JP2016075968A - Information processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve a problem that it is hard to achieve both of cost reduction and performance when conducting communication by using an optical signal in a parallel computer having many computation nodes.SOLUTION: An information processing device includes a plurality of calculation nodes, a communication path using an optical signal, an optical switch switching the communication path, a control unit for allocating a job to the plurality of calculation nodes, and a control unit for generating a signal for switching the optical switch. The information processing device characteristically allocates the jobs and switches the optical switch according to the information in a table where the performance of the calculation nodes, performance of the communication path using the optical signal, and performance of a calculation object are arranged.SELECTED DRAWING: Figure 12

Description

  The present invention relates to a parallel information processing apparatus to which a plurality of calculation nodes are connected, and more particularly to a parallel information processing apparatus in which a plurality of calculation nodes are connected using optical wiring and optical switches.

  A parallel computer is generally used as a computer for performing large-scale information processing. In a parallel computer, a plurality of calculation nodes are connected by wiring, a job to be calculated is assigned to each calculation node, and calculation is performed by exchanging data between the calculation nodes.

  In particular, when computing abstract data composed of nodes (vertices) called graphs and edges (branches) that represent the connection between nodes, the communication performance between computation nodes greatly increases the overall information processing performance. Influence. Therefore, in these parallel computers, it is necessary to improve the communication speed between the computation nodes. One means for improving the communication speed is communication using an optical signal. In conventional communication using electrical signals, for example, if communication between computing nodes of about 1 m is performed at a speed of 25 Gbps, power consumption required for communication increases dramatically, but communication using optical signals reduces power consumption. It becomes possible to suppress.

  Communication using an optical signal requires a communication device using a light emitting element. However, the light emitting element has a problem that it is less reliable than a conventional element that handles an electric signal and is likely to break down.

  Therefore, in order to solve this problem, there has been a technique in which a redundant path is provided in the optical communication path, and communication is performed by switching to the redundant path when the light emitting element fails (Patent Document 1).

JP 2010-514366 A

  However, in the method described in Patent Document 1, it is necessary to prepare a communication device including an extra light emitting element, which increases the cost. In particular, in a parallel computer with a large number of computing nodes, the number of communication paths increases, and the cost of redundant communication paths becomes enormous. From another point of view, if a redundant communication path can be used before an element breaks down, higher-speed communication is possible, and the invention described in Patent Document 1 reduces communication performance if considered at the same cost. Yes. Therefore, when performing communication using an optical signal, it has been a problem to achieve both cost and performance.

  In the present invention, the job is arranged in the calculation node in consideration of the communication amount required for the job to be executed, and is rearranged to the optimum communication path configuration for the arranged job.

  Among the means for solving the problems according to the present invention, typical ones are as follows.

  First, an information processing apparatus, a plurality of calculation nodes, a communication path using optical signals, an optical switch for switching communication paths, a control unit that assigns jobs to the plurality of calculation nodes, and an optical switch A control unit that generates a signal for switching between, and according to the information of the table in which the performance of the calculation node, the communication path using the optical signal and the performance of the target to be calculated are arranged, a job with a large amount of necessary communication has a large communication capacity It is characterized in that the connection of the optical switch is switched so as to be allocated to the calculation nodes and to secure a necessary communication amount between the calculation nodes.

  According to the present invention, in an information processing apparatus using an optical communication path, it is possible to perform information processing without degrading performance as much as possible even if there is a failure in the optical communication path. Furthermore, it is possible to maximize performance by utilizing resources such as communication paths possessed by the information processing apparatus. Furthermore, it is possible to maximize the performance by minimizing the cost of the communication path and the like.

1 is a basic configuration diagram of a parallel computer using the present invention. An example of a computation node. A table holding the network capacity of each compute node. A table that holds the amount of communication between jobs to be calculated. Optical switch and job assignment table. A table that holds the amount of communication between jobs to be calculated. Optical switch and job assignment table. The block diagram after the optical switch switching of the parallel computer using this invention. The figure which shows a state when a failure generate | occur | produces in the communication path of a parallel computer. A table holding the network capacity of each compute node. Optical switch and job assignment table. The state diagram when a failure occurs in the parallel computer using the present invention. (A) And (b) of this figure is a switch switching diagram for specifying a failure location. The flowchart for specifying a failure location. 1 is a basic configuration diagram of a parallel computer using the present invention. A table holding the network capacity of electrical switches. A table that holds the amount of communication between jobs to be calculated. Optical switch and job assignment table. A table holding the network capacity of electrical switches. Optical switch and job assignment table. The block diagram after the optical switch switching of the parallel computer using this invention. A table holding the network capacity of each compute node. A table that holds the amount of communication between jobs to be calculated. Optical switch and job assignment table. The block diagram after the optical switch switching of the parallel computer using this invention. The block diagram after the optical switch switching of the parallel computer using this invention. A table holding network uptime. The block diagram after the optical switch switching of the parallel computer using this invention. The table | surface which hold | maintained the emitted light intensity of the optical element in a network. The block diagram after the optical switch switching of the parallel computer using this invention. A table holding the ambient temperature of the light emitting elements of the network. Job execution flow. Job assignment to compute nodes and optical switch switching flow. Job assignment to compute nodes and optical switch switching flow. The job execution flow when a failure occurs in the communication path. The job execution flow when a failure occurs in the communication path.

  In a typical information processing apparatus of the present invention, in a parallel information processing apparatus using a communication path using a plurality of calculation nodes and optical elements, even if there is a failure in the optical communication path, the information processing is not performed. Can be done. Embodiments of the present invention will be described below in detail with reference to the drawings.

  FIG. 1 shows a configuration of an information processing apparatus using the present invention. In FIG. 1, NODE1 to NODE4 are computation nodes that perform operations, OSW1 is a 12-input x 12-output optical switch, CONT1 assigns a job to the computation node and controls the optical switch, and JCONT1 assigns a job to the computation node The signal OCONT1 represents a control signal for switching the path of the optical switch. Three communication paths are output from each computation node, and are connected to other computation nodes via optical switches. When information processing is performed with this configuration, a job is allocated to each of a plurality of calculation nodes and calculation is performed. When the result calculated in a certain job is transferred to another job, data is transferred through a communication path using an optical switch. Further, when sending data from one node to another node, it is possible to send data via another node even if there is no direct route. For example, even if there is no direct communication path when sending data from NODE1 to NODE4, if there is a route from NODE1 to NODE2 and NODE2 to NODE4, first send the data from NODE1 to NODE2, and then send the data from NODE2 to NODE4. Data communication is possible by sending to. However, when there is no direct route, it takes time to transfer data, and the performance deteriorates as compared with the case where there is a direct route.

  An example of the computation nodes NODE1 to NODE4 is shown in FIG. The computation node is a board (BOARD in the figure) on the board (CPU in the figure), a memory to store calculation data (MEM in the figure), and temporarily stores data with a larger capacity than the memory. Storage (STO in the figure) and communication card (HCA in the figure) that outputs three communication paths (LAN in the figure) are arranged. The CPU is a general CPU, and a DRAM is usually used as a memory, and a hard disk drive or an SSD (Solid State Drive) is used as a storage. Ethernet, Infiniband, PCE Express, etc. are used for the communication path. In this embodiment, three communication paths are output from each computation node, and a communication path using an optical signal is used here. In FIG. 3, the number of communication paths of each computation node, that is, the capacity of the communication path is represented by a table. Since the state of FIG. 1 shows the case where all the communication paths of each node are operating, the network capacity of each node is 3.

  FIG. 4 shows an example of the communication amount between each job in the parallel calculation processed by this information processing apparatus. In the parallel calculation, the processing to be calculated is divided into jobs of the number of nodes, and the job is assigned to each node of the information processing apparatus. At this time, communication is generated because calculation results need to be exchanged between nodes.

  When the calculation process is divided, the communication amount necessary for each job is known, so the communication amount distribution shown in the table of FIG. 4 is obtained. In the information processing apparatus using the present invention, each job is assigned to each node in consideration of the network capacity possessed by each node in FIG. 3 and the traffic required between each job in FIG. The wiring for performing communication is switched by an optical switch, and processing is executed. In this example, because the network capacity of each node is equal and the necessary communication amount is equal, jobs are assigned as shown in FIG. 5 and optical switch connections are also equally assigned between the nodes. The four columns at the center of the table of FIG. 5 indicate the number of communication paths connecting the nodes, and the four nodes are evenly connected one by one. The right column shows job assignment, and JOB1 to JOB4 are assigned to NODE1 to NODE4, respectively. In this case, the assignment of communication paths by the optical switch is as shown in FIG.

  FIG. 32 shows a calculation execution procedure in the information processing apparatus using the present invention. Job assignment to a calculation node and switching of an optical switch are executed according to the characteristics of the job to be executed first, and the job is executed after job assignment and switching of the optical switch are completed. When the execution of the job is completed, the job assignment to the calculation node and the switching of the optical switch are performed again according to the contents of the next job.

  By repeating this, the job is executed and the calculation is executed using the optimum calculation node assignment and network configuration for each job. FIG. 33 shows a flow when executing job assignment and optical switch switching in FIG. First, a communication amount table required for each job is created based on job characteristic information. Next, based on the information in the table, each job is assigned to a calculation node in the information device so that the required information amount matches the network capacity of the information device. Next, the optical switch is switched so as to realize a network configuration necessary for each computation node to which each job is assigned. When the switching of the optical switch is finished, the calculation is performed.

  FIG. 34 shows an example of a flowchart for executing the flow of FIG. In step 101, the total amount of communication required for each job is calculated from the table of required communication amount between jobs. This gives the table in FIG. In step 102, jobs with a large communication capacity are assigned in order from jobs with a large required communication volume. With the conditional branches and loops of steps 102 to 104, all jobs can be assigned in the order of the required amount of communication. Next, in step 106, a table of necessary traffic between nodes is created.

  This can be easily created by writing out the necessary communication amount between the nodes based on the assigned job information. Next, in step 107, the network possessed by each node is allocated to the required communication amount by proportional distribution. For example, the table of FIG. 3 is used as the network capacity of each node used at this time. The branches and loops of steps 108 to 110 are flows for preventing a node that needs a network but is not actually connected to the network when a network capacity is simply distributed proportionally to each node.

  For example, if the required traffic between node 1 and node 2 is 100, the required traffic between node 1 and node 3 is 1, and the communication path (network capacity) that can be allocated to node 1 is 3, A certain communication path is allotted between node 1 and node 2. To prevent this, if there is a node that has no communication path, the network is not connected even though the network connection is necessary by re-allocating the network by making the required communication amount twice larger for convenience. Prevent nodes from forming. The table of FIG. 5 is obtained according to the assignment result. If the optical switch is switched as shown in this table, jobs can be arranged and calculation can be performed.

  Here, a case is considered where the content of calculation processed by the information processing apparatus in the present example changes, and the necessary communication amount between jobs changes. Figure 6 shows the amount of communication required between each job that has changed. Compared to the case of FIG. 4, it can be seen that the required communication volume between JOB1 and JOB3 increases, while the required communication volume from JOB4 to JOB1 and JOB3 decreases. When this process is performed using the communication paths shown in FIGS. 1 and 5, communication between JOB1 and JOB3 becomes a bottleneck. Therefore, this bottleneck can be eliminated by changing the communication path using the optical switch.

  The table in FIG. 7 shows how to assign optical switches for this purpose. While the number of communication paths between NODE1 and NODE3 to which JOB1 and JOB3 are assigned is increased to 2 from the case of Figure 5, the number of paths between NODE4 and NODE1 and NODE3 to which JOB4 that does not require communication is assigned is reduced. 0. FIG. 8 shows a configuration when this is actually switched by an optical switch. Unlike the case of FIG. 1, it can be seen that there are two communication paths between NODE1 and NODE3, but there are no communication paths from NODE1 and NODE3 to NODE4.

  That is, it can be seen that by switching the communication path by the optical switch, a computer configuration corresponding to the processing to be calculated can be taken, and the performance can be improved. Also in this case, it is possible to execute assignment of each job to each calculation node and switching of the optical switch using the flow of FIG. In this case, the table generated in step 101 of FIG. 34 is the table of FIG. 6, and the table of FIG. 3 is used as the traffic volume table of each node. The table shown in FIG. 7 is generated as a table representing switch connections. If the optical switch is switched according to the table of FIG. 11, the job can be arranged in the calculation node and executed.

  Furthermore, there is a feature that a failure is likely to occur in the communication path, and in particular, a light emitting element required when using an optical signal is easily broken. When the present invention is used, a calculation may be performed without degrading performance when a failure occurs in a communication path.

  FIG. 9 shows the state of the information processing apparatus when a failure occurs in the communication path. In this example, one of the three communication paths of NODE3 has failed. At this time, NODE3 senses that the communication path has failed and notifies the control unit. Accordingly, the network capacity table of each node held by the control unit is rewritten as shown in FIG. Along with this, the jobs and optical switches are reassigned based on the necessary communication amount of each job shown in FIG. 6 and the network capacity of FIG.

  An example of job and switch assignment results is shown in FIG. Since NODE3 has a low network capacity, JOB4, which requires less communication, is assigned to NODE3, and JOB3 is assigned to NODE4. Since the amount of communication between JOB1 and JOB3 is large, the number of connections between NODE1 and NODE4 is increased to 2. FIG. 12 shows the job assignment state and the optical switch connection.

  Even when the same processing as in FIG. 8 is executed, it can be seen that the job assignment and the optical switch connection state are different. In addition, at this time, although a part of the communication path is broken in FIG. 12, the necessary communication amount can be secured, so that it can be seen that the performance does not deteriorate in FIG. 12 compared to FIG.

  Also in this case, it is possible to execute assignment of each job to each calculation node and switching of the optical switch using the flow of FIG. In this case, the table generated in step 101 in FIG. 34 is the table in FIG. 6, and the table in FIG. 10 is used as the traffic volume table of each node. The table shown in FIG. 11 is generated as a table representing switch connections. If the optical switch is switched according to the table of FIG. 11, it is possible to place a job on a calculation node and execute the processing without degrading the calculation performance even when there is a failure in the communication path.

  Here, when the communication path fails, it is possible to easily locate the failure point when there is a function such as generating an error signal when a failure occurs in a component constituting the communication path or turning on a lamp. . In particular, when a fault signal is generated, the network capacity table shown in Fig. 3 can be easily rewritten by inputting the signal to the control circuit of the optical switch, and the configuration of the job and optical switch can be changed. I can do it.

  In addition, when a component does not have the function, the failure location can be specified by controlling the communication path. For example, in Fig. 13 (a), if the communication output at the top of NODE3 fails, the exchange of data between NODE1 and NODE3 changes to the path A that connects the top communication path of NODE1 and the top of NODE3. It is possible to detect that a failure has occurred, but it is not known where the failure occurred on the path A. Therefore, in this case, the failure location is specified according to the sequence of FIG. When a failure occurs, the job being executed at that time is stopped. Since it is understood that a failure has occurred on the path A, the optical switch is then switched to the state shown in FIG.

  In this state, if the communication path of route B is not connected, it turns out that the communication channel at the top of NODE3 has failed. If the communication route of route C is not connected, the communication channel at the top of NODE1 has failed. I understand that. Therefore, by reflecting this information on the communication capacity table of each node and switching the job and the optical switch, it is possible to bypass the failure location without degrading the performance.

  FIG. 35 shows a job execution flow when a failure occurs in the communication path. In a situation where no failure has occurred, the job is executed after the job assignment and the optical switch switching, as in FIG. Here, for example, consider a case where a failure occurs in the communication path during execution of job B. For example, when Infiniband is used for communication between nodes, for example, the communication performance is lowered unless the communication path is completely blocked, but the communication itself can be executed. Therefore, the job execution time becomes long, but the job B being executed at that time can be executed. In the information communication apparatus using the present invention, when the job B is completed, the network capacity table of each node is rewritten based on the information of the network where the failure has occurred.

  Using the table, job assignment to each computation node and switching of the optical switch are executed in the flow of FIG. 34, and the next job C is executed.

  Also, when a communication path failure occurs, if the communication path of the connected node is completely lost, or if Ethernet is used for communication between nodes, the communication path will fail. The network is completely shut down when it occurs. FIG. 36 shows the job execution flow in this case. In that case, the job being executed stops, and the node to which the job is assigned detects the stop. At that time, the network capacity table of each node is rewritten, and the information in that table is used to assign jobs to each computation node and switch the optical switch in the flow of FIG. Execution is performed.

  In the present embodiment, the control circuit CONT1 in FIG. 1 is illustrated as a component different from the calculation node, but in reality, any one of the calculation nodes can play the role. Further, if there are a plurality of calculation nodes, communication paths, and optical switch ports, the same effect can be obtained even if the number changes.

  As described above, according to the present invention, in an information processing apparatus having a plurality of computation nodes, it is possible to execute information processing without degrading performance even when there is a failure in a communication path.

  FIG. 15 shows a configuration of an information processing apparatus using the present invention. In FIG. 15, NODE1 to NODE8 are computation nodes that perform computation, OSW1 is an optical switch that switches a communication path using an optical signal, and ESW1 to ESW4 are electrical switches that allocate data when communicating between a plurality of computation nodes and the optical switch , CONT2 is a control circuit that assigns jobs to calculation nodes and controls routing of electrical switches and connection of optical switches, JCONT2 is a signal for assigning jobs to calculation nodes, OCONT2 is a control signal for switching the path of optical switches, ECONT2 represents a control signal for controlling the routing of the electrical switch.

  This embodiment is different from the first embodiment in that an electrical switch is inserted between the calculation node and the optical switch, and communication using an optical signal is performed between the electrical switch and the optical switch. Communication between the calculation node and the electrical switch can be performed using communication using an electrical signal, or can be performed using communication using an optical signal. When data is sent from one computation node to another computation node, the data is transferred to the computation nodes connected to the same electrical switch only through the electrical switch.

  For this reason, the performance of the data transfer is high. On the other hand, when data is exchanged between computation nodes that are not connected to the same electrical switch, for example, when sending data from NODE1 to NODE5 in FIG. 15, data is sent from NODE1 to electrical switch ESW1, where the communication destination is Data is sent to the electrical switch ESW3 via the optical switch via the communication path using the optical signal. In ESW3, NODE5 which is a data transfer destination is selected, and data is transferred from the electrical switch to NODE5.

  In addition, when data is transferred from NODE1 to NODE8 when the optical switch is in the state shown in FIG. 15, NODE1 to ESW1, ESW1 to ESW2 via optical switch OSW1, and ESW2 to optical switch OSW1. The data is transferred to ESW4, and finally the data is transferred from ESW4 to NODE8. In this way, when transferring data between computation nodes, data transfer between nodes connected to the same electrical switch is the fastest, and then connected to a different electrical switch. Communication between nodes connected by a switch switching path is the next fastest, and it is the slowest when connected to different electrical switches and the electrical switches are not connected by a path via an optical switch. .

  Therefore, when performing computations with this information processing device, jobs with the largest amount of traffic are allocated to the nodes connected to the same electrical switch, and jobs with the next largest amount of traffic are connected to the electrical switch connected with the optical switch. Performance increases when allocated to different compute nodes.

  FIGS. 16 and 17 show an example of the optical network capacity of each electrical switch and the amount of communication between jobs to be calculated. The control circuit CONT2 in FIG. 15 has this information as a table, performs optimization, and determines the connection of the optical switch and the assignment of the job to each node as shown in FIG. This determined information is used to control the optical switch with the optical switch control signal OCONT2 in FIG. 15 and to allocate to each job node with the job allocation signal JCONT2. In addition, when actually sending a signal, it is necessary for the electrical switch to have information on how each electrical switch is connected and which destination data is sent via which path. The information is transferred to each electric switch by a signal ECONT2 for controlling

  Here, when a failure occurs in the optical communication path, the information on the capacity of the optical network of each electrical switch is changed as shown in FIG. This failure can be detected by using the same method as in the first embodiment. FIG. 20 shows the state of the optical switch and the assignment of each job determined from the changed information and the information of each job shown in FIG. The state of the optical switch that realizes this is shown in FIG. At this time, although a part of the communication path is broken in FIG. 21, it can be seen that the required communication amount is secured, so that the performance is not lowered as compared with FIG.

  As described above, according to the present invention, in an information processing apparatus having a plurality of computation nodes, it is possible to execute information processing without degrading performance even when there is a failure in a communication path.

  In the first embodiment, a communication path failure using an optical signal is detected and job assignment and optical switch path switching are performed. However, in the current communication path, a part of the communication component fails. However, it is possible to perform communication with reduced performance. For example, in Infiniband, when some parts are broken, it is possible to communicate at a communication speed of 40 Gbps at normal times and at 10 Gbps at the time of partial failure.

  Therefore, it is possible to perform job assignment and optical switch switching in a computer using a state in which the network performance of the communication path is not a binary value indicating whether the communication path is normal or failed, but the performance is degraded.

  The third embodiment is different from the first embodiment in that the performance of the communication path can be seen in detail.

  FIG. 22 shows an example of network capacity information of each computation node. Since each compute node has three network routes, the performance of each route is shown.

  Further, FIG. 23 shows the amount of communication between jobs to be calculated. From these two tables, FIG. 24 shows the optical switch switching and job assignment so as to reduce the traffic as much as possible. A device configuration for realizing this is shown in FIG. Reference numerals 1001 and 1002 in FIG. 25 denote terminals with degraded performance. The difference from the first embodiment is that the performance of each communication path is finely fed back to the control circuit CONT3 by a signal JCONT3 connecting the control circuit CONT3 and each calculation node.

  As described above, according to the present invention, in an information processing apparatus having a plurality of computation nodes, it is possible to execute information processing without degrading performance even when there is a failure in a communication path.

  An optical element used for communication using an optical signal has a property that it easily breaks down when an operation time becomes long. In the present invention, the network is configured so as not to use an element having a long operation time as much as possible. FIG. 26 shows the configuration of an information processing apparatus using the present invention. The difference from FIG. 1 of the first embodiment is that OTIME for storing the operation time of the optical element, a signal OINFO for transmitting the operation status of the optical element to OTIME, and information on the operation time of the optical element are added to the calculation node. It has a control circuit CONT4 for controlling job allocation and optical switch switching. In this configuration, the network operation time is stored as shown in FIG. 27, and the optical switch is switched so as not to use an element having a long operation time. In this example, since the operation time of the network device included in NODE 3 is long, the optical switch OSW1 is switched to suppress the operation time as shown in FIG.

  As described above, according to the present invention, in an information processing apparatus having a plurality of calculation nodes, it is possible to execute information processing without causing a failure of the communication apparatus and further degrading performance.

  An optical element used for communication using an optical signal has a property that light emission intensity decreases before failure. In the present invention, the light emission intensity is measured, and the network is configured so as not to use an element having a low light emission intensity as much as possible. FIG. 28 shows the configuration of an information processing apparatus using the present invention.

  The difference from FIG. 1 of Example 1 is that the OCALC that converts the emission intensity of the optical element into data, the measuring device OMEAS that measures the emission intensity of the optical element, and the calculation node taking into account the information of the emission intensity of the optical element A control circuit CONT5 for controlling job allocation and switching of optical switches.

  In this configuration, the light emission intensity of the optical element is tabulated as shown in FIG. 29, and the optical switch is switched so as not to use the element having the light emission intensity lowered. In this example, since the light intensity of the optical element of the network device possessed by NODE3 is low, the optical switch OSW1 is switched as shown in FIG. 28 to suppress the operation time. In OCALC, it is also possible to control the switch by inputting the measured luminescence intensity information to CONT5 without processing, and it has a memory circuit in OCALC that measures and memorizes the change in luminescence intensity. It is also possible to control the switch so as not to use an optical element whose emission intensity is lower than that of the optical element.

  As described above, according to the present invention, in an information processing apparatus having a plurality of calculation nodes, it is possible to execute information processing without causing a failure of the communication apparatus and further degrading performance.

  An optical element used for communication using an optical signal has a property of easily failing when operated in a high temperature state. In the present invention, the temperature around the optical element is measured, and the network is configured so as not to use the element at a high temperature as much as possible.

  FIG. 30 shows a configuration of an information processing apparatus using the present invention. The difference from FIG. 1 of Example 1 is that the TCALC that converts temperature information around the optical element into data, the thermometer TMEAS that measures the temperature around the optical element, and the temperature information around the optical element are taken into the calculation node. The control circuit CONT6 controls the job allocation and switching of the optical switch. In this configuration, temperature information around the optical element is tabulated as shown in FIG. 31, and the optical switch is switched so as not to use an element having a high ambient temperature. In this example, the optical switch OSW1 is switched as shown in FIG. 30 because the ambient temperature of the optical element of the network device included in the NODE 3 is high. In TCALC, it is possible to control the switch by inputting the measured temperature information into CONT6 without processing it, and control the switch so that the temperature change is measured and the optical element whose temperature has risen is not used. Is also possible.

  As described above, according to the present invention, in an information processing apparatus having a plurality of calculation nodes, it is possible to execute information processing without causing a failure of the communication apparatus and further degrading performance.

CONT1 to CONT6: Job and optical switch control elements
NODE1 ~ NODE8 ... compute node
OSW1 ... Optical switch
JCONT3: Job control signal
OCONT2: Optical switch control signal
CPU: Arithmetic LSI, CPU
MEM ... Memory LSI
STO ... Storage device (hard disk, SSD)
HCA ... Communication card
LAN: Communication path
JOB1 ~ JOB8 ... Calculation job
OTIME: Element that stores the operating time of the communication device
OINFO: Signal indicating the operating status of the communication device
OCALC: An element that converts light intensity in a communication device into data
OMEAS ... Light intensity measuring device for optical elements in communication devices
TCALC: Element that converts the ambient temperature of optical elements in communication equipment into data
TMEAS ... Ambient temperature measuring device for optical elements in communication devices
101-112 ... Job assignment and optical switch switching flow steps

Claims (16)

Multiple compute nodes;
A communication path using optical signals;
An optical switch for switching the communication path;
A controller that assigns jobs to a plurality of compute nodes;
A first control unit for generating a signal for switching the optical switch,
According to the information in the first table that describes the number of communication paths each calculation node has and the required communication volume of the job to be calculated,
Jobs are assigned to compute nodes with the largest number of communication paths in descending order of the required communication volume of jobs,
To perform data communication from the first computation node to which the first job is assigned to the second computation node to which the second job is assigned, and
Information processing characterized by setting an optical switch provided between the plurality of calculation nodes in order to set an optical communication line between the first calculation node and the second calculation node apparatus.   The information processing apparatus according to claim 1, wherein the optical switch is switched so that a faulty communication path is not used when there is a fault in the communication path.   The information processing apparatus according to claim 2, wherein a communication path failure occurs in an optical element that transmits and receives an optical signal.   The information processing apparatus according to claim 3, wherein when there is a failure in the communication path, the control unit that allocates the job assigns the job so that an unused communication path is generated.   The information processing apparatus according to claim 4, wherein the control unit that assigns the job is realized by using a part of a function of the computation node.   5. The information processing apparatus according to claim 4, wherein a plurality of communication paths are used by switching an optical switch in order to identify a place where the communication path has failed. A storage circuit for storing the operating time of the communication path;
And a signal for transmitting the operation status of the communication path,
The information processing apparatus according to claim 1, wherein the optical switch is switched so as not to use a communication path having a long operation status of the communication path.   The information processing apparatus according to claim 7, wherein an operation time of an optical element used for communication is used as an operation time of a communication path. It has a device that measures the light emission intensity of the optical element used for communication,
The information processing apparatus according to claim 1, wherein the optical switch is switched so as not to use a communication path including a light emitting element having a reduced emission intensity. A circuit for storing a history of light emission intensity;
The information processing apparatus according to claim 9, wherein the optical switch is switched so as not to be used with a communication path including a light emitting element whose light emission intensity is lower than before. It has a thermometer that measures the temperature of the peripheral circuits of parts used for communication,
The information processing apparatus according to claim 1, wherein the optical switch is switched so as not to use a communication path including a component having a high temperature. The communication path includes an optical element,
The information processing apparatus according to claim 11, wherein the temperature around the optical element is measured. Have multiple electrical switches,
The communication path output from the computation node is connected to the electrical switch,
The information processing apparatus according to claim 1, wherein a part of a communication path output from the electrical switch is connected to the optical switch. The information processing apparatus according to claim 13, further comprising: a signal that controls a path output from the electrical switch; and a second control unit that outputs a signal that controls the path output from the electrical switch.   15. The information processing apparatus according to claim 14, wherein the second control unit determines an output path based on optical switch switching information using information in the first table. Multiple compute nodes;
An optical switch in which the plurality of calculation nodes are connected by a communication path using an optical signal;
A first controller that assigns jobs to a plurality of compute nodes;
A second control unit for generating a signal for switching the optical switch,
The first controller assigns jobs to the computation nodes so that jobs executed on the computation nodes are assigned to computation nodes with many communication paths using optical signals in the order of traffic, and the second controller is an optical switch. Information processing apparatus characterized by switching connection of.

Images