JP2024069960A - RESOURCE RECONFIGURATION PROGRAM, RESOURCE RECONFIGURATION METHOD, AND INFORMATION PRO - Google Patents

Abstract

To make it possible to appropriately determine whether or not to perform resource reconfiguration.SOLUTION: A processing part 12 calculates a first time required from reconfiguration to job completion in the case of allocating a job to a first node on the basis of a reconfiguration time required for reconfiguration of the first node requiring reconfiguration of resources for executing the job, a job execution time after the reconfiguration, and a first coefficient representing influence on the execution time of communication contention due to communication when the job is executed by the first node. The processing part 12 calculates a second time required till the job completion in the case of allocating the job to a second node on the basis of a second coefficient representing influence on the execution time and the execution time of communication contention due to communication when the job is executed by the second node which does not require the resource configuration. The processing part 12 performs the resource reconfiguration in the first node in the case that the first time is shorter than the second time, and does not perform the resource reconfiguration in the first node in the case that the first time is equal to or longer than the second time.SELECTED DRAWING: Figure 1

Description

The present invention relates to a resource reconfiguration program, a resource reconfiguration method, and an information processing system.

Currently, systems are being used that pool hardware resources and change the node resource configuration by switching the connection between the node and the resource via a switch. Such systems are called disaggregated systems. Examples of resources that can be pooled include GPUs (Graphics Processing Units), FPGAs (Field Programmable Gate Arrays), and SSDs (Solid State Drives).

In addition, a management device has been proposed that takes into account cases where a service is provided not only on one VM (Virtual Machine) but also on multiple VMs, thereby reducing the hardware resources used for the service while still satisfying the SLA (Service Level Agreement).

JP 2018-116556 A

When assigning a job to a node, if the resources required by the job are insufficient on the assigned node, the resources must be reconfigured. Reconfiguring resources takes time and can delay job completion.

On the other hand, if there is a node that already has the resources required by the job configured, you can avoid reconfiguring the resources by assigning the job to that node. However, communications when the job is executed by that node may conflict with communications between other nodes. Such communications conflicts can also cause delays in job completion.

For this reason, even if a job is simply assigned to a node that does not require resource reconfiguration, there is a problem in that the delay in job completion is not necessarily improved compared to the case where reconfiguration is performed.
In one aspect, the present invention aims to make it possible to appropriately determine whether or not to reconfigure resources.

In one aspect, a resource reconfiguration program is provided. This resource reconfiguration program causes a computer to execute the following process. The computer calculates a first time required from reconfiguration to completion of a job when the job is assigned to the first node, based on a reconfiguration time required for reconfiguring a first node that is a candidate for job allocation and requires resource reconfiguration to execute the job, the execution time of the job after reconfiguration, and a first coefficient representing the effect on the execution time of a communication conflict accompanying communication during execution of the job by the first node. The computer calculates a second time required to complete a job when the job is assigned to the second node without reconfiguring resources, based on a second coefficient representing the effect on the execution time of a communication conflict accompanying communication during execution of the job by a second node that is a candidate for job allocation and does not require resource reconfiguration, and the execution time. The computer compares the first time with the second time, and if the first time is shorter than the second time, reconfigures the resources in the first node, and if the first time is equal to or longer than the second time, does not reconfigure the resources in the first node.

In one aspect, a resource reconfiguration method executed by a computer is provided. In another aspect, an information processing system having a storage unit and a processing unit is provided.

In one aspect, it is possible to appropriately decide whether or not to reconfigure resources.

FIG. 1 is a diagram illustrating an information processing system according to a first embodiment. FIG. 1 illustrates an example of an information processing system according to a second embodiment. FIG. 2 illustrates an example of hardware of a management apparatus. FIG. 2 is a diagram illustrating an example of connections between nodes. FIG. 1 illustrates an example of a connection between a node in a rack and a resource pool. FIG. 13 illustrates an example of free nodes that are candidates for job allocation. FIG. 13 illustrates an example in which communication contention occurs. FIG. 13 is a diagram illustrating an example of reconfiguration of node resources. FIG. 13 is a diagram illustrating an example in which communication contention is avoided. FIG. 2 illustrates an example of functions of a management device. FIG. 4 illustrates an example of a job management table. FIG. 13 is a diagram illustrating an example of a reference communication time table. FIG. 13 illustrates an example of a communication performance table. FIG. 13 is a diagram illustrating an example of an evaluation value table. 13 is a flowchart showing an example of processing by a job scheduler. 13 is a flowchart illustrating an example of a node selection process. 13 is a flowchart illustrating an example of a process for extracting a node combination. 13 is a flowchart illustrating an example of a process for evaluating a node combination. 11A and 11B are diagrams illustrating an example of a difference in total time related to the execution of a job.

The present embodiment will be described below with reference to the drawings.
[First embodiment]
A first embodiment will be described.

FIG. 1 is a diagram illustrating an information processing system according to a first embodiment.
The information processing system 1 has an information processing device 10 and nodes 20, 20a, 20b, 20c, 20d, .... The information processing device 10 and the nodes 20, 20a, ... are connected to a management network 30. The nodes 20, 20a, ... are connected to an inter-node network 40. The management network 30 is a network used for communication between the information processing device 10 and the nodes 20, 20a, .... The inter-node network 40 is a network used for communication between the nodes.

The information processing device 10 has a storage unit 11 and a processing unit 12. The storage unit 11 may be a volatile semiconductor memory such as a random access memory (RAM), or may be a non-volatile storage such as a hard disk drive (HDD) or a flash memory. The processing unit 12 is, for example, a processor such as a central processing unit (CPU), a GPU, or a digital signal processor (DSP). However, the processing unit 12 may include an electronic circuit for a specific purpose such as an application specific integrated circuit (ASIC) or an FPGA. The processor executes a program stored in a memory such as a RAM (which may be the storage unit 11). A collection of multiple processors is sometimes called a "multiprocessor" or simply a "processor."

The nodes 20, 20a, ... also have the same hardware as the information processing device 10. In addition, the functions of the information processing device 10 may be provided by any of the nodes 20, 20a, .... In other words, any of the nodes 20, 20a, ... may operate as the information processing device 10.

The information processing system 1 is a disaggregation system. That is, the information processing system 1 is capable of changing the resource configuration of each of the nodes 20, 20a, .... The information processing device 10 manages the resource configuration of each of the nodes 20, 20a, ... and controls the reconfiguration of the resources of each of the nodes 20, 20a, .... The resources include, for example, a GPU, an FPGA, and an SSD.

The processing unit 12 accepts a job execution request. The job execution request includes the number of nodes required for execution, information on the resources required for job execution, and information on the job execution time. The resource information includes the type and amount of resources required per node, such as the number of GPUs and SSD capacity. The job execution time is the time to execute the job. The processing unit 12 assigns the job to the requested number of nodes out of nodes 20, 20a, ..., and has the assigned nodes execute the job. The assigned nodes are the nodes that have the requested resources.

The resources connected to the nodes 20, 20a, ... are managed by pooling them in a resource pool provided for each group of the nodes 20, 20a, .... The resource pool is a device that aggregates multiple resources. The group is, for example, a group of racks in which the nodes are mounted. For example, each node belonging to a certain group and each resource of the resource pool corresponding to the group are connected via a connection device such as a PCIe (Peripheral Component Interconnect-Express) switch corresponding to the group. In this case, the connection device changes the connection relationship between the nodes and resources within the group, thereby reconfiguring the resources in the nodes. Note that the resource pool and connection device are not shown in FIG. 1.

When a job is assigned to a node that does not have the requested resources, the resources of that node must be reconfigured. In reconfiguring resources, for example, configuration changes are made, such as increasing the number of GPUs available to the node. Reconfiguring resources takes, for example, several seconds to several minutes. Specifically, if a node supports hot plug/remove, resource reconfiguration takes about several seconds. On the other hand, if a node does not support hot plug/remove and the node is rebooted, resource reconfiguration takes several minutes.

Information indicating the time required for reconfiguring resources is stored in advance in the storage unit 11 depending on whether the nodes of the information processing system 1 support hot plugging/removal. In addition, information on the job execution time included in the job execution request is also stored in the storage unit 11.

When there are a first node and a second node among the nodes 20, 20a, ... that are candidates for the assignment of a job, the processing unit 12 selects the node to which the job is assigned from the first node and the second node as follows: The first node is a node that requires resource reconfiguration to execute the job. The second node is a node that has already configured the resources required by the job and does not require resource reconfiguration to execute the job.

The processing unit 12 calculates a first time required from reconfiguration to completion of the job when the job is assigned to the first node. The first time is the total time required for reconfiguring the resources of the first node and for starting and completing job execution on the first node after the reconfiguration. The processing unit 12 calculates the first time based on the reconfiguration time, the execution time of the job after the reconfiguration, and a first coefficient that represents the effect on the execution time of communication contention associated with communication during job execution by the first node.

The first time is T _Total1 , the reconfiguration time is T _reconf , the job execution time is T _job , and the first coefficient is a1, where a1 is a real number equal to or greater than 1. The first time T _Total1 is calculated, for example, by equation (1).

T _Total1 = T _reconf + a1 × T _job ... (1)
The processing unit 12 also calculates a second time required to complete the job when the job is assigned to the second node. The second time is the total time required from the start of job execution on the second node to the completion of execution. The second time differs from the first time in that it does not include the time required for reconfiguring resources. The processing unit 12 calculates the second time based on a second coefficient representing the effect on execution time of communication contention associated with communication during job execution by the second node and the execution time of the job.

The second time period is T _Total2 and the second coefficient is a2, where a2 is a real number equal to or greater than 1. The second time period T _Total2 is calculated, for example, by the formula (2).
T _Total2 = a2 × T _job ... (2)
The processing unit 12 compares the first time with the second time. If the first time is shorter than the second time (if T _Total1 <T _Total2 ), the processing unit 12 reconfigures resources in the first node. If the first time is equal to or longer than the second time (if T _Total1 ≧T _Total2 ), the processing unit 12 does not reconfigure resources in the first node.

The processing unit 12 may measure the first communication time of communication between nodes belonging to a first group of nodes including the first node, which are candidates for job allocation, using a predetermined communication benchmark program, and calculate the first coefficient a1 based on the measurement result. More specifically, the processing unit 12 may obtain in advance a reference communication time between nodes in a case where there is no communication contention, using the same number of nodes as the number of nodes belonging to the first group, and calculate the first coefficient a1 based on the ratio between the first communication time and the reference communication time. Similarly, the processing unit 12 may calculate the second coefficient a2 based on the measurement result of the second communication time of communication between nodes belonging to a second group of nodes including the second node, which are candidates for job allocation.

The processing unit 12 may also calculate the first coefficient a1 and the second coefficient a2 based on the amount of resources required per node for job execution. It is estimated that the greater the amount of resources required per node, the greater the amount of calculations at the node and the greater the amount of communication between nodes. For this reason, the processing unit 12 may increase the first coefficient a1 and the second coefficient a2 as the amount of resources increases.

When the processing unit 12 reconfigures resources in the first node, it instructs the first node to reconfigure the resources, causes the first node to execute the reconfiguration, and assigns the requested job to the first node. When the processing unit 12 does not reconfigure resources in the first node, it assigns the requested job to the second node, not the first node. Then, the processing unit 12 causes the assigned node to execute the job.

Figure 1 shows examples of candidate nodes for allocation. Assume that the number of nodes required for a job is two. For example, the first candidate #1 is a combination of nodes 20a and 20b. The second candidate #2 is a combination of nodes 20a and 20c. For simplicity of explanation, it is assumed that node 20a does not require resource reconfiguration to execute the job and has been confirmed as the allocation destination for the job.

Node 20b requires resource reconfiguration to execute the job. In other words, node 20b corresponds to the first node. Node 20c does not require resource reconfiguration to execute the job. In other words, node 20c corresponds to the second node.

When node 20b is selected as the job allocation destination, for example, it is assumed that no communication conflict occurs due to conflict with existing communication paths between other nodes in the internode network 40. As an example, it is assumed that nodes 20 and 20d are executing other jobs and communicating using a communication path 41 that is a part of the internode network 40. A communication path 42 that is a part of the internode network 40 used for communication with nodes 20a and 20b is separate from communication path 41. Therefore, there is no conflict between communication between nodes 20 and 20d and communication between nodes 20a and 20b. In this case, the first coefficient a1 is a1=1. The first coefficient a1=1 corresponds to a case where there is no effect of communication conflict on the job execution time.

On the other hand, if node 20c is selected as the destination for the job, a communication conflict occurs due to a conflict with the communication path of existing communication between other nodes in internode network 40. As an example, nodes 20a and 20c communicate using communication path 41 of internode network 40, and a communication conflict occurs on communication path 41 of the existing communication between nodes 20 and 20d. In this case, the second coefficient a2 is a2>1.

When a1=1 or a1 is close to 1, the first time T _Total1 is easily influenced by the reconfiguration time T _reconf . On the other hand, the second time T _Total2 is easily influenced by the second coefficient a2. Therefore, when the reconfiguration time T _reconf is relatively short and the second coefficient a2 is relatively large, T _Total1 <T _Total2 . In this case, reconfiguring the node 20b and assigning the job to the node 20b can reduce the delay until the job is completed. On the other hand, when the reconfiguration time T _reconf is relatively long and the second coefficient a2 is relatively small, T _Total2 ≦T _Total1 . In this case, assigning the job to the node 20c without reconfiguring the node 20b can reduce the delay until the job is completed.

In the above example, a1=1 is exemplified. However, even if a1>1, the processing unit 12 can evaluate the total required time for job execution using T _reconf , a1, and a2, and determine whether or not to reconfigure resources based on the evaluation result.

As described above, according to the information processing device 10, the first time T Total1 is calculated based on the reconfiguration time T _reconf required for reconfiguring the resources of the first node, the job execution time T _job , and the first coefficient a1. The first coefficient a1 represents the influence of the communication contention caused by the communication during the execution of the job by the first node on the execution time T _job . In addition, the second time T _Total2 is calculated based on the second coefficient a2 and the job execution time T _job . The second coefficient a2 represents the influence of the communication contention caused by the communication during the execution of _the job by the second node on the execution time T _job . Then, the first time T _Total1 and the second time T _Total2 are compared. If the first time T _Total1 is shorter than the second time T _Total2 , the resources in the first node are reconfigured. If the first time T _Total1 is greater than or equal to the second time T _Total2 , the reconfiguration of resources in the first node is not performed.

This allows the information processing device 10 to appropriately decide whether or not to reconfigure resources, taking into account the impact of communication contention. In addition, by taking into account the time required to reconfigure resources and the impact of communication contention, the information processing device 10 becomes able to select a node to which a job is assigned so as to shorten the time required to complete job execution.

[Second embodiment]
Next, a second embodiment will be described.
FIG. 2 illustrates an example of an information processing system according to the second embodiment.

The information processing system 2 has a management device 100 and nodes 200, 200a, 200b, .... The information processing system 2 is a disaggregation system that can change the configuration of the hardware resources of each of the nodes 200, 200a, 200b, .... The management device 100 and the nodes 200, 200a, 200b, ... are connected to a management network 50. The nodes 200, 200a, 200b, ... are connected to an inter-node network 60. The management network 50 is, for example, an Ethernet (registered trademark) network. The inter-node network 60 is, for example, an InfiniBand network. However, the inter-node network 60 may be another type of network, such as Ethernet.

The management device 100 is a server computer that receives a job execution request and controls the allocation of jobs to the nodes 200, 200a, 200b, ... based on the execution request. The job execution request includes information on the number of nodes required to execute the job, the amount of hardware resources required for each node, and the execution time of the job. The management device 100 controls the reconfiguration of resources of the nodes 200, 200a, 200b, ... as required for job allocation. The management device 100 is an example of the information processing device 10 of the first embodiment.

Nodes 200, 200a, 200b, ... are server computers that execute assigned jobs. Nodes 200, 200a, 200b, ... can communicate with each other via inter-node network 60. For example, when a job is executed using two nodes, the two nodes execute the job while communicating with each other via inter-node network 60.

FIG. 3 illustrates an example of hardware of the management apparatus.
The management device 100 has a CPU 101, a RAM 102, a HDD 103, a GPU 104, an input interface 105, a media reader 106, and a communication interface 107. These units of the management device 100 are connected to a bus inside the management device 100. The CPU 101 corresponds to the processing unit 12 in the first embodiment. The RAM 102 or the HDD 103 corresponds to the storage unit 11 in the first embodiment.

The CPU 101 is a processor that executes program instructions. The CPU 101 loads at least a portion of the programs and data stored in the HDD 103 into the RAM 102 and executes the programs. The CPU 101 may include multiple processor cores. The management device 100 may also have multiple processors. The processes described below may be executed in parallel using multiple processors or processor cores. A collection of multiple processors may also be called a "multiprocessor" or simply a "processor."

RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by CPU 101 and data used by CPU 101 for calculations. Note that management device 100 may include a type of memory other than RAM, and may include multiple memories.

The HDD 103 is a non-volatile storage device that stores software programs such as the OS (Operating System), middleware, and application software, as well as data. Note that the management device 100 may also include other types of storage devices, such as flash memory and SSDs, and may also include multiple non-volatile storage devices.

The GPU 104 outputs an image to a display 111 connected to the management device 100 in accordance with an instruction from the CPU 101. The display 111 may be any type of display, such as a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD), a plasma display, or an organic electro-luminescence (OEL) display.

The input interface 105 acquires an input signal from an input device 112 connected to the management device 100 and outputs the signal to the CPU 101. The input device 112 may be a pointing device such as a mouse, a touch panel, a touch pad, or a trackball, a keyboard, a remote controller, or a button switch. In addition, multiple types of input devices may be connected to the management device 100.

The media reader 106 is a reading device that reads programs and data recorded on the recording medium 113. For example, a magnetic disk, an optical disk, a magneto-optical disk (MO: Magneto-Optical disk), a semiconductor memory, etc. can be used as the recording medium 113. Magnetic disks include flexible disks (FD: Flexible Disks) and HDDs. Optical disks include compact discs (CDs) and digital versatile discs (DVDs).

For example, the media reader 106 copies the program or data read from the recording medium 113 to another recording medium such as the RAM 102 or the HDD 103. The read program is executed by the CPU 101, for example. Note that the recording medium 113 may be a portable recording medium, and may be used to distribute programs and data. The recording medium 113 and the HDD 103 may also be referred to as computer-readable recording media.

The communication interface 107 is connected to the management network 50 and communicates with other information processing devices including the nodes 200, 200a, 200b, ... via the management network 50. The communication interface 107 may be a wired communication interface connected to a wired communication device such as a switch or a router, or a wireless communication interface connected to a wireless communication device such as a base station or an access point.

The nodes 200, 200a, 200b, ... are also realized by the same hardware as the management device 100. The nodes 200, 200a, 200b, ... are connected not only to the management network 50 but also to the inter-node network 60, and therefore also have a communication interface that connects to the inter-node network 60.

FIG. 4 is a diagram showing an example of connections between nodes.
Each node is mounted on one of racks R1, R2, R3, and R4. For example, some of all the nodes, including nodes 200 and 200a, are mounted on rack R1. Some of all the nodes, including nodes 200m and 200n, are mounted on rack R2. Although not shown in the figure, some of all the nodes are also mounted on racks R3 and R4.

Each of the racks R1, R2, R3, and R4 also has a resource pool. A resource pool is a device that aggregates hardware resources. In the second embodiment, a GPU is exemplified as the resource. However, the resources pooled in the resource pool may include other types of hardware such as FPGAs and SSDs. For example, the rack R1 has a resource pool 300. The rack R2 has a resource pool 300a. Although not shown in the figure, the racks R3 and R4 are also equipped with resource pools.

Nodes in the same rack are connected by an inter-node connection switch mounted on the rack. Also, nodes in the same rack and a resource pool are connected by a PCIe switch mounted on the rack. For example, rack R1 has an inter-node connection switch 61 and a PCIe switch 71. The inter-node connection switch 61 connects the nodes 200, 200a, ... mounted on rack R1. The PCIe switch 71 connects the nodes 200, 200a, ... mounted on rack R1 and the resource pool 300. Also, rack R2 has an inter-node connection switch 62 and a PCIe switch 72. The inter-node connection switch 62 connects the nodes 200m, 200n, ... mounted on rack R2. Also, the PCIe switch 72 connects the nodes 200m, 200n, ... mounted on rack R2 and the resource pool 300a. Although not shown in the figure, racks R3 and R4 are also equipped with inter-node connection switches and PCIe switches.

The node connection switches of each rack, including the inter-node connection switches 61 and 62, are connected to the upper switches 65, 66, 67, and 68. The upper switches 65, 66, 67, and 68 enable communication between nodes across racks. The node connection switches of each rack, including the inter-node connection switches 61 and 62, and the upper switches 65, 66, 67, and 68 are all InfiniBand switches, and form the inter-node network 60.

The topology of the inter-node network 60 is, for example, FatTree. Specifically, the number of links connecting to the nodes in each inter-node connection switch is equal to the number of links on the upper switch side of the inter-node connection switch. The usage of each link is balanced by using different links on the upper switch side used to send packets from a node to a communication partner node in another rack depending on the communication partner node. However, the topology of the inter-node network 60 may be a topology other than FatTree. Here, a link is a communication path included in the inter-node network 60.

FIG. 5 is a diagram illustrating an example of a connection between nodes in a rack and a resource pool.
The node 200 includes a CPU 201, a RAM 202, a HDD 203, and a NIC (Network Interface Card) 204. The node 200a includes a CPU 201a, a RAM 202a, a HDD 203a, and a NIC 204a. The NICs 204 and 204a are communication interfaces that connect to the inter-node connection switch 61. The other nodes in the rack R1 also include hardware similar to that of the nodes 200 and 200a. The CPUs of each node in the rack R1, including the CPUs 201 and 201a, are connected to a PCIe switch 71. The PCIe switch 71 recognizes the CPUs of each node in the rack R1, including the CPUs 201 and 201a, as root complexes, and connects them to GPUs 301, 302, ..., which are endpoint devices included in the resource pool 300. The PCIe switch 71 changes the connection relationship between the CPU of each node in the rack R1 and the GPUs 301, 302, . . . of the resource pool 300 in response to an instruction from the node.

Here, when the management device 100 receives a job allocation request, it sets free nodes to which no job is currently assigned as candidates for job allocation.
FIG. 6 is a diagram showing an example of free nodes that are candidates for job allocation.

Each node is identified by a node number. For example, node 200 has a node number of "0", and node 200a has a node number of "1". Below, a node with node number n will be written as node "n". The number of nodes mounted on one rack is assumed to be four. In other words, four nodes are connected to one inter-node connection switch. The total number of racks is assumed to be four, and the total number of nodes is assumed to be 16.

In this case, in the FatTree topology of the inter-node network 60, there are four links for one inter-node connection switch, connecting it to upper switches 65, 66, 67, and 68, respectively. Here, the inter-node connection switch 63 is mounted on rack R3. The inter-node connection switch 64 is mounted on rack R4.

Nodes "0" to "3" are connected to inter-node connection switch 61. Nodes "4" to "7" are connected to inter-node connection switch 62. Nodes "8" to "11" are connected to inter-node connection switch 63. Nodes "12" to "15" are connected to inter-node connection switch 64.

In the figure, nodes "3," "4," "7," "8" to "15," indicated by black circles, are nodes to which jobs have already been assigned. Nodes to which jobs have already been assigned have been assigned a previous job and are currently executing that job, and are excluded from being candidates for allocation of new jobs. Nodes "0" to "2," "5," and "6," indicated by white circles, are nodes to which jobs have not yet been assigned. Nodes to which jobs have not yet been assigned are candidates for allocation of new jobs. Furthermore, in the figure, the rectangles marked with the letter "G" connected to the nodes indicate the GPUs connected to those nodes.

For example, when allocating nodes to a job that requires four nodes and one GPU per node, the management device 100 selects four node combinations from nodes "0", "1", "2", "5", and "6" as candidate combinations of nodes to which the job is to be allocated. The node combinations may also be referred to as node groups.

In this way, the management device 100 can select candidate nodes to which a job is assigned across racks. In this case, depending on the selection of the node to which the job is assigned in another rack, communication contention may occur in some links of the inter-node network 60 when the job is executed.

FIG. 7 illustrates an example in which communication contention occurs.
7, links between switches are shown by lines connecting inter-node connection switches 61, 62, 63, 64 and upper level switches 65, 66, 67, 68. Numbers written on the links, for example "4, 8, 12", are node numbers of communication partners and indicate the link selected when communicating with the node with the node number from the inter-node connection switch 61 side.

For example, when node "1" communicates with node "5" and node "3" communicates with node "9", the same link connecting the inter-node switch 61 and the upper switch 66 is used. If the bandwidth of the link is shared and used, performance degradation due to communication contention may occur. This problem can also occur in topologies other than FatTree.

Incidentally, the connection configuration of the GPUs to the nodes can be changed.
FIG. 8 is a diagram illustrating an example of reconfiguration of node resources.
8A shows an example of reconfiguration in which a GPU connected to node "5" is reconnected to node "6." Fig. 8B shows an example of reconfiguration in which a GPU that is not being used in the resource pool 300a is connected to node "6."

For example, in resource management in each resource pool, resources such as GPUs can be returned to the resource pool after use, or left as is until a configuration change is required. In the former case, GPUs from the resource pool can be connected to the node to which the job is assigned according to the requested number of GPUs. In the latter case, as shown in FIG. 8(A), a GPU connected to node "5" can be reconnected to node "6" as necessary. Alternatively, as shown in FIG. 8(B), if there is a free GPU in resource pool 300a, the existing connection of the GPU to node "5" can be maintained and the free GPU can be connected to node "6."

In this way, in the information processing system 2, it is possible to prepare a node equipped with a GPU required for a job by reconfiguring the GPU in the node. This reconfiguration can sometimes avoid the communication contention illustrated in FIG. 7.

FIG. 9 is a diagram illustrating an example in which communication contention is avoided.
8A or 8B, a GPU is connected to node "6", which allows the management device 100 to assign a job to node "6". By using node "6" for job execution instead of node "5", a link connecting node "1" and upper switch 67 is used for communication between node "1" and node "6", for example. The link connecting node "6" and upper switch 67 is different from the link connecting node "3" and upper switch 66, which is used when node "9" communicates, and communication contention is avoided.

Communication contention affects the delay until the job execution is completed. On the other hand, even if resources are reconfigured to avoid communication contention, the reconfiguration takes time and affects the delay until the job execution is completed. The time required for reconfiguration, i.e., the reconfiguration time, is determined in advance, for example, depending on whether or not the node supports hot plug/remove. Hot plug/remove is a function that makes it possible to change the connection of resources such as GPUs without rebooting the node. For example, if the node supports hot plug/remove, the reconfiguration of the GPU takes about a few seconds. On the other hand, if the node does not support hot plug/remove and the node is rebooted, the reconfiguration of the GPU takes several minutes.

Therefore, the management device 100 provides a function that allows jobs to be executed efficiently by determining whether or not to reconfigure resources while taking communication contention into account and allocating nodes to jobs.

FIG. 10 illustrates an example of functions of the management device.
The management device 100 includes a storage unit 120 and a job scheduler 130. The storage unit 120 uses storage areas of the RAM 102 and the HDD 103. The job scheduler 130 is realized by the CPU 101 executing a program stored in the RAM 102.

The memory unit 120 stores information used in the processing of the job scheduler 130. The information stored in the memory unit 120 includes a job management table 121, a reference communication time table 122, a communication performance table 123, and an evaluation value table 124.

The job management table 121 is a table that holds job information such as the number of nodes required for a job, the number of GPUs per node, and the execution time of the job.
The reference communication time table 122 is a table that holds a reference communication time used to determine the presence or absence of communication contention by communication benchmark measurement. The reference communication time is measured in advance before the operation of the system starts, and is registered in the reference communication time table 122. The communication benchmark measurement is performed by having each target node execute a predetermined communication benchmark program for a short period of time and measuring the communication time between the nodes.

The communication performance table 123 is a table that holds information on the measurement results of the communication performance for each candidate combination of nodes to which a job is assigned. The communication performance for each candidate combination of nodes is obtained by performing a communication benchmark measurement for that candidate combination.

The evaluation value table 124 is a table that holds evaluation values of the total time until a job is completed for each candidate combination of nodes to which the job is assigned. In addition to the information in the communication performance table 123, the evaluation of the total time also takes into account the reconfiguration time if the GPU is reconfigured on the node.

In addition to the above information, the memory unit 120 also stores information such as the rack on which each node is mounted, the job allocation status for each node, and the number of GPUs connected to each node.

The job scheduler 130 assigns nodes 200, 200a, 200b, ... to a job and has the assigned node execute the job. The job scheduler 130 has a job information acquisition unit 131, a node assignment unit 132, and a node selection unit 133.

The job information acquisition unit 131 accepts input of a job execution request. The job execution request is input to the management device 100 from, for example, a client device connected to the management network 50. The job information acquisition unit 131 acquires job information included in the execution request, such as the number of nodes required for the job, the number of GPUs per node, and the job execution time, and registers the information in the job management table 121.

When the job information acquisition unit 131 accepts job information, the node allocation unit 132 requests the node selection unit 133 to select a node to be assigned to the job based on the job information. The node allocation unit 132 acquires the selection result of the node to be assigned from the node selection unit 133, and assigns the job to the assigned node. In other words, the node allocation unit 132 instructs the assigned node to execute the job.

The node selection unit 133 selects nodes to be assigned to the job in response to a node selection request from the node assignment unit 132. The node selection unit 133 has a node combination extraction unit 133a and an evaluation unit 133b.

The node combination extraction unit 133a extracts a combination of nodes to be assigned to a job. Specifically, the node combination extraction unit 133a first extracts candidate combinations of nodes to be assigned to a job from among available nodes. The node combination extraction unit 133a requests the evaluation unit 133b to evaluate the total time until the job execution is completed for the extracted candidate combinations.

Here, the total time is the time from the start of GPU reconfiguration on the node to the completion of job execution. The total time is the sum of the GPU reconfiguration time on the node and the job execution time after reconfiguration. However, if GPU reconfiguration is not required, the reconfiguration time = 0.

Then, the node combination extraction unit 133a determines the combination candidate with the best evaluation value of the total time obtained by the evaluation unit 133b as the combination of nodes to be assigned to the job. The node combination extraction unit 133a responds to the node assignment unit 132 with the determined combination of nodes.

The evaluation unit 133b evaluates the total time of jobs for node combination candidates in response to a request from the node combination extraction unit 133a. In evaluating the total time, the evaluation unit 133b calculates the total time for job execution for each combination candidate. The effect of communication contention is taken into account in the job execution time. The presence or absence of the effect of communication contention is determined by a specified communication benchmark measurement that is performed in a relatively short time.

The evaluation unit 133b registers the results of the communication benchmark measurement in the communication performance table 123. The evaluation unit 133b also calculates the total time for each candidate node combination based on the reference communication time table 122 and the communication performance table 123, and registers the calculated total time in the evaluation value table 124.

Here, the evaluation unit 133b calculates the total time by the following formula (3).
T _Total = T _reconf + α × β × T _job ... (3)
T _Total is the total time. T _Total may be called an evaluation value of the total time. T _reconf is the reconfiguration time of the GPU in the node. T _reconf is given in advance to the evaluation unit 133b according to the hot plug/remove support status of the node. T _job is the main body of the job execution time. If reconfiguration is not required, T _reconf =0. For T _job , the upper limit value of the job execution time at the time of job execution input by the user is used. The units of T _Total , T _reconf and T _job are, for example, seconds.

α is a coefficient representing the influence of communication contention on the job execution time T _job . α is a magnification of the measurement result of the communication time by a communication benchmark program that is completed in a short time to the reference communication time (a value when there is no influence of communication contention). The reference communication time is a reference communication time obtained in advance using the communication benchmark program in a state where no job is assigned to each node (a state where there is no communication contention). Specifically, α = (communication time measured by the communication benchmark) / reference communication time. Note that, unlike the communication benchmark program, the job that is actually executed includes calculation processing other than communication. Therefore, α does not have to be the magnification itself to the reference communication time, and may be a value obtained by adjusting the magnification to the reference communication time. For example, an adjustment method for reducing the influence can be considered, such as setting the value obtained by multiplying the magnification by 0.5 to α.

β is a coefficient that represents the effect of communication contention on the job execution time T _job . β indicates the effect of communication contention according to the number of GPUs per node of the job to be executed. This is because it is considered that the more GPUs per node of the job, the more inter-node communications occur, and the greater the effect on the job execution time.

For example, if the number of GPUs per node is _NG , then β=1+(1/8)× _NG .
In this case, when the number of GPUs per node is 1, β=1.1. When the number of GPUs per node is 4, β=1.5. The constant in the denominator of the coefficient (1/8) multiplied by _NG may be, for example, the maximum number of GPUs that can be configured per node. The constant in the denominator is determined in advance. Note that when (communication time measured in communication benchmark measurement)/reference communication time≦1, this is a case where there is no effect of communication contention, and the evaluation unit 133b sets α=1 and β=1.

The coefficient expressed by α×β in equation (3) corresponds to the first coefficient a1 and the second coefficient a2 in the first embodiment.
FIG. 11 is a diagram illustrating an example of a job management table.

The job management table 121 includes fields for job number, program name, number of nodes, number of GPUs per node, and execution time. The job number field registers the job number, which is the identification number of the job. The program name field registers the program name of the job. The number of nodes field registers the number of nodes required to execute the job. The number of GPUs per node field registers the number of GPUs per node required to execute the job. The execution time field registers the upper limit of the execution time of the job specified by the user in the execution request.

For example, job management table 121 has a record with job number "1", program name "A", number of nodes "2", number of GPUs per node "1", and execution time "1:00:00". This record indicates that the program name of the job with job number "1" is "A", the requested number of nodes is "2", the requested number of GPUs per node is "1", and the upper limit of the job execution time is "1:00:00" (1 hour, 0 minutes, 0 seconds). Note that the job execution time specified in the job execution request does not take into account the effects of communication contention. The job execution time may be extended due to the effects of communication contention during actual job execution.

In the job management table 121, records of other job numbers are also registered.
FIG. 12 is a diagram illustrating an example of a reference communication time table.
The reference communication time table 122 includes items for the number of nodes and the reference communication time (msec). The number of nodes to which a job is assigned is registered in the number of nodes item. The reference communication time is registered in the reference communication time item. The reference communication time is a reference time for determining whether a communication contention will occur in communication between nodes. The reference communication time for each number of nodes is obtained in advance by measuring a communication benchmark before the system starts operating, and is registered in the reference communication time table 122. The unit of the reference communication time is msec (milliseconds).

For example, if a job is assigned to two nodes, and the communication time obtained by benchmarking the communication between the nodes is less than or equal to the reference communication time, it is determined that there is no communication contention. On the other hand, if the communication time obtained by benchmarking the communication between the nodes is longer than the reference communication time, it is determined that there is a communication contention.

For example, the reference communication time table 122 has a record with the number of nodes "2" and the reference communication time "4.70." This record indicates that when a job is executed on two nodes, the reference communication time for the two nodes is 4.70 msec.

Records for other numbers of nodes, such as when the number of nodes is "3", are also registered in the reference communication time table 122. Note that the communication time between nodes (including the reference communication time), for example, when there are three or more nodes, may be the average of the communication times between two nodes obtained in a communication benchmark measurement using those nodes, or it may be the maximum communication time, or it may be the minimum communication time.

FIG. 13 illustrates an example of a communication performance table.
The communication performance table 123 includes fields for item number, combination, communication time (msec), and ratio to standard. The item number, which is an identification number of a record, is registered in the item number field. The combination field registers a candidate combination of nodes to be assigned to a job. The communication time field registers the communication time obtained in a communication benchmark measurement for the node combination. The unit of communication time is msec. The ratio to standard field registers the ratio of the communication time to the standard communication time (= communication time ÷ standard communication time). This ratio corresponds to α in equation (3).

The example of the communication performance table 123 shows the results of measuring the communication performance by the evaluation unit 133b for candidate node combinations in which four nodes are selected from the nodes "0", "1", "2", "5", and "6" shown in FIG. 6.

For example, the communication performance table 123 has a record with item number "1", combination "0,1,2,5", communication time "5.30", and magnification "1.08" relative to the standard. This record indicates that the communication time between nodes obtained in a communication benchmark measurement of nodes "0", "1", "2", and "5", which are candidate node combinations, is 5.30 msec, and the magnification α of the communication time relative to the standard communication time for node number "4" is 1.08. The magnification α = 1.08 is calculated as α = 5.30 ÷ 4.90 = 1.08 using the standard communication time of 4.90 msec for node number "4" in standard communication time table 122.

In the communication performance table 123, records for other combination candidates are also registered.
FIG. 14 is a diagram illustrating an example of an evaluation value table.
The evaluation value table 124 includes fields for item number, combination, magnification to criterion, and evaluation value. In the field for item number, an item number that is an identification number of a record is registered. In the field for combination, a candidate combination of nodes to be assigned to a job is registered. In the field for magnification to criterion, a magnification α of the communication time to the reference communication time is registered. In the field for evaluation value, T _Total calculated by the evaluation unit 133b using formula (3) is registered.

The evaluation value table 124 shows the calculation result of T _Total for a job using four nodes (one GPU per node) as exemplified in Fig. 6. As an example, T _job = 3600 seconds, and T _reconf = 150. For the combination of nodes "0", "1", "2", and "5", no GPU reconfiguration is required.

For example, the evaluation value table 124 has a record of item number "1", combination "0,1,2,5", magnification with respect to the reference "1.08", and evaluation value "4276.80". This record indicates that the magnification α for the nodes "0", "1", "2", and "5" which are candidate node combinations is 1.08, and T _reconf calculated using the magnifications α and β is "4276.80". In the above example, β=1.1.

That is, for the combination "0, 1, 2, 5", T _Total = 1.1 * 1.08 * 3600 = 4276.80.
The evaluation value table 124 also registers T _Total for other combination candidates of nodes.

For the combination "0,1,2,6", T _Total =150+1.1*1.02*3600=4189.20.
For the combination "0,1,5,6", T _Total =150+1.1*1.12*3600=4585.20.

For the combination "0,2,5,6", T _Total = 150 + 1.1 * 1.10 * 3600 = 4506.00.
For the combination "1, 2, 5, 6", T _Total = 150 + 1.1 * 1.06 * 3600 = 4347.60.

In the example of evaluation value table 124, record number "1" corresponds to the case where the GPU is not reconfigured. Records numbered "2" to "5" correspond to the case where the GPU is reconfigured.

In the example of the evaluation value table 124, the node combination candidates "0", "1", "2", and "6" have the smallest T _Total . Therefore, the node allocation unit 132 acquires nodes "0", "1", "2", and "6" from the node selection unit 133 as node combinations to which the job is to be assigned, and assigns nodes "0", "1", "2", and "6" to the job. Node "6" is a node that requires GPU reconfiguration. Therefore, the GPU will be reconfigured in node "6" in order to assign the job.

In addition, when there are a plurality of combinations with the smallest T _Total , among which some combinations include a node requiring reconfiguration and some do not, the node combination extraction unit 133a may select an arbitrary combination from among the combinations that do not include a node requiring reconfiguration. In this way, unnecessary reconfiguration can be avoided.

Next, the processing procedure of the management device 100 will be described.
FIG. 15 is a flowchart showing an example of processing by the job scheduler.
(S10) When the job information acquisition unit 131 receives a job execution request, it acquires job information from the execution request and registers it in the job management table 121. The node allocation unit 132 acquires the job information registered in the job management table 121 and determines whether or not there are free nodes in the configuration required to execute the job. If there are free nodes in the configuration required to execute the job, the process proceeds to step S11. If there are not free nodes in the configuration required to execute the job, the process ends. For example, if there are not free nodes in the configuration required to execute the job, the node allocation unit 132 re-determines whether or not there are free nodes in the configuration required to execute the job for that job after a predetermined time has elapsed.

(S11) The node allocation unit 132 determines whether or not allocation within one rack is possible. If allocation within one rack is possible, processing proceeds to step S12. If allocation within one rack is not possible, processing proceeds to step S13. Allocation within one rack means that all nodes to which the job is to be assigned can be selected from the same rack.

(S12) The node allocation unit 132 allocates the job to a node in the corresponding rack. Then, the process proceeds to step S14.
(S13) The node allocation unit 132 requests node selection from the node selection unit 133. The node selection unit 133 executes node selection and returns the selection result to the node allocation unit 132. The node allocation unit 132 assigns the job to the node selected by the node selection unit 133. Details of node selection by the node selection unit 133 will be described later.

(S14) The node allocation unit 132 instructs the node assigned in step S12 or step S13 to execute the job. The node that received the instruction executes the job. Then, the processing of the job scheduler 130 ends.

FIG. 16 is a flowchart illustrating an example of a node selection process.
The node selection process corresponds to step S13.
(S20) The node combination extraction unit 133a extracts node combinations. The node combination extraction will be described in detail later.

(S21) The evaluation unit 133b evaluates the node combination. The evaluation of the node combination will be described in detail later. An evaluation value table 124 is created based on the evaluation of the node combination.
(S22) The node combination extraction unit 133a selects a combination of nodes to which the job is to be assigned based on the evaluation value (T _Total ) in the evaluation value table 124. At this time, the node combination extraction unit 133a compares each evaluation value T _Total in the evaluation value table 124, and selects the combination with the smallest T _Total . The node combination extraction unit 133a responds to the node assignment unit 132 with the nodes that belong to the selected combination. Then, the node selection process ends.

FIG. 17 is a flowchart illustrating an example of a process for extracting a node combination.
The process of extracting node combinations corresponds to step S20.
(S30) The node combination extraction unit 133a extracts a combination of configured nodes that is the same as the configuration required for job execution, based on the job information in the job management table 121. For example, the storage unit 120 holds information on the number of GPU connections in free nodes. The node combination extraction unit 133a can extract a combination of configured nodes based on the information stored in the storage unit 120.

For example, in the example of Figure 6, of the four node combinations selected from the free nodes "0", "1", "2", "5", and "6", the combination of nodes "0", "1", "2", and "5" is a combination of configured nodes that has the same configuration as the configuration required for job execution.

(S31) The node combination extraction unit 133a extracts node combinations that require a configuration change. For example, the node combination extraction unit 133a may extract node combinations that require a configuration change based on information on the number of GPU connections at free nodes stored in the storage unit 120. For example, in the example of FIG. 6, of the four node combinations selected from the free nodes "0", "1", "2", "5", and "6", any combination other than the combination of nodes "0", "1", "2", and "5" is a node combination that requires a configuration change. This is because node "6" requires a GPU reconfiguration. Then, the node combination extraction process ends.

The node combinations extracted in steps S30 and S31 become candidate combinations of nodes to which the job is assigned.
FIG. 18 is a flowchart illustrating an example of a process for evaluating a node combination.

The node combination evaluation process corresponds to step S21.
(S40) The evaluation unit 133b evaluates the total time T _Total for each node combination candidate using the evaluation formula (3). At this time, the evaluation unit 133b measures the communication time for each node combination candidate by performing a predetermined communication benchmark measurement for a relatively short time, registers the measured communication time in the communication performance table 123, and obtains α in formula (3). The communication benchmark measurement is performed by having each node included in the combination candidate execute a communication benchmark program for a relatively short time. At this time, if (communication time measured by the communication benchmark measurement)/reference communication time≦1, the evaluation unit 133b determines that there is no effect of communication contention, and sets α=1 and β=1 in formula (3). If (communication time measured by the communication benchmark measurement)/reference communication time>1, the evaluation unit 133b determines that there is an effect of communication contention, and sets α=(communication time measured by the communication benchmark measurement)/reference communication time, and β=1+(1/8)× _NG .

(S41) The evaluation unit 133b records T _Total calculated in the evaluation in step S40 for each combination candidate in the evaluation value table 124. Then, the evaluation process of the node combination ends.

FIG. 19 is a diagram showing an example of a difference in the total time related to the execution of a job.
Time charts 80, 81, and 82 each show an example of the total time from time t0 until the completion of job execution. In Fig. 19, the positive direction of time is from the left to the right in the figure.

A time chart 80 shows a total time T _TotalA in the case where the GPU configuration is not changed but is affected by communication contention.
The time chart 81 shows the total time T _TotalB when the configuration change eliminates the influence of communication contention and improves the execution time of the job itself. T _TotalB < T _totalA . In this way, even if there is an initial overhead due to the configuration change (reconfiguration), the total time may also be improved by improving the execution time of the job itself.

The time chart 82 shows the total time T _TotalC when the effect of communication contention is small and the increase in execution time due to the effect of communication contention is small even without making a configuration change. T _TotalC < T _TotalB . In this way, depending on the situation, if the effect of communication contention is small, the increase in execution time may be small and it may be better not to make a configuration change.

Therefore, the management device 100 calculates the total time for each candidate node combination, taking into account the resource reconfiguration time and the impact of communication contention between nodes when the job is executed, and selects the candidate combination with the shortest total time as the node combination to which the job is assigned.

This allows the management device 100 to appropriately decide whether or not to reconfigure resources such as GPUs in a node. In addition, the management device 100 can select a node to which a job is assigned so as to shorten the total time until the job is completed.

The functions of the management device 100 may be realized by any of the nodes 200, 200a, ... In that case, the information processing system 2 does not need to include the management device 100 or the management network 50.

As described above, the management device 100 executes the following process.
The job scheduler 130 calculates a first time required from reconfiguration to completion of a job when the job is assigned to a first node that is a candidate for the job assignment and requires resource reconfiguration to execute the job. At this time, the job scheduler 130 calculates the first time based on a reconfiguration time required for resource reconfiguration, an execution time of the job after the reconfiguration, and a first coefficient representing the effect on the execution time of a communication contention accompanying communication during execution of the job by the first node. The job scheduler 130 also calculates a second time required to complete a job when the job is assigned to a second node that is a candidate for the job assignment and does not require resource reconfiguration without resource reconfiguration. At this time, the job scheduler 130 calculates the second time based on a second coefficient representing the effect on the execution time of a job of a communication contention accompanying communication during execution of the job by the second node, and the execution time. The job scheduler 130 compares the first time with the second time, and if the first time is shorter than the second time, reconfigures the resources on the first node, and if the first time is equal to or longer than the second time, does not reconfigure the resources on the first node.

This allows the management device 100 to appropriately determine whether or not to reconfigure resources, taking into account the effect of communication contention. The above-mentioned coefficient α×β is an example of the first coefficient and the second coefficient.
The job scheduler 130 may calculate a first time for a first group of nodes that are candidates for job allocation and include the first node, and may calculate a second time for a second group of nodes that are candidates for job allocation and do not include the first node.

This allows the management device 100 to appropriately determine whether or not to reconfigure resources, taking into account the effects of communication contention. It also allows the management device 100 to determine the node group to which a job is assigned, i.e., the combination of nodes, so as to shorten the time it takes to complete the job.

For example, the job scheduler 130 calculates a first coefficient based on the measurement result of a first communication time of communication between nodes belonging to a first node group. The job scheduler 130 also calculates a second coefficient based on the measurement result of a second communication time of communication between nodes belonging to a second node group.

This allows the management device 100 to appropriately determine the first coefficient and the second coefficient. For example, the first coefficient can be calculated based on the ratio between the first communication time and a predetermined reference communication time. Also, the second coefficient can be calculated based on the ratio between the second communication time and the reference communication time.

At this time, the job scheduler 130 determines whether or not there is a communication conflict associated with the communication when the first node executes the job based on the first communication time and a predetermined reference communication time, and sets the first coefficient to 1 if there is no communication conflict. The job scheduler 130 also determines whether or not there is a communication conflict associated with the communication when the second node executes the job based on the second communication time and the reference communication time, and sets the second coefficient to 1 if there is no communication conflict.

This allows the management device 100 to appropriately determine the first coefficient and the second coefficient. For example, if the first communication time is equal to or less than the reference communication time, the job scheduler 130 determines that there is no communication contention and sets the first coefficient to 1. If the first communication time is longer than the reference communication time, the job scheduler 130 determines that there is a communication contention and sets the first coefficient based on the ratio between the first communication time and the reference communication time. Similarly, if the second communication time is equal to or less than the reference communication time, the job scheduler 130 determines that there is no communication contention and sets the second coefficient to 1. If the second communication time is longer than the reference communication time, the job scheduler 130 determines that there is a communication contention and sets the second coefficient based on the ratio between the second communication time and the reference communication time.

The job scheduler 130 also calculates the first and second coefficients based on the amount of resources per node required for the nodes used to execute the job.

Here, the greater the amount of resources per node used to execute a job, the greater the communication volume during job execution. For this reason, it is estimated that the greater the number of resources per node, the more susceptible the job execution time is to communication contention. Thus, by determining the first and second coefficients based on the amount of resources required for one node, the management device 100 can appropriately determine the first and second coefficients. Note that the number of GPUs and SSD capacity required per node for job execution are examples of the amount of resources.

The job scheduler 130 also calculates the first time as the sum of the product of the job execution time and the first coefficient and the resource reconfiguration time, and calculates the product of the execution time and the second coefficient as the second time.

This allows the management device 100 to appropriately calculate the first time and the second time.
Furthermore, the job scheduler 130 assigns the job to the first node when the resources in the first node are to be reconfigured, whereas the job scheduler 130 assigns the job to the second node when the resources in the first node are not to be reconfigured.

This enables the management apparatus 100 to shorten the total time until the execution of the job is completed.
The information processing in the first embodiment can be realized by causing the processing unit 12 to execute a program. The information processing in the second embodiment can be realized by causing the CPU 101 to execute a program. The program can be recorded in a computer-readable recording medium 113.

For example, the program can be distributed by distributing the recording medium 113 on which it is recorded. The program may also be stored in another computer and distributed via a network. For example, the computer may store (install) the program recorded on the recording medium 113 or a program received from another computer in a storage device such as the RAM 102 or HDD 103, and read and execute the program from the storage device.

Reference Signs List 1 Information processing system 10 Information processing device 11 Storage unit 12 Processing unit 20, 20a, 20b, 20c, 20d, . . . Node 30 Management network 40 Inter-node network 41, 42 Communication path

Claims (9)

On the computer,
calculate a first time required from the reconfiguration to the completion of the job when the job is allocated to the first node, based on a reconfiguration time required for the reconfiguration of a first node that is a candidate for allocation of the job and requires resource reconfiguration to execute the job, an execution time of the job after the reconfiguration, and a first coefficient that represents an influence on the execution time of a communication contention accompanying communication during execution of the job by the first node; calculate a second time required for the completion of the job when the job is allocated to the second node without the reconfiguration of the resources, based on the execution time and a second coefficient that represents an influence on the execution time of a communication contention accompanying communication during execution of the job by a second node that is a candidate for allocation of the job and does not require the resource reconfiguration;
comparing the first time with the second time, and performing the reconfiguration of the resource in the first node if the first time is shorter than the second time, and not performing the reconfiguration of the resource in the first node if the first time is equal to or longer than the second time;
A resource reconfiguration program that accompanies job execution to execute processing. calculating the first time for a first node group which is a candidate for allocation of the job and which includes the first node, and calculating the second time for a second node group which is a candidate for allocation of the job and which does not include the first node;
2. The resource reconfiguration program according to claim 1, which causes the computer to execute a process. calculating the first coefficient based on a measurement result of a first communication time of communication between nodes belonging to the first node group, and calculating the second coefficient based on a measurement result of a second communication time of communication between nodes belonging to the second node group;
3. The resource reconfiguration program according to claim 2, which causes the computer to execute a process. determining whether or not a communication conflict occurs in communication by the first node when the job is executed based on the first communication time and a predetermined reference communication time, and setting the first coefficient to 1 if the communication conflict does not occur; determining whether or not a communication conflict occurs in communication by the second node when the job is executed based on the second communication time and the reference communication time, and setting the second coefficient to 1 if the communication conflict does not occur;
4. The resource reconfiguration program according to claim 3, which causes the computer to execute a process. calculating the first coefficient and the second coefficient based on an amount of the resource per node required for a node used to execute the job;
2. The resource reconfiguration program according to claim 1, which causes the computer to execute a process. calculating a sum of a product of the execution time and the first coefficient and the reconstruction time as the first time, and calculating a product of the execution time and the second coefficient as the second time;
2. The resource reconfiguration program according to claim 1, which causes the computer to execute a process. assigning the job to the first node when the reconfiguration of the resources in the first node is to be performed, and assigning the job to the second node when the reconfiguration of the resources in the first node is not to be performed;
2. The resource reconfiguration program according to claim 1, which causes the computer to execute a process. The computer
calculate a first time required from the reconfiguration to the completion of the job when the job is allocated to the first node, based on a reconfiguration time required for the reconfiguration of a first node that is a candidate for allocation of the job and requires resource reconfiguration to execute the job, an execution time of the job after the reconfiguration, and a first coefficient that represents an influence on the execution time of a communication contention accompanying communication during execution of the job by the first node; calculate a second time required for the completion of the job when the job is allocated to the second node without the reconfiguration of the resources, based on the execution time and a second coefficient that represents an influence on the execution time of a communication contention accompanying communication during execution of the job by a second node that is a candidate for allocation of the job and does not require the resource reconfiguration;
comparing the first time with the second time, and performing the reconfiguration of the resource in the first node if the first time is shorter than the second time, and not performing the reconfiguration of the resource in the first node if the first time is equal to or longer than the second time;
A method for reconfiguring resources when a job is executed. a storage unit that stores a reconfiguration time required for a first node that is a candidate for a job and requires resource reconfiguration for executing the job, and an execution time of the job after the reconfiguration;
a processing unit that calculates a first time required from the reconfiguration to completion of the job when the job is assigned to the first node, based on the reconfiguration time, the execution time, and a first coefficient representing an influence on the execution time of a communication contention accompanying communication during execution of the job by the first node, and calculates a second time required to complete the job when the job is assigned to the second node without performing the reconfiguration of the resource, based on the execution time and a second coefficient representing an influence on the execution time of a communication contention accompanying communication during execution of the job by a second node that is an assignment candidate for the job and does not require the reconfiguration of the resource, compares the first time with the second time, and performs the reconfiguration of the resource in the first node if the first time is shorter than the second time, and does not perform the reconfiguration of the resource in the first node if the first time is equal to or longer than the second time;
An information processing system having the above configuration.

Images