Architecture reconstruction method and system in computing cluster and storage medium

Abstract

The invention discloses an architecture reconstruction method and system in a computing cluster and a storage medium, and relates to the technical field of computers. The method is applied to a computing cluster, the computing cluster at least comprises a cluster management node and a computing chip, and the method comprises the following steps: acquiring a cluster segmentation instruction issued by a management scheduling platform through the cluster management node; a target computing chip is determined based on the cluster segmentation instruction and the chip identification information of the computing chip through the cluster management node, and the chip identification information of the target computing chip is fed back to the management scheduling platform; and the computing chip responds to a networking instruction issued by the management scheduling platform based on the chip identification information, and a cluster subsystem is constructed based on the chip identification information target of the target computing chip. According to the embodiment of the invention, the computing cluster is segmented into a plurality of cluster subsystems, and the computing system is quickly reconstructed, so that the computing cluster carries out parallel computing on multiple tasks, different computing power and business scene requirements are met, flexible adaptation is realized, and the computing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, system and storage medium for architecture reconstruction in a computing cluster. Background Art

[0002] With the vigorous development of artificial intelligence, biomedicine, and new materials research, the demand for computing power has shown explosive growth, especially in the direction of three-dimensional simulation computing, which requires large-scale computing clusters to adapt.

[0003] However, in the existing technology, tasks with small computing loads will also monopolize computing power, and multiple computing tasks cannot be calculated in parallel. This is not flexible enough, resulting in computing task accumulation and reduced computing efficiency. Therefore, how to divide a large computing system into several independent sub-computing systems without changing the hardware architecture to avoid one computing task occupying the resources of an entire computing cluster has become an urgent problem to be solved. Summary of the invention

[0004] The present invention provides a method, system and storage medium for reconstructing the architecture in a computing cluster to solve the problem that tasks with small computing amounts will monopolize computing power and multiple computing tasks cannot be calculated in parallel.

[0005] According to one aspect of the present invention, a method for reconstructing architecture in a computing cluster is provided, wherein the method is applied to a computing cluster, the computing cluster at least comprising a cluster management node and a computing chip, and the method comprises:

[0006] Obtain the cluster splitting instruction issued by the management and scheduling platform through the cluster management node;

[0007] Determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip through the cluster management node, and feed back the chip identification information of the target computing chip to the management and scheduling platform;

[0008] The computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and constructs a cluster subsystem based on the chip identification information of the target computing chip.

[0009] According to another aspect of the present invention, a system for reconfiguring architecture in a computing cluster is provided, wherein the system comprises: a computing cluster; the computing cluster comprises at least a cluster management node and a computing chip;

[0010] The cluster management node is used to obtain the cluster splitting instruction issued by the management and scheduling platform;

[0011] The cluster management node is used to determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip, and feed back the chip identification information of the target computing chip to the management and scheduling platform;

[0012] The computing chip is used to respond to the networking instruction issued by the management and scheduling platform based on the chip identification information and to target and construct a cluster subsystem based on the chip identification information of the target computing chip.

[0013] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement a method for reconstructing architecture in a computing cluster as described in any embodiment of the present invention when executed.

[0014] The technical solution of the embodiment of the present invention obtains the cluster splitting instruction issued by the management and scheduling platform through the cluster management node, determines the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip through the cluster management node, feeds back the chip identification information of the target computing chip to the management and scheduling platform, and the computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and builds a cluster subsystem based on the chip identification information of the target computing chip, so as to split the computing cluster into multiple cluster subsystems and quickly reconstruct the computing system so that the computing cluster can calculate multiple tasks in parallel, meet the needs of different computing power and business scenarios, flexibly adapt, and improve computing efficiency.

[0015] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 is a flowchart of a method for architecture reconstruction in a computing cluster provided according to Embodiment 1 of the present invention;

[0018] Figure 2 This is an example diagram of a route change provided according to Embodiment 1 of the present invention;

[0019] Figure 3 is a flowchart of another method for architecture reconstruction in a computing cluster provided according to Embodiment 2 of the present invention;

[0020] Figure 4 is a structural diagram of an architecture reconstruction system in a computing cluster provided according to Embodiment 3 of the present invention;

[0021] Figure 5 is a timing diagram of a computing cluster construction method provided according to Embodiment 4 of the present invention;

[0022] Figure 6 is a schematic diagram of a computing cluster provided according to Embodiment 4 of the present invention;

[0023] Figure 7 is a schematic diagram of the structure of a computing chip provided according to a fourth embodiment of the present invention;

[0024] Figure 8 It is a timing diagram of an architecture reconstruction method in a computing cluster provided according to Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0027] Embodiment 1

[0028] Figure 1This is a flowchart of a method for reconstructing the architecture in a computing cluster provided according to the first embodiment of the present invention. This embodiment is applicable to the case of reconstructing the architecture of a computing cluster. The method can be executed by a method for reconstructing the architecture in a computing cluster. The method for reconstructing the architecture in a computing cluster can be implemented in the form of hardware and / or software. The method for reconstructing the architecture in a computing cluster can be configured in a computing cluster, wherein the computing cluster at least includes a cluster management node and a computing chip. Figure 1 As shown, the method includes:

[0029] S110. Obtaining a cluster splitting instruction issued by a management and scheduling platform through a cluster management node.

[0030] Among them, the cluster management node can be understood as a node for coordinating the operation of the computing cluster, which is a key component of the core control and management function in the computing cluster. In practical applications, the computing cluster may include a cluster management node and a computing chip. For the computing cluster in the present application, it can be a cube computing cluster composed of a cluster management node and a plurality of computing chips, which is used for three-dimensional simulation calculations. In one embodiment, the computing chips in the computing cluster can be configured in multiples of 64 according to demand. For example, the computing cluster can be composed of 512 computing chips to form an 8×8×8 cube computing cluster, and the cluster management node is connected to the computing chip. The management and scheduling platform can be a platform for centralized management and coordination of resources, tasks and workflows, and can be used to schedule computing clusters to perform computing tasks. In the actual operation process, the management and scheduling platform can run in a computer. In one embodiment, the computing cluster can be a massively parallel processing cluster (Massive Parallel Processing, MPP). The cluster splitting instruction can be understood as the indication information of the splitting cluster. When the computing cluster receives the splitting instruction, the computing cluster can be split according to the splitting instruction. Generally speaking, the cluster splitting instruction may include the target number of target computing chips and the indication information of detecting the status of the computing chip.

[0031] In an embodiment, the management and scheduling platform can determine the computing resources required by the computing task, determine the target number of computing chips according to the computing resources, and generate cluster splitting instructions. The cluster management node can receive the cluster splitting instructions issued by the management and scheduling platform to split the cluster subsystem.

[0032] S120 . Determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip through the cluster management node, and feed back the chip identification information of the target computing chip to the management and scheduling platform.

[0033] Among them, the target computing chip can be understood as a computing chip used to build a cluster subsystem. Generally speaking, the target computing chip is an adjacent computing chip to reduce the distance of data transmission between computing chips. In the actual operation process, the computing chip that is closest and in a non-operating state can be used as the target computing chip. The chip identification information can be understood as information used to identify and distinguish each target computing chip. In the actual operation process, the chip identification information can be determined by the cluster management node according to the location information of the computing chip. Since the computing cluster is a cube cluster, a coordinate system can be established in the computing cluster, and chip identification information is established for each computing chip according to the position of each computing chip in the coordinate system. The chip identification information may include but is not limited to a unique chip identification (Identity document, ID). Exemplarily, taking the chip identification information consisting of three digits as an example, the first digit represents the position of the computing chip in the X-axis direction, the second digit represents the position of the computing chip in the Y-axis direction, and the second digit represents the position of the computing chip in the Z-axis direction. The chip identification information can be 007, 103, etc.

[0034] In an embodiment, the target number contained in the cluster splitting instruction can be extracted by the cluster management node, and the chip identification information of the computing chip can be determined. The adjacent target number of computing chips is determined as the target computing chip, and the chip identification information of the target computing chip is fed back to the management and scheduling platform. In the actual operation process, the location of the computing chip can be determined according to the chip identification information of the computing chip, and the adjacent computing chip can be determined as the target computing chip according to the location. When there are multiple groups of adjacent target number computing chips, the computing chip with a closer distance between the computing chips in each group can be determined as the target computing chip according to the chip identification information. For example, when the target number is 64, the target computing chip is determined according to the chip identification information of the computing chip, including a computing chip group with a 4×4×4 cube position relationship and a computing chip group with an 8×2×4 cube position relationship. It can be determined that the distance between the computing chips in the 4×4×4 cube position relationship is closer, and the computing chip in the 4×4×4 cube position relationship is used as the target computing chip. Alternatively, the preset position relationship can be determined according to the target number, and the computing chip searched according to the preset position relationship is used as the target computing chip. Exemplarily, when the number of targets is 8, the preset position relationship may include a 2×2×2 cube position relationship; when the number of targets is 16, the preset position relationship may include a 2×2×4 cube position relationship; when the number of targets is 32, the preset position relationship may include a 2×4×4 cube position relationship; when the number of targets is 64, the preset position relationship may include a 4×4×4 cube position relationship; when the number of targets is 128, the preset position relationship may include a 4×4×8 cube position relationship; when the number of targets is 256, the preset position relationship may include a 4×8×8 cube position relationship.

[0035] In one embodiment, after the cluster management node obtains the cluster splitting instruction, it may also first obtain the performance parameters of the computing chip in response to the cluster splitting instruction, and when it is determined that the performance parameters of the computing chip are good, determine the adjacent target number of computing chips as the target computing chips.

[0036] S130. The computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and constructs a cluster subsystem based on the chip identification information of the target computing chip.

[0037] The networking instruction can be understood as an instruction for instructing to construct a cluster subsystem. The cluster subsystem can be understood as a sub-computing system divided in a computing cluster based on a target computing chip.

[0038] In an embodiment, in response to the networking instructions issued by the management and scheduling platform based on the chip identification information, the target link relationship of the target computing chip can be determined according to the chip identification information of the target computing chip, and the routing of each target computing chip can be changed according to the target link relationship, and the communication connection between the target communication chips can be established to form a cluster subsystem. In the actual operation process, each axial communication interface preset by the target computing chip can be determined, and each target computing chip can be routed according to the target link relationship according to each axial communication interface. Among them, each axial communication interface can include an X+ axial communication interface, an X- axial communication interface, a Y+ axial communication interface, a Y- axial communication interface, a Z+ axial communication interface, and a Z- axial communication interface.

[0039] In one embodiment, Figure 2 FIG. 1 is an example diagram of a route change provided according to the first embodiment of the present invention. Figure 2 As shown, take the X-axis as an example, for example, there are 8 chips in the X-axis of the computing cluster, and the chip identification information is 000 to 007 respectively. The X-axis high-speed input / output cables (IO) between the chips are connected back and forth to form a closed loop. When the computing cluster performs computing tasks, the communication route of the X-axis communication interface in the 000 chip points to the X+axis communication interface of the 007 chip, which is responsible for confirming the connection before data exchange and data processing. When a 4×4×4 cluster subsystem is cut out, the X-axis communication interface of the 000 chip can point the communication route to the X+axis communication interface of the 003 chip to establish communication, and the data is forwarded to the X+axis communication interface of the 003 chip through the 007, 006, 005, and 04 chips to complete the data transmission. By analogy, the Y and Z axes also complete the change of communication routes in this way, establish communication connections to build cluster subsystems.

[0040] In an embodiment of the present invention, a cluster splitting instruction issued by a management and scheduling platform is obtained through a cluster management node, a target computing chip is determined through the cluster management node based on the cluster splitting instruction and the chip identification information of the computing chip, the chip identification information of the target computing chip is fed back to the management and scheduling platform, and the computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and a cluster subsystem is constructed based on the chip identification information of the target computing chip, so as to split the computing cluster into multiple cluster subsystems, and quickly reconstruct the computing system, so that the computing cluster can calculate multiple tasks in parallel, meet the requirements of different computing power and business scenarios, flexibly adapt, and improve computing efficiency.

[0041] In one embodiment, before obtaining the cluster splitting instruction issued by the management and scheduling platform through the cluster management node, the method further includes:

[0042] The cluster management node responds to the cluster status detection instruction issued by the management and scheduling platform to obtain the performance parameters of the computing chip;

[0043] The performance parameters are sent to the management and scheduling platform through the cluster management node, so that the management and scheduling platform determines the computing chip for forming the target computing cluster as the first computing chip according to the performance parameters, allocates chip identification information of each first computing chip according to the position of the first computing chip contained in the performance parameters, and issues a cluster networking instruction;

[0044] Receive cluster networking instructions and chip identification information through the cluster management node;

[0045] Determine each preset axial communication interface through the computing chip, and determine the link relationship between the first computing chips based on the cluster networking instruction, the chip identification information and each axial communication interface;

[0046] The communication connection between the first computing chips is established through the computing chips according to the link relationship, a target computing cluster is constructed, and the target computing cluster is used as a new computing cluster.

[0047] The cluster status detection instruction can be understood as the indication information for detecting the cluster status. The performance parameter refers to the index parameter of the computing chip for executing the computing task. For example, the performance parameter can include the basic status of the chip, such as the operating voltage, temperature and online status, etc. The performance parameter can also include various performance indicators, such as the interface transmission rate, etc. The cluster networking instruction can be understood as the instruction for instructing to build the target computing cluster.

[0048] In an embodiment, the management and scheduling platform may issue a cluster status detection instruction to the computing cluster. After receiving the status detection instruction, the cluster management node obtains the performance parameters of the computing chip in response to the status detection instruction, and sends the performance parameters to the management and scheduling platform through the cluster management node. The management and scheduling platform determines the computing chip for forming the target computing cluster according to the performance parameters, and uses the computing chip for forming the target computing cluster as the first computing chip. The management and scheduling platform allocates the chip identification information of each first computing chip according to the position of the first computing chip contained in the performance parameters, and issues a cluster networking instruction. The cluster management node may receive the cluster networking instruction and the chip identification information. The computing chip determines the preset axial communication interfaces, determines the position of each first computing chip according to the chip identification information through the computing chip, and determines the link relationship between the first computing chips according to the position and each axial communication interface. The computing chip then establishes a communication connection between the first computing chips according to the link relationship, constructs the target computing cluster, and uses the target computing cluster as a new computing cluster.

[0049] In the actual operation process, the management and scheduling platform can determine the computing chips to be used to form the target computing cluster according to the operating voltage, temperature, online status and interface transmission rate, and use the computing chips that form the target computing cluster as the first computing chip. Since the computing cluster is a cube cluster, a coordinate system can be established in the computing cluster, and chip identification information can be established for each computing chip according to the position of each computing chip in the coordinate system. Exemplarily, taking the chip identification information as a three-digit number as an example, the first digit represents the position of the computing chip in the X-axis direction, the second digit represents the position of the computing chip in the Y-axis direction, and the second digit represents the position of the computing chip in the Z-axis direction. The position of the first computing chip can be assigned the chip identification information of each first computing chip. Exemplarily, the chip identification information can be 007, 103, etc. The computing chip can pre-configure each axial communication interface, determine the link relationship between the first computing chips through the previous position of the first computing chip, and establish a communication connection between the first computing chips according to the link relationship, build the target computing cluster, and use the target computing cluster as a new computing cluster.

[0050] In one embodiment, determining the link relationship between the first computing chips based on the cluster networking instruction, chip identification information, and each axial communication interface by the computing chip includes:

[0051] Determine the location of each first computing chip according to the chip identification information by using the computing chip;

[0052] The link relationship between the first computing chips is determined by the computing chips according to the positions of the first computing chips and the axial communication interfaces.

[0053] The location may be understood as the location of the first computing chip in the computing cluster.

[0054] In an embodiment, since the chip identification information of the first computing chip is allocated according to the position of the first computing chip, the computing chip can identify the position of the first computing chip according to the chip identification information. Adjacent first computing chips can be connected according to the position of the first computing chip and each axial communication interface, and the connected links can be used as the link relationship between the first computing chips.

[0055] In one embodiment, before obtaining the networking instruction issued by the management and scheduling platform based on the chip identification information through the computing chip and constructing the cluster subsystem based on the networking instruction and the chip identification information of the target computing chip, it also includes:

[0056] The subsystem identification information issued by the management and scheduling platform based on the chip identification information of the target computing chip is obtained through the cluster management node, and the subsystem identification information is used as the identification information of the cluster subsystem to be constructed.

[0057] The subsystem identification information may be understood as identification information of a cluster subsystem, and is used to identify and distinguish the cluster subsystem.

[0058] In an embodiment, when the networking instructions issued by the management and scheduling platform based on the chip identification information are obtained through the computing chip, before the cluster subsystem is constructed based on the networking instructions and the chip identification information of the target computing chip, the management and scheduling platform can first issue the subsystem identification information according to the chip identification information of the target computing chip, and send the subsystem identification information to the cluster management node. The cluster management node can use the subsystem identification information as the identification information of the cluster subsystem to be constructed.

[0059] In one embodiment, after the cluster subsystem is constructed, the management and scheduling platform can re-issue chip identification information to the computing chips according to the position relationship of the computing chips in the cluster subsystem to facilitate subsequent routing changes of the cluster subsystem.

[0060] In one embodiment, after obtaining the networking instruction issued by the management and scheduling platform based on the chip identification information through the computing chip, and building the cluster subsystem based on the networking instruction and the chip identification information of the target computing chip, it also includes:

[0061] Send the cluster subsystem construction completion information to the management and scheduling platform through the cluster management node;

[0062] The computing chip obtains the computing tasks issued by the management and scheduling platform based on the construction completion information, and executes the computing tasks through the cluster subsystem.

[0063] The construction completion information may be understood as information indicating that the cluster subsystem has been constructed. The construction completion information may include data information of the cluster subsystem, such as relevant information of computing chips in the cluster subsystem. The computing task may refer to a large computing task that is collaboratively processed by a cluster subsystem of multiple computing chips. For example, the computing task may include a three-dimensional simulation computing task.

[0064] In an embodiment, after the construction of the cluster subsystem is completed, the construction completion information of the cluster subsystem can be sent to the management and scheduling platform through the cluster management node. The management and scheduling platform publishes computing tasks based on the construction completion information, and executes the computing tasks through the cluster subsystem, so that the computing tasks can be independently executed by the cluster subsystem, so that the computing cluster can execute multiple computing tasks in parallel and improve computing efficiency.

[0065] Embodiment 2

[0066] Figure 3 FIG. 1 is a flowchart of another method for reconfiguring the architecture in a computing cluster provided according to Embodiment 2 of the present invention. This embodiment further illustrates a method for reconfiguring the architecture in a computing cluster based on the above embodiment. Figure 3 As shown, the method includes:

[0067] S210: Obtain, through the cluster management node, a cluster splitting instruction issued by the management and scheduling platform.

[0068] S220 . Extract the target quantity included in the cluster splitting instruction through the cluster management node, and determine the chip identification information of the computing chip.

[0069] In an embodiment, the cluster management node may parse the cluster split instruction, extract the target quantity contained in the cluster split instruction, and at the same time, extract the chip identification information of the computing chip that is pre-issued.

[0070] S230. Determine the location of each computing chip according to the chip identification information through the cluster management node, and determine a preset position relationship based on the target number.

[0071] The preset position relationship may be understood as a pre-set position relationship between target computing chips, and the preset position relationship may correspond to the target quantity.

[0072] In an embodiment, the cluster management node may determine the location corresponding to each computing chip according to the chip identification information, and query the corresponding preset position relationship according to the target number. Exemplarily, when the target number is 8, the preset position relationship may include a 2×2×2 cube position relationship; when the target number is 16, the preset position relationship may include a 2×2×4 cube position relationship; when the target number is 32, the preset position relationship may include a 2×4×4 cube position relationship; when the target number is 64, the preset position relationship may include a 4×4×4 cube position relationship; when the target number is 128, the preset position relationship may include a 4×4×8 cube position relationship; when the target number is 256, the preset position relationship may include a 4×8×8 cube position relationship.

[0073] S240. Determine, through the cluster management node, a computing chip corresponding to a preset position relationship as a target computing chip according to the location, and send the chip identification information of the target computing chip to the management and scheduling platform, so that the management and scheduling platform generates a networking instruction based on the chip identification information.

[0074] In an embodiment, after the cluster management node determines the preset position relationship, it can query the location that satisfies the preset position relationship and find the computing chip corresponding to the location as the target computing chip, and determine the chip identification information of the target computing chip and send it to the management and scheduling platform, so that the management and scheduling platform generates a networking instruction based on the chip identification information. In an embodiment, when there are multiple groups of target computing chips, a group of target computing chips with smaller chip identification information can be determined as the target computing chip.

[0075] S250. The computing chip responds to the networking instruction issued by the management and scheduling platform, and determines the target link relationship of the target computing chip according to the chip identification information of the target computing chip.

[0076] Among them, the target link relationship can be understood as the connection link of the target computing chip.

[0077] In an embodiment, when the computing chip receives the networking instruction issued by the management and scheduling platform, the location of the target computing chip can be determined according to the chip identification information of the target computing chip, and the adjacent target computing chips can be connected according to the location, and the connection links between the target computing chips are used as the target link relationship.

[0078] S260, using the computing chip to change the routing of each target computing chip according to the target link relationship based on the preset axial communication interfaces, and establish communication connections between the target communication chips to form a cluster subsystem.

[0079] Among them, each axial communication interface can be understood as a communication interface for each axis of the computing chip. By way of example, each axial communication interface can include an X+ axial communication interface, an X- axial communication interface, a Y+ axial communication interface, a Y- axial communication interface, a Z+ axial communication interface and a Z- axial communication interface.

[0080] In an embodiment, the computing chip can determine each preset axial communication interface, perform routing changes according to the target link relationship through each axial communication interface, connect each target communication chip, and establish a communication connection between the target communication chips to form a cluster subsystem. In the actual operation process, the link relationship in the computing cluster can be determined, the target link relationship and the different link relationships in the link can be determined, and only the different link relationships can be routed and changed, and the communication connection between the target communication chips can be established to form a cluster subsystem.

[0081] In an embodiment of the present invention, a cluster splitting instruction issued by a management and scheduling platform is obtained through a cluster management node. The cluster management node extracts the target number contained in the cluster splitting instruction and determines the chip identification information of the computing chip. The cluster management node determines the location of each computing chip according to the chip identification information, and determines the preset position relationship based on the target number. The cluster management node determines the computing chip corresponding to the preset position relationship as the target computing chip according to the location, and sends the chip identification information of the target computing chip to the management and scheduling platform. The computing chip responds to the networking instruction, determines the target link relationship of the target computing chip according to the chip identification information of the target computing chip, and changes the routing of each target computing chip according to the target link relationship based on the preset axial communication interfaces through the computing chip, establishes communication connections between the target communication chips to form a cluster subsystem, and realizes that a large computing system can be divided into several independent cluster subsystems without changing the hardware computing cluster architecture, so as to avoid a computing task occupying too many computing resources.

[0082] Embodiment 3

[0083] Figure 4 1 is a schematic diagram of a system for reconfiguring an architecture in a computing cluster according to Embodiment 3 of the present invention. Figure 4 As shown, the architecture reconstruction system in the computing cluster includes: a computing cluster 10; the computing cluster at least includes a cluster management node 11 and a computing chip 12.

[0084] The cluster management node 11 is used to obtain the cluster splitting instruction issued by the management and scheduling platform.

[0085] The cluster management node 11 is used to determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip, and feed back the chip identification information of the target computing chip to the management and scheduling platform.

[0086] The computing chip 12 is used to respond to the networking instruction issued by the management and scheduling platform based on the chip identification information, and to target and build a cluster subsystem based on the chip identification information of the target computing chip.

[0087] In an embodiment of the present invention, a cluster splitting instruction issued by a management and scheduling platform is obtained through a cluster management node. The cluster management node determines a target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip, and feeds back the chip identification information of the target computing chip to the management and scheduling platform. The computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and builds a cluster subsystem based on the chip identification information of the target computing chip, thereby splitting the computing cluster into multiple cluster subsystems and rapidly reconstructing the computing system so that the computing cluster can calculate multiple tasks in parallel, meet the requirements of different computing power and business scenarios, flexibly adapt, and improve computing efficiency.

[0088] Furthermore, in the above embodiment of the invention, the cluster management node 11 is also used for:

[0089] In response to the cluster status detection instruction issued by the management and scheduling platform, the performance parameters of the computing chip are obtained.

[0090] Furthermore, in the above embodiment of the invention, the cluster management node 11 is also used for:

[0091] The performance parameters are sent to the management and scheduling platform through the cluster management node, so that the management and scheduling platform determines the computing chip for forming the target computing cluster as the first computing chip according to the performance parameters, allocates the chip identification information of each first computing chip according to the position of the first computing chip contained in the performance parameters, and issues cluster networking instructions.

[0092] Furthermore, in the above-mentioned embodiment of the invention, the cluster management node 11 is used to:

[0093] Receive cluster networking instructions and chip identification information.

[0094] Furthermore, in the above-mentioned embodiment of the invention, the computing chip 12 is used for:

[0095] Determine each preset axial communication interface, and determine the link relationship between the first computing chips based on the cluster networking instruction, the chip identification information and each axial communication interface;

[0096] It is used to establish a communication connection between the first computing chips according to the link relationship, build a target computing cluster, and use the target computing cluster as a new computing cluster.

[0097] Furthermore, in the above-mentioned embodiment of the invention, the computing chip 12 is also used for:

[0098] Determine the location of each first computing chip according to the chip identification information;

[0099] The link relationship between the first computing chips is determined according to the locations of the first computing chips and the axial communication interfaces.

[0100] Furthermore, in the above-mentioned embodiment of the invention, the cluster management node 11 is used to:

[0101] Extract the target quantity contained in the cluster splitting instruction and determine the chip identification information of the computing chip;

[0102] Determine the location of each computing chip according to the chip identification information, and determine the preset position relationship based on the target quantity;

[0103] According to the current location, a computing chip corresponding to a preset position relationship is determined as a target computing chip, and chip identification information of the target computing chip is sent to the management and scheduling platform, so that the management and scheduling platform generates a networking instruction based on the chip identification information.

[0104] Furthermore, in the above-mentioned embodiment of the invention, the computing chip 12 is used for:

[0105] In response to the networking instruction issued by the management and scheduling platform, the target link relationship of the target computing chip is determined according to the chip identification information of the target computing chip;

[0106] Based on the preset axial communication interfaces, the routing of each target computing chip is changed according to the target link relationship, and the communication connection between the target communication chips is established to form a cluster subsystem.

[0107] Furthermore, in the above-mentioned embodiment of the invention, the cluster management node 11 is used to:

[0108] The subsystem identification information issued by the management and scheduling platform based on the chip identification information of the target computing chip is obtained, and the subsystem identification information is used as the identification information of the cluster subsystem to be constructed.

[0109] Furthermore, in the above-mentioned embodiment of the invention, the cluster management node 11 is used to:

[0110] Send the cluster subsystem construction completion information to the management and scheduling platform.

[0111] Furthermore, in the above-mentioned embodiment of the invention, the computing chip 12 is used for:

[0112] Obtain computing tasks issued by the management and scheduling platform based on the build completion information, and execute computing tasks through the cluster subsystem.

[0113] In one embodiment, the computing chip 12 includes: a computing core and a management core;

[0114] The management kernel is used to determine each axial communication interface of the computing chip, respond to the networking instruction issued by the management and scheduling platform based on the chip identification information, and build a cluster subsystem based on the chip identification information of the target computing chip;

[0115] The computing kernel is used to execute computing tasks issued by the management and scheduling platform.

[0116] The system for reconfiguring the architecture in a computing cluster provided by an embodiment of the present invention can execute the method for reconfiguring the architecture in a computing cluster provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0117] Embodiment 4

[0118] Figure 5 1 is a timing diagram of a computing cluster construction method provided according to Embodiment 4 of the present invention. This embodiment is based on the above embodiment and takes a computing chip including a computing core and a management core as an example to further illustrate a method for reconstructing the architecture in a computing cluster. Figure 5 As shown, the method includes:

[0119] The computing chips in the computing cluster are connected by 6 axial high-speed IO cables as axial communication interfaces. The management core in the chip is responsible for the communication link routing management and defines the polarity of each axial communication interface. Before performing computing tasks, the management and scheduling platform of the computing cluster sends the cluster status detection instruction to the cluster management node, and the cluster management node forwards the cluster status detection instruction to the management core of the computing chip to check the performance parameters of the computing chips in the computing cluster, where the performance parameters include the basic status and each single performance indicator (such as interface transmission rate, etc.), and transmit the performance parameters to the management and scheduling platform. After the management and scheduling platform checks that the computing chip passes, it assigns chip identification information (chip ID) to each computing chip, and sends networking instructions to the management core of each computing chip. The management core defines the polarity of each axial communication interface of the chip, performs routing changes and establishes communication with the specified link, and then the management and scheduling platform inspects the computing cluster to obtain the communication status. After the inspection passes, a cubic computing cluster connected by high-speed cables in the positive and negative XYZ axes is formed. The cluster management node sends the cluster subsystem construction completion information to the management and scheduling platform, and the management and scheduling platform issues computing tasks based on the construction completion information, and the computing tasks are executed by the cluster subsystem. Figure 6 Schematic diagram of a computing cluster provided according to Embodiment 4 of the present invention. Figure 6 As shown, it is an 8×8×8 512-cube computing cluster.

[0120] In one embodiment, Figure 7 Schematic diagram of the structure of a computing chip provided according to the fourth embodiment of the present invention. Figure 7 As shown in the figure, the computing chip has 48 cores, including 36 computing cores and 12 management cores. These cores are divided into 6 groups, each of which consists of 2 management cores and 6 computing cores. The management core is responsible for task scheduling, routing management, data location, search and binding scheduling, etc.; the computing core is responsible for data processing and transmission. The coordination mechanism between the management core and the computing core ensures the rapid mobilization and processing of data, and maintains the characteristics of high efficiency and low latency in data processing. The chip uses a multi-level packet interconnection network, which is interconnected through internal routing. High-speed IO and internal routing are responsible for transmitting high-speed real-time data, and high-speed cache is responsible for storing internal cache data.

[0121] When performing whole-machine computing tasks, the management and scheduling platform can split the computing tasks into multiple sub-computing tasks and distribute them to all chip management cores. After receiving the task, the management core schedules the managed computing core to start computing. At the same time, it defines the polarity of the managed high-speed IO interface and communicates with the designated chip interface. The data generated and processed by the computing core in each axis is transmitted through the high-speed IO interface to realize the short-range force calculation of 512 neighborhoods.

[0122] In one embodiment, Figure 8 1 is a timing diagram of a method for reconfiguring architecture in a computing cluster according to Embodiment 4 of the present invention. Figure 8 As shown, the method includes:

[0123] When multiple computing tasks or small computing tasks need to be performed simultaneously, the computing cluster can be divided into multiple cluster subsystems, which can be divided into three-dimensional cube sub-computing systems with the number of chips such as 8 (2×2×2), 16 (2×2×4), 32 (2×4×4), 64 (4×4×4), 128 (4×4×8), and 256 (4×8×8). One 512 cube can be divided into 2 256 or 4 128 or 8 64 and other computing combinations, so as to quickly reconstruct the computing system and achieve rapid elastic scaling. When it is determined to split a cluster subsystem, such as splitting a 64-cube cluster subsystem, the management and scheduling platform will send a cluster splitting instruction to the computing cluster to confirm the adjacent computing chips that can form a 4×4×4 cube, and allocate a cluster subsystem identification information (cluster subsystem ID number) based on the chip ID allocated by the cluster. Each chip management core will change the route according to the allocated cluster subsystem ID number. The routing links of adjacent chips remain unchanged. The communication routes of the head and tail chips are forwarded through non-sub-computing system chips to complete the interconnection. Non-sub-computing system chips do not participate in computing tasks and are responsible for completing data forwarding. In the actual operation process, the management and scheduling platform can send cluster subsystem networking instructions and health check instructions to confirm the system status of the cluster subsystem.

[0124] In one embodiment, if Figure 2 In the example diagram of a route change shown, there are 8 chips in the X-axis of the computing cluster, and the chip identification information is 000 to 007 respectively. The high-speed input / output cables (IO) between the chips are connected back and forth to form a closed loop. When the computing cluster performs computing tasks, the management core of the communication route of the X-axis communication interface in the 000 chip points to the management core of the X+axis communication interface of the 007 chip. The management core responsible for the X+axis communication interface of the 007 chip confirms the connection before performing data exchange and data processing. When a 4×4×4 cluster subsystem is cut out, the management core of the X-axis communication interface of the 000 chip will point the communication route to the management core of the X+axis communication interface of the 003 chip to establish communication, and the data is forwarded to the management core of the X+axis communication interface of the 003 chip through the 007, 006, 005, and 04 chips to complete the data transmission. By analogy, the Y and Z axes also complete the change of communication routes in this way, establish communication connections to build cluster subsystems.

[0125] The embodiment of the present invention is a method for rapidly reconfiguring the system of an MPP architecture simulation computing cluster, which is mainly used in the field of three-dimensional simulation computing. By changing the routing of the communication link of a high-performance simulation computing chip, flexible computing system reconstruction can be achieved. A large computing cluster can be divided into multiple sub-computing systems, and the computing system can be rapidly reconfigured to achieve rapid elastic scaling. At the same time, multiple tasks can be calculated in parallel to meet the needs of different computing power and business scenarios, and flexible adaptation can be achieved to improve computing efficiency. Flexible computing system reconstruction can be achieved by calculating the routing of the chip communication link without changing the physical architecture of the cluster. The architecture is flexible, efficient, and easy to maintain.

[0126] Embodiment 5

[0127] Embodiment 5 of the present invention further provides a storage medium including computer executable instructions, wherein the computer executable instructions are used to execute a method for reconfiguring architecture in a computing cluster when executed by a computer processor, and the method is applied to a computing cluster, wherein the computing cluster includes at least a cluster management node and a computing chip, and the method includes:

[0128] Obtain the cluster splitting instruction issued by the management and scheduling platform through the cluster management node;

[0129] Determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip through the cluster management node, and feed back the chip identification information of the target computing chip to the management and scheduling platform;

[0130] The computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and constructs a cluster subsystem based on the chip identification information of the target computing chip.

[0131] Of course, the computer executable instructions of a storage medium containing computer executable instructions provided in an embodiment of the present invention are not limited to the operations of the method described above, and can also execute related operations in the architecture reconstruction method in the computing cluster provided in any embodiment of the present invention.

[0132] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0133] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0134] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0135] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0136] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for reconfiguring architecture in a computing cluster, characterized in that: Applied to a computing cluster, the computing cluster includes at least a cluster management node and a computing chip, the method includes: Obtain the cluster splitting instruction issued by the management and scheduling platform through the cluster management node; Determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip through the cluster management node, and feed back the chip identification information of the target computing chip to the management and scheduling platform; The computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information, and constructs a cluster subsystem based on the chip identification information of the target computing chip.

2. The method according to claim 1, characterized in that: Before obtaining the cluster splitting instruction issued by the management and scheduling platform through the cluster management node, it also includes: Acquiring performance parameters of computing chips by the cluster management node in response to the cluster status detection instruction issued by the management and scheduling platform; The performance parameter is sent to the management and scheduling platform through the cluster management node, so that the management and scheduling platform determines the computing chip for forming the target computing cluster as the first computing chip according to the performance parameter, allocates chip identification information of each first computing chip according to the position of the first computing chip contained in the performance parameter, and issues a cluster networking instruction; Receiving the cluster networking instruction and the chip identification information through a cluster management node; Determine each preset axial communication interface through the computing chip, and determine the link relationship between the first computing chips based on the cluster networking instruction, the chip identification information and the each axial communication interface; The communication connection between the first computing chips is established by the computing chips according to the link relationship, a target computing cluster is constructed, and the target computing cluster is used as a new computing cluster.

3. The method according to claim 2, characterized in that Determining the link relationship between the first computing chips based on the cluster networking instruction, the chip identification information, and the axial communication interfaces through the computing chip includes: Determining the location of each of the first computing chips according to the chip identification information by using the computing chip; The link relationship between the first computing chips is determined by the computing chips according to the positions of the first computing chips and the axial communication interfaces.

4. The method according to claim 1, characterized in that: The determining the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip by the cluster management node, and feeding back the chip identification information of the target computing chip to the management and scheduling platform, includes: Extracting the target quantity contained in the cluster splitting instruction through the cluster management node, and determining the chip identification information of the computing chip; Determine the location of each computing chip according to the chip identification information through the cluster management node, and determine the preset position relationship based on the target number; The cluster management node determines the computing chip corresponding to the preset position relationship as the target computing chip according to the location, and sends the chip identification information of the target computing chip to the management and scheduling platform, so that the management and scheduling platform generates a networking instruction based on the chip identification information.

5. The method according to claim 1, characterized in that: The method of constructing a cluster subsystem based on the chip identification information of the target computing chip in response to the networking instruction issued by the management and scheduling platform based on the chip identification information by the computing chip includes: The computing chip responds to the networking instruction issued by the management and scheduling platform, and determines the target link relationship of the target computing chip according to the chip identification information of the target computing chip; The computing chip changes the routing of each target computing chip according to the target link relationship based on the preset axial communication interfaces, and establishes communication connections between the target communication chips to form a cluster subsystem.

6. The method according to claim 1, characterized in that Before the computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information and constructs the cluster subsystem based on the chip identification information of the target computing chip, the method further includes: The subsystem identification information issued by the management and scheduling platform based on the chip identification information of the target computing chip is obtained through the cluster management node, and the subsystem identification information is used as the identification information of the cluster subsystem to be constructed.

7. The method according to claim 1, characterized in that After the computing chip responds to the networking instruction issued by the management and scheduling platform based on the chip identification information and constructs the cluster subsystem based on the chip identification information of the target computing chip, the method further includes: Sending the construction completion information of the cluster subsystem to the management and scheduling platform through the cluster management node; The computing task issued by the management and scheduling platform based on the construction completion information is obtained through the computing chip, and the computing task is executed through the cluster subsystem.

8. A system for reconfiguring architecture in a computing cluster, characterized in that: The system includes: a computing cluster; the computing cluster includes at least a cluster management node and a computing chip; The cluster management node is used to obtain the cluster splitting instruction issued by the management and scheduling platform; The cluster management node is used to determine the target computing chip based on the cluster splitting instruction and the chip identification information of the computing chip, and feed back the chip identification information of the target computing chip to the management and scheduling platform; The computing chip is used to respond to the networking instruction issued by the management and scheduling platform based on the chip identification information and to target and construct a cluster subsystem based on the chip identification information of the target computing chip.

9. The system according to claim 8, characterized in that The computing chip includes: a computing core and a management core; The management core is used to determine each axial communication interface of the computing chip, respond to the networking instruction issued by the management and scheduling platform based on the chip identification information, and target-build a cluster subsystem based on the chip identification information of the target computing chip; The computing kernel is used to execute the computing tasks issued by the management and scheduling platform.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the method for reconstructing the architecture in a computing cluster according to any one of claims 1 to 7 when executed.

Images