) × T_max (j)) (10)
Is a virtual usage rate of the j-th node group.

  By determining n (j) so that the virtual usage rate of the node group calculated in this way becomes a substantially constant value with respect to the task type j included in the task group A, an appropriate number of allocated nodes can be determined. Can be calculated.

However,
n (1) +... + n (J) = Na (11)
It is.

Regarding the number of nodes n (j) obtained in step 124 above,
n (j) ≧ N_h (j) (Λ (j)) (12)
Is checked (step 125).

  The value calculated by N_h (j) (Λ (j)) is the minimum necessary number of nodes. If such a condition is not met, it means that an illegal assignment has occurred, so the system administrator is notified that there are not enough nodes to be assigned.

  Before executing the procedure P, the number of nodes assigned to each node group j is NG (j), and the number of assigned nodes obtained by executing the procedure P is NG ′ (j). However, j = 1 to J are node group numbers.

Here, the change in the number of nodes is
ΔNG (j) = NG ′ (j) −NG (j) (12)
It is determined.

In order to realize the obtained node assignment NG ′ (j), j = 1 to J,
ΔNG (j) <0 (13)
Select one node from the node group j such that and delete it from the node group,
ΔNG (i)> 0 (14)
It is preferable to repeat the procedure of adding the node to the node group i such that the node allocation number for the node group finally satisfies NG ′ (j).

Here, as a method of selecting one node from the node group j,
・ Random selection method or
-A method of selecting the node with the minimum load by using an index such as the number of staying tasks, average response time, or average throughput as the load can be used.

Therefore, in this embodiment, from the node group configuration table 193 (see FIG. 19),
ΔNG (j) <0 (15)
A node group j is selected, and one node is selected from the j rows in the node group configuration table 193 using the method described above, and the node ID is deleted from the table.

Furthermore, from the node group configuration table 193,
ΔNG (i)> 0 (16)
The node group i is selected, and the deleted node ID is added to the i row in the table.

  The task transmission means 15 refers to the task type / node group correspondence table 194 (see FIG. 14) and transfers the task belonging to the task type j to the node belonging to the task group j.

  When there are a plurality of nodes belonging to the task group j, it is necessary to select one of the plurality of nodes to which the request is transferred.

As a method of such node selection,
・ Round robin method to select multiple nodes in sequence,
・ Random selection method that randomly selects one node from a plurality of nodes,
・ Minimum request node selection method to select the node with the smallest number of requests staying on the node,
・ Minimum response selection method to select the node with the shortest response time,
・ Maximum throughput selection method to select the node with the highest throughput,
Either of these can be used.

  For performance values such as the number of staying requests, response time, and throughput required for each node, values managed in the node performance management table 192 (see FIG. 18) may be referred to.

  In addition, since the value of the node performance management table 192 includes the latest L measurement results, not only the latest measurement result but also a weighted average of L measurement results can be used.

In this embodiment, a function N_h (j) for obtaining the number of nodes assigned to the task type j is used.
When the arrival rate Λ (j) of the task type j is less than the limit arrival rate Λ_lim (j),
Assign Ceil (Λ (j) / λ (j)) nodes,
When the arrival rate Λ (j) of the task type j is greater than or equal to the limit arrival rate Λ_lim (j),
It operates to allocate Ceil (Λ_lim (j) / λ (j)) nodes.

  By operating as described above, according to the present embodiment, a limit is imposed on the number of nodes assigned to a task type having an arrival rate that is equal to or higher than the allowable arrival rate, and a sufficient amount of nodes is provided for other task types. There is an effect of being assigned.

  In the above embodiment, tabular data as shown in FIG. 12 is assumed as the TN characteristic accepted by the system configuration input means 13.

  However, if the single profile method described above is used, the throughput and response time for the request arrival rate can be calculated using the utilization time of resources in the node (CPU and disk) required to execute one task. Can do. In such a case, as the input data, the CPU consumed when executing one task and the disk usage time are required.

  In the above embodiment, a condition that one task type is associated with one node group is provided.

  In the following, an embodiment in which this condition is relaxed and a plurality of task types are associated with one group will be described.

  The node group construction means 12 divides all nodes constituting the system into, for example, two node groups (node group 1 and node group 2). In the node group 1, the task type belonging to the task group A is executed, In group 2, control is performed to execute the task type belonging to task group B (task groups A and B can be configured by step 122 of procedure P).

  As a data structure for this purpose, a task type / node group correspondence table 194 (see FIG. 14) is used.

  In FIG. 14, since the arrival rate Λ of the task type (task ID) 2 is equal to or greater than the allowable arrival rate Λ_lim, it is classified into the task group B, and the arrival rates of the task types 1 and 3 are less than the allowable arrival rate. The example of the correspondence table when task types 1 and 3 are assigned to node group 1 and task type 2 is assigned to node group 2 when classified into node group A is shown.

  As the number of nodes allocated to the node group B, Nb calculated in step 123 of FIG. 13 is allocated, and as the number of nodes allocated to the node group A, N−Nb is allocated (where N is the total number of nodes). Show).

  The task transmitting unit 15 (see FIG. 1) refers to the task type / node group correspondence table 194 (see FIG. 14) and determines a node group to which a task belonging to the task type j is transferred.

  In the above embodiment, it is assumed that the processing capabilities of all nodes are equal.

  In the following, an embodiment will be described in which nodes have different processing capabilities, that is, nodes are classified according to processing capabilities, and there are K types of nodes in the entire system.

  The number of node types k assigned to the node group j is expressed as N (j, k).

  An allowable arrival rate of one node is given for each node type, and an allowable arrival rate of task type j for node type k is expressed as λ (j, k).

  Furthermore, a situation is assumed where task type j is executed on node group j.

  In this case, instead of the function N_h (j) for obtaining the number of nodes to be assigned to the j-th node group used in the above embodiment, the number of nodes to be assigned is determined according to the following procedure. This procedure is obtained by replacing step 123 of FIG. 13 with steps 1230 to 1232 of FIG. 15 and step 124 of FIG. 13 with 1241 to 1242 of FIG.

Referring to FIG. 15, in step 1230,
The following two steps (1231 and 1232) are executed for all task types j in task group B.

In the next step 1231,
All N ′ (j, k) (k = 1 to K) pairs of nodes satisfying the following conditions are obtained.

    Λ_Lim (j) ≧ N ′ (j, 1) λ (j, 1) + N ′ (j, 1) λ (j, 1) +... + N ′ (j, K) λ (j, K) (17)

In the next step 1232,
The value (distance) of the following equation (18) calculated using the set N (j, k) of the number of nodes immediately before the execution of the procedure is the smallest among the obtained sets N ′ (j, 1) of the number of nodes. Let N (j, k) be a set of new allocation node numbers.
| N ′ (j, 1) −N (j, 1) |
+ | N ′ (j, 2) −N (j, 2) |
+ ...
+ | N ′ (j, K) −N (j, K) | (18)

  For the node type k, the sum of N (j, k) is Nb (k) (where j is a task type belonging to task group B).

Referring to FIG. 16, in step 1241,
The number of nodes that can be assigned to the task type of task group A is calculated for each node type k. That is,
Na (k) = N (k) -Nb (k) (k = 1 to K) (19)
And

In step 1242,
Na (k) is appropriately assigned to each task type j of task group A. An example of such an allocation method is an allocation method in which virtual usage rates for the k-th node type constituting the node group are balanced. Such conditions are specifically described as follows.

  The throughput of the task type j when using the node type k is Φ (j, k), and the number of nodes of the node type k of the jth node group is n (j, k).

At this time, the maximum value of the throughput that can be achieved by n (j, k) nodes is the maximum throughput T_max (j, k) that can be achieved by using one node of the node type k.
n (j, k) × T_max (j, k) (20)
It can be expressed as.

The ratio of the current throughput to the maximum throughput is β (j, k) = Φ (j, k) / (n (j, k) × T_max (j, k)) (21)
And

  At this time, when β (j, k) takes substantially the same value for j = 1 to J, the virtual usage rate for the k-th node type constituting the node group is called “balance”. By determining n (j, k) so as to satisfy such a condition, the optimum number of allocated nodes can be calculated.

However,
n (1, k) +... + n (j, k) = Na (k) (22)
It is.

  Next, an extension of the embodiment in the case where the global load distribution apparatus 1 has the global queuing means 16 will be described. In this case, when the task transmitting unit 15 transmits a task to a node, and the number of staying tasks on the destination candidate node exceeds a certain value, the node is referred to as a transmission target node. Without selecting another destination candidate node. When the number of staying tasks on all nodes that have risen as destination candidates exceeds a certain value, the task transmission unit 15 stores the task on the global queuing unit without transmitting the task to the node. .

Such a task is transmitted on a node when task processing on that node is completed. Note that the number of staying tasks used for determining whether to transmit a task is a value calculated using Little's formula from the allowable response R (j) and the allowable arrival rate λ (j). ,
(J) × λ (j) (23)
Can be considered. Or, expect some room,
α × R (j) × λ (j), (24)
However, 0 ≦ α <1 may be satisfied.

  Next, a second embodiment of the present invention will be described with reference to the drawings. This example corresponds to the second embodiment of the present invention. However, the following premise (premise 1) is set first.

Assumption 1:
Arriving tasks are processed in a cluster of n nodes with absolute node IDs from 0 to n-1.

  A relative node ID is defined for the node. The relative node ID is calculated such that, for example, when viewed from the node i (i is an integer not less than 0 and not more than n−1), the node i + 1 is a node whose relative node ID is “1”.

The formula for calculating the relative node ID of node j as seen from node i is:
(i−j + m) mod m (25)
It becomes. However, m is a natural number. Hereinafter, simply speaking, a node ID indicates a relative node ID.

  The load information is information regarding a resource that becomes a bottleneck, and here, the queue length of the local queuing means is used as the load information. Although the load information is represented by a real number, the load transfer unit is an integer.

  Of the tasks connected to the queue, only a fixed number of tasks are executed from the top, and the tasks that can be transported are those that have not yet been executed. However, it is assumed here that only the first task from the top of the queue can be executed.

  In this embodiment, the load information sharing means 32 shares information of all other nodes, and the node priority assignment means 37 assigns a node ID as a priority (the smaller the value, the higher the priority). ).

  The balance degree calculation means 38 (see FIG. 4) of the load distribution means 31 calculates the smallest integer equal to or greater than the average value of the loads of all nodes.

  In this embodiment, the load distribution means 31 operates according to the flowchart of FIG.

FIG. 20 shows an example of the operation of this embodiment in a cluster of 8 nodes. FIG. 20 shows a table format for each load transfer target node.
・ Initial value of load information of target node (initial load info.)
-Node load information after modified task transfer (modified load info.),
・ Under value before task transfer judgment (under (b)),
・ Under value (under (a)) after task transfer judgment,
-Over value after task transfer judgment (over)
It is shown.

  Node 0 (own node) recognizes that its own node is overloaded because the difference between its load (= 5) and the threshold threshold (= “3”) calculated from the load information is positive. , Start load transfer.

  Then, the node 0 searches for the destination node in order from the node 1 having the highest priority.

  When the difference between the load information of the load transfer target node and the threshold “threshold” is added to “under” in order, “under” becomes “1” or more for the first time in node 6, so One task is transferred, the load information of the own node and the node 6 is updated, and load distribution ends.

  The above is the operation of node 0, but it should be noted that other nodes also perform the same operation. In FIG. 20, the task transfer from the overloaded node 1, node 2, and node 3 is also indicated by broken lines.

  Assuming the same load state, the node L, which is the lightest node among all nodes, is compared with the load of its own node, and the difference is Δ (here, Δ = 1). In addition, when tasks are transferred to the node L one by one, four tasks are transferred to the node 6 to generate a crowd effect.

  In this embodiment, the crowd effect does not occur as shown in FIG.

For three nodes x, y, z
If the priority that x assigns to y is higher than the priority that x assigns to z,
The priority y assigns to z is higher than the priority y assigns to x,
The priority that z assigns to x is higher than the priority that z assigns to y.
In the present specification (explanation part) and claims, priority assignment such as follows is referred to as “circular monotonically increasing priority assignment”. In the present embodiment, cyclic monotone priority assignment is established for any three nodes.

  Therefore, when performing load transfer to a certain node, the crowd effect can be prevented by considering whether load transfer by a node closer to that node occurs.

  On the other hand, if the priority among some of the three nodes does not increase in a cyclic monotonous manner, the crowd effect will occur.

  How much crowd effect occurs depends on the load state and the number of combinations of three nodes that cannot be assigned to a monotonically increasing priority.

  Next, a third embodiment of the present invention will be described with reference to the drawings. This example corresponds to the third embodiment.

  In this embodiment, the above-mentioned “Premise 1” is assumed. Like the second embodiment, the load information sharing means 32, the node priority assignment means 37, and the balance degree calculation means 38 are provided, but the transfer task The determination means 39 operates by replacing step A9 in FIG. 6 with steps B1 to B5 in FIG.

  FIG. 21 shows an operation example of this embodiment in a cluster of 8 nodes. The table shown in FIG. 21 is the same as that in the second embodiment.

  Unlike the operation in the second embodiment, the node 0 transfers a task by “under” to a node having a load amount “under” of 2 or more.

  Then, when all nodes are selected as transfer target nodes or the overload state is resolved (“over” becomes “0”), the load distribution ends.

  Therefore, when a node whose “under” is initially “1” or more cannot accept all overloaded tasks, the task is transferred to a plurality of nodes. In FIG. 21, since “under” is 3 in the node 2, three tasks are transferred to the node 2. However, since “over” is still “1” or more, a task transfer destination node is further searched, and two tasks are transferred to nodes 3 and 4 and one task is transferred to node 7.

  Assuming the same load state, comparing the load of the node L, which is the lightest node among all nodes, with the load of its own node, the difference is Δ (here, Δ = 1), When transferring tasks one by one to the node L, the overload state of the node 0 is hardly eliminated even after the end of load distribution.

  In this embodiment, since the transfer of a plurality of tasks is permitted, the load is balanced after the end of load distribution. Usually, when a plurality of tasks are moved, the crowd effect is strengthened, but in this embodiment, the crowd effect is not generated. This is because the task transfer of the own node is determined in consideration of the task transfer of other nodes.

  Next, a fourth embodiment of the present invention will be described with reference to the drawings. This example corresponds to the fourth embodiment.

  In the present embodiment, load sharing is performed when load information with reduced spatial accuracy is shared under assumption 1.

  Spatial accuracy indicates how many nodes of load information each node represents.

  For example, if an average value of a load of a certain four nodes is transmitted as the load information and this average value is used as the load information of these nodes, the spatial accuracy of this information is ¼. That is, the spatial accuracy is defined by the following equation (26).

    Spatial accuracy = 1 / (number of nodes represented by the information) (26)

In the load information sharing means 32 (see FIG. 3) of this embodiment, each node
The load on node 0 (own node),
The average load transmission node g ₁ to g ₂ -1 from node g _1,
The average load of the nodes g ₂ to g ₃ −1 transmitted from the node g ₂ ,...
With an average load of the node g node sent from _{k-1 g k-1 ~g} k -1 as information.

However,
g _k-1 <n <= g _k and m = g _k
And
When i <j
g _{i + 1} −g _i <= g _{j + 1} −g _j
Holds.

G _i is a statically determined value. The load information is periodically updated (this information sharing method is realized by the method of the prior application 1).

  Calculation of the threshold threshold by the balance degree calculation means 38 in step A1 of FIG. 6 is performed as follows. However, load (i) is the load information of the node i possessed by the own node.

threshold = {load (0) + Σ _{i = 1} ^k-2 load (g _i ) × (g _{i + 1} -g _i ) + load (g _k-1 ) × (ng _k-1 )} / n
... (27)

However, the term beginning with the operation symbol Σ _{i = 1} ^k−2 represents the sum of load (g _i ) × (g _{i + 1} −g _i ) from i = 1 to i = k−2.

  Further, in the determination of the inter-node distance in step A2 in FIG. 6, the node priority assigning means 37 assigns a node ID as in the third embodiment. The transfer task determination means 39 operates in the same manner as in the third embodiment.

Only the information on the node g _i is pushed from the load information sharing means 32, and the information on the nodes g _i +1 to g _{i + 1} −1 is not pushed.

Therefore, when the load information management means 36 receives an information inquiry about the nodes g _i +1 to g _{i + 1} −1, it basically reconstructs the load information by returning the information on the node g _i. To do. This operation is referred to as “simple space accuracy complement” in the present specification (explanation part) and claims.

However, for information has been updated nodes from the transport task determining means 39, after the update, unless information of the node g _i is pushed, returns the information updated by the transport task determining means 39.

  An example of operation in a cluster of 8 nodes according to this embodiment will be described. First, FIG. 22 shows an information sharing method used in the operation example.

In the example shown in FIG.
Node 0 is
・ From node 1, information on node 1
-From node 2, the average of the information of node 2 and node 3 is
・ Information on node 4 and node 7 is received from node 4.

  Node 2 can transmit the average of the information of nodes 2 and 3 because node 2 has received the information of node 3 from node 3 and node 4 can transmit the average of the information of nodes 4 and 7 This is because the information of the node 5 is received from the node 5 and the average of the nodes 6 and 7 is received from the node 6.

  FIG. 23 is a diagram illustrating an example of the operation of the present embodiment. What is shown in the table format at the bottom of FIG. 23 is the same as in the second embodiment.

  The node 0 (own node) transfers the task of the load amount “over” (= “3”) to another node from the threshold threshold (= “2.5”) calculated from the load information. Then, the destination node is searched in order from node 1. As the node ID of the node is younger, the spatial accuracy is higher.

  By doing so, the fact that the load imbalance can be solved using only the nearby nodes leads to the fact that the load imbalance can be solved using only highly accurate information, and more efficient load distribution becomes possible.

  Furthermore, even when a load is thrown to a distant node, the influence of a decrease in information accuracy can be suppressed. This is because each node performs load balancing more accurately with nearby nodes, so that it is basically possible to consider the loads of nodes far away as the same.

  In the example shown in FIG. 23, since simple space accuracy complementation is performed, the loads on the nodes 2 and 3 are regarded as “1.5”, and the loads on the nodes 4 to 7 are regarded as “2”.

  The actual load on each node varies. For example, the load on the node 4 is “4”, but the load on the node 5 is “0”. However, since the node 4 also performs load balancing and transfers tasks to the node 5, this difference is alleviated. Therefore, the node 0 can transfer tasks by assuming that these nodes have the same load.

Assuming that the number of nodes is “n” and the information update period is “p”, the cost of the load information sharing method used in this embodiment is O (pn log n). Cost is lower than O (pn ² ).

  Nevertheless, in this embodiment, the load distribution can be performed relatively accurately in the present embodiment. In this embodiment, the information on the nearby nodes is relatively accurate information, and priority is given to the nearby nodes. The purpose is to transfer tasks.

  Furthermore, in this embodiment, load transfer of other nodes is taken into consideration, and transfer of a plurality of tasks is allowed. Therefore, load balance can be achieved at high speed without being affected by the crowd effect.

  Next, a fifth embodiment of the present invention will be described with reference to the drawings. This example corresponds to the fifth embodiment. This embodiment is based on “Premise 1”.

  It is assumed that Q tasks can be executed from the head of the queue. At this time, the local queuing means 35 (see FIG. 3) of the node holds the total of response times and the number of measurements when the number of executions is “1” to “Q” for each number of executions. The effective sampling number is “10” and the additional execution number is “1”.

  In step C3 of FIG. 9 in the fifth embodiment of the present invention, it is assumed that an effective average response time is obtained, and a set of execution numbers from which the average response time is obtained is V ′.

  Then, the largest element of V ′ is a, and V is the union of V ′ and {a + 1}. By dividing the number of executions by the average response time according to the Little formula (the number of staying tasks is equal to the product of the task arrival rate and the average response time), the throughput for the number of executions belonging to V ′ is obtained.

  If k is the number of executions belonging to V and the throughput when the number of executions is k is T (k), the minimum k is found such that T (k) and T (j) are approximately equal, and this is the desired execution. Let it be the number R.

However, j is the smallest number of executions larger than k belonging to V. Also,
T (a + 1) = T (a) (28)
If V is empty, the value of R does not change.

  FIG. 24 is a response time table when Q = 4. Therefore, the sum of response times and the number of measurements when the number of executions is “4” or less are effective. At this time, the sum of response times equal to or greater than the effective sampling number is recorded when the number of simultaneous executions is “1” and “2” (a value greater than “5” is not considered).

  Therefore, when the total response time is divided by the number of measurements to obtain the response time, the average response time when the number of executions is “1” is “0.94”, and the average response time when the number of executions is “2” is “1.63”.

Therefore,
T (1) <T (2) = T (3)
So,
Let R = 2.

  FIG. 25 is a response time table when Q = 3. At this time, the sum of the response times exceeding the effective sampling number is recorded when the number of simultaneous executions is “1”, “2”, and “3”.

T (1) <T (2) <T (3) = T (4)
So,
Let R = 3.

  Usually, due to input / output (I / O) waiting or the like, there is a time during which a task cannot be executed even if resources are allocated to the task. Therefore, if a plurality of tasks are executed at the same time, even if one task cannot be executed, another task can be executed, thereby improving the throughput.

  On the other hand, when transferring a task once started, it is necessary to either execute the task again from the beginning or transfer the task status and execute it from the middle. Since overhead is large compared to transferring the previous task, in the present invention, only the task before the start of execution is targeted for transfer.

  Therefore, if the number of tasks to be executed simultaneously is increased, the throughput of each node is improved, but task transfer for balancing the load is limited accordingly.

  In this embodiment, by dynamically determining the number of tasks to be simultaneously executed, the throughput of the entire system is improved by making a trade-off between improvement of throughput for each node and reduction of throughput due to load imbalance.

  Next, a sixth embodiment of the present invention will be described in detail with reference to the drawings. This example corresponds to the sixth embodiment.

  In this embodiment, when the node group construction unit 12 (see FIG. 1) changes the node group, what is the node included in the same group in each node included in the changed node group (node group data) Send.

For example, in a cluster composed of 16 nodes (nodes 0 to 15, in this embodiment, the node ID is an absolute node ID),
Nodes 0-4 are the first node group,
Nodes 5-7 are the second node group,
Nodes 8-15 are the third node group,
When configuring
The global load balancer 1 is
When a decision is made to move node 5 to the first node group,
The global load balancer 1
An integer from 0 to 5 is transmitted to nodes 0 to 5 as node group data,
An integer from 6 to 7 is transmitted to the nodes 6 to 7 as node group data.

  The load information sharing means 32 (see FIG. 3) of the node that has received the node group data changes the information sharing method based on this information.

  Assuming a situation in which each node transmits load information to all other nodes in the node group, node 0 belonging to the first node group transmits information to nodes 1 to 4 before the node group is changed. However, after the change, information is transmitted to the nodes 1 to 5.

  Since a node without information is not selected as a load distribution target node, load transfer between groups is not performed.

  Since the load information sharing means 32 of the node that has received the node group data is provided with an information sharing method in a form that does not depend on the number of nodes that are information sharing targets, as shown in the prior application 1, specifically, The node to which information is transmitted or from which node the information is received can be reconstructed from the node group data.

  FIG. 26 is a diagram illustrating an example of an operation when a node group is changed.

  FIG. 26 shows the operation when three nodes (nodes 3, 4, 5) are added to a node group consisting of three nodes, nodes 0, 1, and 2.

  Immediately after the node is added, the load imbalance is large, but the load distribution means 31 (see FIGS. 3 and 4) in the present invention instantaneously shifts to the load balance state.

  As a result, even if the load due to a certain task type suddenly increases, not only the execution of other task types is not hindered, but also the load of each changing task type can be followed quickly. It is possible to satisfy the response time specified in.

  Although the present invention has been described with reference to the above-described embodiments and examples, the present invention is not limited to the above-described examples, and is within the scope of the invention of each claim. It goes without saying that various modifications and corrections that can be made by those skilled in the art are included.

  According to the present invention, a cluster system composed of a plurality of computing nodes is divided into a plurality of groups according to the use, and the number of nodes constituting the group is appropriately maintained according to the load situation at the time of execution, and the entire system It can be applied to applications that improve the performance.

  In addition, even if an overload condition occurs unexpectedly, it can be applied to an application that improves the load imbalance caused by changing the group configuration and improves the performance of the entire system.

  In particular, the present invention can also be applied to load balancing for web servers and application servers, and for keeping the number of servers used as web servers and application servers at an optimum number during execution.

It is a figure which shows the structure of the 1st Embodiment of this invention. It is a flowchart explaining the process sequence of the load distribution apparatus of the 1st Embodiment of this invention. It is a figure which shows the structure of the 2nd Embodiment of this invention. It is a figure which shows the structure of the load distribution means of the 2nd Embodiment of this invention. It is a flowchart which shows the operation | movement of the 2nd Embodiment of this invention. It is a flowchart which shows operation | movement at the time of determination of the transfer destination node of a task in the operation | movement of the 2nd Embodiment of this invention. It is a flowchart which shows the operation | movement at the time of determination of the transfer destination node of a task in the operation | movement of the 3rd Embodiment of this invention. It is a figure which shows the response time table of the 5th Embodiment of this invention. It is a flowchart which shows the operation | movement at the time of the response time table update in operation | movement of the 5th Embodiment of this invention. It is a block diagram which shows the structure of the 6th Embodiment of this invention. (A), (B) is the figure which displayed the graph of an example of the task-node characteristic used by embodiment of this invention. In embodiment of this invention, it is a figure which shows an example of the system operation conditions and task-node characteristic which are input into the system configuration input means 13. 5 is a flowchart illustrating an example of a procedure for calculating the number of nodes allocated to a node group in the embodiment of the present invention. In an embodiment of the invention, it is a figure showing an example of composition of a table which manages correspondence with a task kind and a node group. 14 is a flowchart showing an example of a procedure for extending step 123 in FIG. 13 in the embodiment of the present invention. 14 is a flowchart illustrating an example of a procedure for extending step 124 in FIG. 13 in the embodiment of the present invention. In an embodiment of the invention, it is a figure showing an example of composition of a table which manages a task arrival rate. In an embodiment of the invention, it is a figure showing an example of composition of a table which manages node performance. In an embodiment of the invention, it is a figure showing an example of the composition of the table which manages the composition of a node group. It is a figure for demonstrating the task transfer in the 2nd Example of this invention. It is a figure for demonstrating the task transfer in the 3rd Example of this invention. It is a figure for demonstrating information sharing to 4th Example of this invention. It is a figure for demonstrating the task transfer in the 4th Example of this invention. It is a figure which shows the response time table in the 5th Example of this invention. It is a figure which shows a response time table in the 5th Example of this invention. It is a figure for demonstrating the task transfer in the 6th Example of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Global load distribution apparatus 11 Task receiving apparatus 12 Node group construction means 13 System configuration input means 14 Execution performance measurement means 15 Task transmission means 16 Global queuing means 191 Task arrival rate management table 192 Node performance management table 193 Node group configuration table 194 Task type-node group correspondence table 2 Switching device 3-i Node 31 Load distribution means 32 Load information sharing means 33 Task transmission / reception means 34 Task execution means 35 Local queuing means 36 Load information management means 37 Node priority assignment means 38 Balance degree Calculation means 39 Transfer task determination means 310 Response time table

Claims (36)

In the nodes that make up the computer system,
Task transmission / reception means for transmitting / receiving tasks or task execution results; and
Queuing means for queuing tasks received by the task transmitting / receiving means (referred to as "local queuing means");
Load information sharing means for sharing load information;
Load distribution means for determining which task is to be transferred to which node based on load information of other nodes provided by the load information sharing means, and requesting task transfer to the task transmission / reception means;
A task execution unit that processes a task requested to be executed and returns an execution result of the task to the local queuing unit;
Only including,
The load balancing means includes
A balance calculation means for dynamically calculating at least the amount of tasks to be executed in the own node;
A node priority assignment means for assigning a priority to a node;
A node characterized by containing. The load balancing means includes
Load information management means for managing load information;
A transport task determination means for selecting nodes to be transported of loads in order of priority and determining the number of tasks to be transported to the selected nodes;
Node according to claim 1 including, characterized in that the. The node priority allocation means the allocation of priorities for said nodes, performed according to a cyclic monotonically increasing priority assignments, node according to claim 1 or 2, this and characteristics. The load information sharing means shares load information with different accuracy,
The load information management means of the load distribution means performs simple space accuracy complementation,
The node according to claim 2 , wherein the node priority assignment unit of the load distribution unit assigns a higher priority to a node having high load information accuracy. There is a response time table for calculating throughput for each number of task executions,
Said local queuing means of the load distribution unit, said with reference to the response table, the number of tasks to be executed at the same time, dynamically changing, according to any one of claims 1 to 4, characterized in that node. A switching device;
A plurality of nodes according to any one of claims 1 to 5 ,
The plurality of nodes are connected to the switching device and communicate via the switching device. A load balancer for a computer system including a plurality of nodes,
System configuration input means for inputting system configuration information, system operating conditions, and task-node characteristic information measured in advance;
A task receiving means for determining the type of task arriving at the system and measuring the task arrival rate for each task;
An execution performance measuring means for measuring execution performance information of a node at the time of execution;
At least one of the configuration information of the system, the operating conditions of the system, the task-node characteristics measured in advance, the task arrival rate measured at the time of execution, and the runtime performance information of the node A node group constructing means for configuring a node group and determining a task type to be executed on the node group;
A task transfer means for selecting a node to execute a task from the node group and transferring the task to the node;
A load balancer including:
A plurality of nodes according to any one of claims 1 to 5 ,
The plurality of nodes constitute a cluster system,
The load balancer is connected to the node and notifies the node group constituting the cluster system of configuration change information of the node group;
A load distribution system characterized by sharing load information between nodes belonging to the same node group. The load balancer further includes queuing means (referred to as “global queuing means”) for queuing tasks so that the number of staying tasks in a node to which tasks are transferred does not exceed a predetermined number. The load distribution system according to claim 7 . The node group construction means associates a task type that satisfies the operating conditions of the system with one node group, and associates a task type that does not satisfy the operating conditions of the system with another node group; The load distribution system according to claim 7 or 8 , wherein The node group construction means configures a node group so that one kind of task processing is performed on one node group, and stores and manages the node group and a task type corresponding to the node group in association with each other. The load distribution system according to claim 7 or 8 , wherein The load distribution system according to claim 7 , wherein the task-node characteristic information includes the task arrival rate and throughput and / or the task arrival rate and response characteristic information. 12. The load distribution system according to claim 7 , wherein tasks are executed in different node groups for each task type, and tasks are not transferred between the node groups. In the nodes that make up the computer system,
A task transmission / reception step for transmitting / receiving a task or a task execution result; and
A local queuing step for queuing a task received by the task transmitting / receiving means;
A load distribution step of determining which task is to be transferred to which node based on load information of another node provided by load information sharing means for sharing load information, and requesting the task transmitting / receiving means to transfer the task;
A task execution step for processing a task requested to be executed and returning an execution result of the task to the local queuing step;
Including
The load balancing step includes
A balance calculation step for dynamically calculating at least the amount of tasks to be executed on the own node;
A node priority assignment step for assigning priorities to the nodes;
A load balancing method comprising: The load balancing step includes
Load information management step for managing load information;
A transport task determination step for selecting nodes to be transported of loads in order of priority and determining the number of tasks to be transported to the selected nodes;
The load distribution method according to claim 13, further comprising: 15. The load distribution method according to claim 13, wherein the node priority assignment step assigns a priority to the node according to a cyclic monotone increasing priority assignment. The load information sharing means shares load information with different accuracy,
The load information management step of the load distribution step performs simple space accuracy complementation,
The load distribution method according to claim 14 , wherein the node priority assignment step of the load distribution step assigns a higher priority to a node having high load information accuracy. The local queuing step of the load distribution step refers to a response table for calculating throughput for each execution number of tasks, and dynamically changes the number of tasks executed simultaneously. Item 16. The load distribution method according to any one of Items 13 to 15 . A plurality of nodes according to any one of claims 13 to 17,
The plurality of nodes are connected to a switching device and communicate via the switching device. A load balancing method for a computer system including a plurality of nodes,
In the load balancer, a system configuration input step for inputting system configuration information, system operating conditions, and task-node characteristic information measured in advance,
A task receiving step for determining the type of task arriving at the system and measuring the task arrival rate for each task;
An execution performance measurement step for measuring execution performance information of a node at the time of execution;
At least one of the configuration information of the system, the operating conditions of the system, the task-node characteristics measured in advance, the task arrival rate measured at the time of execution, and the runtime performance information of the node A node group construction step for configuring a node group and determining a task type to be executed on the node group;
A task transfer step of selecting a node to execute the task from the node group and transferring the task to the node;
Including
A plurality of nodes according to any one of claims 13 to 17,
The plurality of nodes constitute a cluster system,
The load balancer is connected to the node, and notifies the node constituting the cluster system of configuration change information of the node group,
A load distribution method characterized by sharing load information between nodes belonging to the same node group. The load balancing apparatus further comprises a global queuing step for queuing a task so that the number of staying tasks in a task transfer destination node does not exceed a predetermined number. The load balancing method described . The node group construction step corresponds to one node group with a task type that satisfies the operating conditions of the system, and to another node group with a task type that does not satisfy the operating conditions of the system; The load balancing method according to claim 19 or 20, wherein The node group construction step configures a node group so that one type of task processing is performed on one node group, and stores and manages the node group and a task type corresponding to the node group in association with each other. The load distribution method according to claim 19 or 20, wherein: The load distribution according to any one of claims 19 to 22, wherein the task-node characteristic information includes the task arrival rate and throughput, and / or the task arrival rate and response characteristic information. Way . The load distribution method according to any one of claims 19 to 23, wherein tasks are executed in different node groups for each task type, and tasks are not transferred between node groups. To the computers that make up the nodes that make up the computer system,
Task transmission / reception processing for transmitting / receiving tasks or task execution results;
A queuing process for connecting a task received by the task transmitting / receiving means to a queue;
Load distribution processing for determining which task is transferred to which node based on load information of other nodes provided by a load information sharing unit that shares load information, and requesting task transfer to the task transmission / reception processing;
A task execution process for processing a task requested to be executed and returning an execution result of the task to the local queuing means;
Was executed,
As the load balancing process,
Balance degree calculation processing that dynamically calculates the amount of tasks to be executed at least on its own node;
Node priority assignment processing for assigning priorities to nodes;
A program for running The program according to claim 25 ,
Load information management processing for managing load information;
A transport task determination process for selecting a load transport target node in order of priority and determining the number of tasks to be transported to the selected node;
A program for causing the computer to execute the load distribution process including: The program according to claim 26 ,
A program for causing the computer to execute the node priority assignment processing for assigning priority to another node according to cyclic monotonically increasing priority assignment. The program according to claim 25 ,
The load information sharing unit shares load information with different accuracy,
A program for causing the computer to execute processing for reconstructing the load information from the value and accuracy of the load information sent from another node. The program according to claim 25 ,
A program for causing the computer to execute a process of dynamically changing the number of tasks to be executed simultaneously with reference to a response time table for calculating throughput for each number of tasks executed. A plurality of nodes according to any one of claims 25 to 29,
A program for causing the computer of the node to execute a process connected to a switching device and communicating via the switching device. In a computer constituting a load balancer for a computer system including a plurality of nodes,
System configuration input processing for inputting system configuration information, system operating conditions, and task-node characteristic information measured in advance;
Task reception processing that determines the type of task that arrives at the system and measures the task arrival rate for each task;
An execution performance measurement process that measures the execution performance information of a node at the time of execution;
At least one of the configuration information of the system, the operating conditions of the system, the task-node characteristics measured in advance, the task arrival rate measured at the time of execution, and the runtime performance information of the node A node group construction process for configuring a node group and determining a task type to be executed on the node group;
A task transfer process for selecting a node to execute a task from the node group and transferring the task to the node;
And execute
A plurality of nodes according to any one of claims 25 to 29,
The plurality of nodes constitute a cluster system,
The load balancer is connected to the node and notifies the node group constituting the cluster system of configuration change information of the node group;
A program for causing the computer of the node to execute a process of sharing load information between nodes belonging to the same node group. 32. The program according to claim 31, further causing the load balancer to further execute a global queuing process for queuing tasks so that the number of staying tasks in a task transfer destination node does not exceed a predetermined number . The node group construction process associates a task type that satisfies the operating condition of the system with one node group, and associates a task type that does not satisfy the operating condition of the system with another node group. The program according to claim 31 or 32, wherein The node group construction process configures a node group so that one type of task processing is performed on one node group, and stores and manages the node group and a task type corresponding to the node group in association with each other. The program according to claim 31 or 32. 32. The program according to claim 31, wherein the task-node characteristic information includes the task arrival rate and throughput and / or the task arrival rate and response characteristic information . The program according to any one of claims 31 to 34, wherein tasks are executed in different node groups for each task type, and tasks are not transferred between node groups.

Images