A computer network system and a computer system

Abstract

The present invention discloses a computer network system and a computer system, including a high-performance computing network, the high-performance computing network includes at least 4 computers, multiple computers in the high-performance computing network each solve a large area in the computing domain, the computing domain large area topological relationship corresponds to the high-performance computing network topological relationship, the high-performance computing network topological relationship is that multiple computers in the high-performance computing network are connected by wires to form a ring topology or a daisy chain topology; and a management network, the management network includes all computers in the high-performance computing network and computers for storing and post-processing computing results, all computers in the management network are connected by switches or routers to form a star network topology to solve the problem of poor performance due to the influence of bandwidth and delay, achieve the effect of approximately equal communication delays of adjacent computing nodes, and enhance the communication performance of adjacent computers to improve computing efficiency.

Description

Technical Field

[0001] The invention relates to a computer network system and a computer system. Background Art

[0002] Finite element method and finite volume method are commonly used numerical methods for solving continuous medium problems such as solid deformation and fluid motion. They belong to grid methods; that is, the computational domain is divided into small grid units. Correspondingly, the control equations acting on the computational domain are also discretized from partial differential equations to a set of algebraic equations. This set of algebraic equations is solved using a parallel algorithm, that is, the computational domain is partitioned, and each partition is assigned to a process to solve the algebraic equations of the corresponding computational domain. There is an interface between adjacent partitions, and the data on the interface is synchronized through inter-process communication. Different processes calculate the physical quantities of all partitions in parallel. Usually, each process is responsible for a central processing unit core, and the entire computational domain is divided into hundreds or thousands of parts, which are handed over to multiple multi-core computers in the same network for calculation. In turn, the solution speed of the entire computational domain depends on the computing performance of a single core and the communication capability between cores.

[0003] In the same computing domain, communication only occurs between adjacent partitions, that is, on the partition interface, so communication is local rather than global. Communication occurs between the cores of the computer's central processing unit. The cores of communication may be located in the same computer or distributed in different computers on the same network, communicating with the help of network cards. The communication performance between cores in the same computer is good, but the performance between cores communicating with the help of network cards is greatly affected by bandwidth and latency, and is poor. In addition, due to data consistency requirements, slower links in the network are performance shortcomings. Summary of the invention

[0004] In order to address the deficiencies in the prior art, the present application provides a computer network system and a computer system to solve the problem of poor performance due to the significant impact of bandwidth and delay, achieve the effect of approximately equal communication delays between adjacent computing nodes, and enhance the communication performance of adjacent computers to improve computing efficiency.

[0005] To achieve the above-mentioned object, the present invention provides a computer network system, including a high-performance computing network, wherein the high-performance computing network includes at least 4 computers, wherein the multiple computers in the high-performance computing network each solve a large area in a computing domain, and the topological relationship of the large area in the computing domain corresponds to the topological relationship of the high-performance computing network, and the topological relationship of the high-performance computing network is that the multiple computers in the high-performance computing network are connected by wires to form a ring topology or a daisy chain topology;

[0006] and a management network, the management network includes all computers in the high-performance computing network and computers for storage and post-processing of computing results, and all computers in the management network are connected by switches or routers to form a star network topology.

[0007] Furthermore, the multiple computers of the high-performance computing network are all provided with two single-port high-speed network cards, and the multiple computers of the high-performance computing network are individually placed in the corresponding layers of a single cabinet in sequence and a single cabinet wiring method is adopted to form a single ring network topology;

[0008] In the single cabinet wiring method, n computers are numbered from top to bottom or from bottom to top, namely Ai, where i is (1, 2, ..., n), and then the computers with the corresponding numbers are connected one by one through high-speed copper cables according to Ai and A(i+k)%n, where k is the integer closest to n / 2 and coprime to n. When k contains two integers closest to n / 2 and coprime to n, k takes an integer greater than n / 2 or an integer less than n / 2; % is the modulo operation. When (i+k)%n is equal to 0, Ai is connected to An.

[0009] Furthermore, the multiple computers of the high-performance computing network are all equipped with dual-port high-speed network cards, and the multiple computers of the high-performance computing network are placed in two adjacent cabinets A and B and form a single ring network topology in a double-cabinet wiring manner;

[0010] In the dual-cabinet wiring method, the computers in cabinets A and B are numbered Ai and Bi from top to bottom or from bottom to top, where i is (1, 2, ..., n). The computers with the corresponding numbers are then connected via high-speed copper cables as Ai to Ai+1, An to Bn, Bi to Bi+1, and A1 to B1. When i+1 is equal to n, An is connected to Bn.

[0011] Furthermore, the high-speed network card is a programmable network card, adopts the infiniband communication protocol, and uses data processing unit (DPU) hardware.

[0012] The present invention also provides a method for dividing a computing domain. In the first level of division, the computing domain is divided into N large regions, where N is the number of computers participating in the solution in the high-performance computing network. The number of units in each large region is approximately equal and each large region has one or two adjacent large regions. Computers belonging to large regions that are spatially adjacent in the computing domain are also adjacent in the topology of the high-performance computing network.

[0013] The second level of division divides N large regions into M small regions, where M is the number of threads involved in solving the problem on a single computer. The number of units in each small region is approximately equal and each small region has one or more adjacent small regions. The small regions in the same large region are bound to the same computer.

[0014] Furthermore, the specific method of the first layer division is as follows:

[0015] S1: Orthogonal normalization of the computational domain grid, that is, converting the grid cells into regular polygons with a side length of 1;

[0016] S2: Determine the partition path. On the orthogonal normalized grid, determine a straight line through the centroids of the two farthest planes as the partition path.

[0017] S3: Select N-1 equally divided points of the partition path, determine the normal plane, and divide the computational domain into N large regions, where N is the number of computers involved in the solution in the high-performance computing network, so that the number of units in each large region is approximately equal.

[0018] Furthermore, the specific method of the second layer division is as follows:

[0019] S1: Determine the number of cells M in each region, where M is less than or equal to the number of threads on a single computer involved in solving the problem. For a CPU using hyperthreading technology, M is equal to the number of physical cores of the CPU.

[0020] For a CPU that does not use or enable hyperthreading technology, M is equal to the number of physical cores of the CPU of a single computer minus the number of CPUs of a single computer;

[0021] S2: Use metis or scotch algorithm to divide the large area into M cells, the number of cells in each cell is approximately equal, and each cell has one or more adjacent cells.

[0022] The present invention also provides a computer system, including a computer network system to implement the steps of the computing domain partitioning method. Beneficial Effects

[0023] 1. The present invention uses a high-performance computing network with a ring network topology or a daisy chain topology to realize the inter-thread communication required for parallel computing of finite element method and finite volume method solvers, and uses a management network with a star topology for general communication needs such as network attached storage, network file system, remote login, remote desktop, etc.

[0024] 2. As the existing cores that communicate with the help of network cards are greatly affected by bandwidth and delay and have poor performance, adjacent computing domains are allocated to adjacent cores and the communication performance of adjacent cores is enhanced. The advantages of low latency of adjacent computers are brought into play through a ring topology or daisy chain topology network. Therefore, a wiring method is proposed to achieve the effect of approximately equal communication delays of adjacent computing nodes. Combined with the programmable characteristics of high-speed smart network cards, the load originally borne by the CPU is unloaded to the smart network card, further exerting the performance of the CPU.

[0025] 3. Compared with the existing technology, this application eliminates the need for high-speed switch equipment and reduces the possibility of network paralysis caused by single device failure, thereby reducing costs, saving computer room space, and facilitating the transformation of existing computer clusters.

[0026] 4. This application is compatible with existing FEM / FVM solvers to further improve computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a corresponding schematic diagram of the network system and the computing domain;

[0028] Figure 2 This is a wiring diagram for a single cabinet;

[0029] Figure 3 This is a schematic diagram of the wiring of the dual cabinets;

[0030] Figure 4 Schematic diagram of the computational domain partitioning method in curved pipe flow simulation.

[0031] Figure numerals: 1. computer; 2. switch; 3. high-speed copper cable; 4. shielded twisted pair; 5. computer for storage and post-processing of calculation results; 6. high-speed network card; 7. Ethernet interface; 8. cabinet; 9. large area; 10. small area; 11. computational domain grid; 12. orthogonal normalization; 13. partition path; 14. trisection point; 15. normal plane; 16. centroid. DETAILED DESCRIPTION

[0032] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0033] Refer to Figure 1-Figure 3 As shown, the present application provides a computer network system, including a high-performance computing network and a management network, and the high-performance computing network is used for inter-thread communication required for parallel computing of finite element method and finite volume method solvers. Multiple computers 1 in the high-performance computing network have corresponding large areas 9 in the computing domain, and the topological relationship of the large areas 9 corresponds to the topological relationship of the high-performance computing network. The topological relationship of the high-performance computing network is that multiple computers 1 in the high-performance computing network form a ring topology or a daisy chain topology through wired connections. The high-performance computer 1 network includes at least 4 computers 1, and the central processing unit and memory specifications of the computers 1 are the same. The computers 1 are all equipped with high-speed network cards 6 and the high-speed network cards 6 have the function of direct memory access. The port bandwidth of the high-speed network card 6 is greater than or equal to 10Gbps. The high-speed network card 6 is a programmable network card, adopts the infiniband communication protocol, and uses data processing unit (DPU) hardware. The data processing unit is used to realize enhanced functions such as transparent compression and sparse matrix compression, as well as simulation calculation result monitoring, real-time post-processing, file writing and other computing auxiliary functions.

[0034] The high-speed network cards 6 are connected by high-speed copper cables 3 or optical fibers to form a ring network topology or a daisy chain network topology. The present application preferably uses high-speed copper cables 3. The management network is used for general communication needs such as network attached storage, network file system, remote login and remote desktop, and specifically includes all computers 1 in the high-performance computing network and other computers 5 for storage and post-processing of calculation results. The Ethernet interface 7 bandwidth of the computer 1 is less than or equal to 10Gbps and is connected to the switch 2 or router through a shielded twisted pair 4 to form a star network topology. Through the high-performance computing network of the ring topology or daisy chain topology and the management network of the star topology, the communication overhead is further reduced, the communication performance is improved, and the calculation efficiency of FEM / FVM is thereby improved. Communication occurs between cores, and the cores of the communication may be located in the same computer 1 or may be distributed in different computers 1 of the same network. With the help of the network card for communication, the communication performance between the cores in the same computer 1 is better, and the performance between the cores communicating with the help of the network card is greatly affected by bandwidth and delay, and is poor. Therefore, the present application has come up with an effective idea of ​​allocating adjacent computing domain partitions to adjacent cores and enhancing the communication performance of adjacent cores, thereby improving the calculation efficiency. In order to achieve the same communication delay effect between adjacent computers 1, a wiring method of a single cabinet 8 and a double cabinet 8 is proposed. Either of the two is selected, and the wiring method of the double cabinet 8 is preferred.

[0035] Refer to Figure 3 As shown, the wiring method of the double cabinet 8, the wiring method of the double cabinet 8 numbers the computers 1 in A and B in the cabinet 8 from top to bottom or from bottom to top Ai and Bi, where i is (1, 2, ..., n), and then connects the corresponding numbered computers 1 through the high-speed copper cable 3 according to Ai and Ai+1, An and Bn, Bi and Bi+1, A1 and B1, where when i+1 is equal to n, An and Bn are connected; this application takes n as 16 as an example, and the equipment models are shown in Table 1 below. Each server is equipped with a high-speed network card 6, and the high-speed network card 6 is connected to the ocp interface of the server. Each high-speed network card 6 has two infiniband ports, and the rate of each port is 100Gbps. The high-speed network card 6 is installed on the ocp interface on the right side of the server in the A cabinet 8 and on the ocp interface on the left side of the server in the B cabinet 8 to reduce the length of the high-speed copper cable 3. This arrangement makes the high-speed copper cable 3 not exceed 0.5 m, standard finished wire can be used, and the delay is low. The first layer of servers in cabinet 8A (denoted as A1, and so on) are connected to A2, A2 is connected to A3, and so on, A8 is connected to B8, B8 is connected to B7, and so on, B2 is connected to B1, and B1 is connected to A1, forming a ring topology high-performance computing network. The server motherboard is equipped with two Gigabit Ethernet interfaces 7, and shielded twisted pair cables 4 are used to connect any of them to the switch 2. The twisted pair cables are arranged along the edge of the cabinet 8 to form a star topology management network.

[0036] Table 1 Dual cabinet equipment model list

[0037]

[0038] Refer to Figure 2 As shown, in the wiring method of a single cabinet 8, n computers 1 (device models are shown in Table 2) are numbered from top to bottom or from bottom to top, namely Ai, where i is (1, 2, ..., n), and then the computers 1 with corresponding numbers are connected one by one through high-speed copper cables 3 according to Ai and A(i+k)%n, where (i+k)%n is equal to 0, that is, Ai is connected to An; k is the integer closest to n / 2 and coprime with n, and % is a modulo operation; when k contains two integers closest to n / 2 and coprime with n, k takes an integer greater than n / 2 or an integer less than n / 2; the copper cables are arranged crosswise, and the intersections are fixed with plastic brackets and nylon tie straps. This application takes n as 16 as an example. The 16 servers are placed in the same cabinet 8. Each server is equipped with two high-speed network cards 6, which are installed on the PCIe interfaces on the left and right sides of the server, respectively. This can achieve a larger cable spacing and less crossing, reduce cable signal interference, and the PCIe interfaces on the left and right are respectively connected to different CPUs in the server. Each CPU is connected to a network card, so that the CPU load is more balanced. When there are 16 servers, k is taken as 7 or 9. The copper cable connection formed by substituting the values ​​of 7 and 9 is the same. Therefore, in the calculation, you can choose one according to the actual situation. This application takes k equal to 7 as an example. A1 is connected to A(1+7)%16, and A(1+7)%16 is A8. % is the modulus operation for finding the remainder. That is, A1 is connected to A8, and so on, until A9 is connected to A(9+7)%16. When A(9+7)%16 is equal to A0, A9 is connected to A16, and the connection forms a ring topology, such as Figure 2 shown.

[0039] Table 2 Single cabinet equipment model table

[0040]

[0041] Refer to Figure 1As shown, the present application also provides a method for dividing a computing domain, which performs two-layer division. The first layer of division divides the computing domain into N large regions 9, where N is the number of computers 1 participating in the solution in the high-performance computing network. The number of units in each large region 9 is approximately equal, and each large region 9 has one or two adjacent large regions 9, that is, one or two interfaces. The second-level partitioning divides the region 9 into M small areas 10. The number of small areas 10 is equal to the number of threads involved in the solution of a single computer 1. The number of units in each small area 10 is approximately equal. Each small area 10 has one or more adjacent small areas 10. The small areas 10 in the same region 9 are bound to the same computer 1, so that the topological relationship of the region 9 corresponds to the network topological relationship of the computer 1, that is, the computers 1 belonging to the spatially adjacent regions 9 in the computing domain are also adjacent in the high-performance computing network topology, that is, the partitioning of the region 9 can satisfy the rotation symmetry or mirror consistency with the high-performance computing network topology. For example, if the network topology is A1-A4-A3-A2-A1, the computing grid partition can be A4-A3-A2-A1 or A1-A2-A3-A4, which is a mirror situation. It can be A2-A1-A4-A3 or A1-A4-A3-A2 and other rotation symmetric situations.

[0042] The specific steps for the first level of division are as follows: 1) orthogonal normalization 12 of the computational domain grid 11, i.e., converting the grid cells into regular polygons with a side length of 1; 2) determining the partition path 13, and on the grid after orthogonal normalization 12, determining a straight line through the centroids 16 of the two farthest planes as the partition path 13; 3) selecting N-1 equally divided points of the partition path 13, determining the normal plane 15, and dividing the computational domain into N large regions 9, where N is the number of computers 1 involved in the solution, so that the number of cells in each large region 9 is approximately equal.

[0043] The specific steps for the second-level division are as follows: 1) Determine the number M of cells 10 in each region 9, where M is less than or equal to the number of threads of a single computer 1 involved in the solution. For a CPU using hyperthreading technology, M is equal to the number of physical cores of the CPU. For a CPU that does not use or has not enabled hyperthreading technology, M is equal to the number of physical cores of the CPU of a single computer 1 minus the number of CPUs of a single computer 1, that is, each CPU has one core that does not participate in the solution; 2) Use metis or scotch algorithms to divide the region 9 into M cells 10, where the number of units in each cell 10 is approximately equal, and each cell 10 has one or more adjacent cells 10. The present application also provides an embodiment for the method of dividing the computational domain, refer to Figure 4 Schematic diagram of the calculation domain division of the bent pipe.

[0044] The present application also provides a computer system, including a computer network system to implement the steps of the computational domain partitioning method.

[0045] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A computer network system, Features: A high-performance computing network is included, wherein the high-performance computing network includes at least 4 computers, wherein each of the multiple computers in the high-performance computing network solves a large area in a computing domain, and the topological relationship of the large area in the computing domain corresponds to the topological relationship of the high-performance computing network, and the topological relationship of the high-performance computing network is that the multiple computers in the high-performance computing network are connected by wires to form a ring topology or a daisy chain topology; and a management network, the management network including all computers in the high performance computing network and computers for storing and post-processing computing results, all computers in the management network are connected by switches or routers to form a star network topology; The computational domain is divided into N large regions by the computational domain partitioning method. N is the number of computers involved in solving the problem in the high-performance computing network. The number of units in each large region is approximately equal and each large region has one or two adjacent large regions. Computers belonging to large regions that are spatially adjacent in the computational domain are also adjacent in the high-performance computing network topology. The second level of division divides N large regions into M small regions, where M is the number of threads involved in solving the problem on a single computer. The number of units in each small region is approximately equal and each small region has one or more adjacent small regions. The small regions in the same large region are bound to the same computer.

2. The computer network system according to claim 1, Features: Multiple computers in the high-performance computing network are equipped with two single-port high-speed network cards. Multiple computers in the high-performance computing network are individually placed in the corresponding layers of a single cabinet in sequence and use a single cabinet wiring method to form a single ring network topology; In the single cabinet wiring method, n computers are numbered from top to bottom or from bottom to top, namely Ai, where i is 1, 2, ..., n, and then the computers with the corresponding numbers are connected one by one through high-speed copper cables according to Ai and A(i+k)%n, where k is the integer closest to n / 2 and coprime to n. When k contains two integers closest to n / 2 and coprime to n, k takes an integer greater than n / 2 or an integer less than n / 2; % is a modulo operation. When (i+k)%n is equal to 0, Ai is connected to An.

3. The computer network system according to claim 1, Features: Multiple computers in the high-performance computing network are all equipped with dual-port high-speed network cards. Multiple computers in the high-performance computing network are placed in two adjacent cabinets A and B and form a single ring network topology in a double-cabinet wiring mode. In the dual-cabinet wiring method, the computers in cabinets A and B are numbered Ai and Bi from top to bottom or from bottom to top, where i is 1, 2, ..., n. The computers with the corresponding numbers are then connected via high-speed copper cables as Ai to Ai+1, An to Bn, Bi to Bi+1, and A1 to B1. When i+1 is equal to n, An is connected to Bn.

4. The computer network system according to claim 2 or 3, Features: The high-speed network card is a programmable network card, adopts the infiniband communication protocol, and uses data processing unit (DPU) hardware.

5. The computer network system according to claim 4, Features: The specific method of the first layer division is as follows: S1: Orthogonal normalization of the computational domain grid, converting the grid cells into regular polygons with a side length of 1; S2: Determine the partition path. On the orthogonal normalized grid, determine a straight line through the centroids of the two farthest planes as the partition path. S3: Select N-1 equally divided points of the partition path, determine the normal plane, and divide the computational domain into N large regions, where N is the number of computers involved in the solution in the high-performance computing network, so that the number of units in each large region is approximately equal.

6. The computing domain network system according to claim 5, Features: The specific method of the second layer division is as follows: S1: Determine the number of cells M in each region, where M is less than or equal to the number of threads on a single computer involved in solving the problem. For a CPU using hyperthreading technology, M is equal to the number of physical cores of the CPU. For a CPU that does not use or enable hyperthreading technology, M is equal to the number of physical cores of the CPU of a single computer minus the number of CPUs of a single computer; S2: Use metis or scotch algorithm to divide the large area into M cells, the number of cells in each cell is approximately equal, and each cell has one or more adjacent cells.

7. A computer system for implementing the computer network system according to claim 6.

Images