WO2013086861A1

WO2013086861A1 - Method for accessing multi-path input/output (i/o) equipment, i/o multi-path manager and system

Info

Publication number: WO2013086861A1
Application number: PCT/CN2012/079307
Authority: WO
Inventors: 雕峻峰; 刘云海
Original assignee: 华为技术有限公司
Priority date: 2011-12-13
Filing date: 2012-07-28
Publication date: 2013-06-20
Also published as: CN102497432B; CN102497432A

Abstract

A method for accessing multi-path input/output (I/O) equipment, an I/O multi-path manager and system relate to the field of information technologies (IT). With the invention, Peripheral component interconnect-express (PCI-E) links between all computing nodes and a PCI-E switch can be effective, and slave nodes also can access the I/O equipment through own PCI-E links, so that the time delay and the bottleneck under the condition of high service traffic are eliminated, and the utilization rate of the bandwidth is increased. The method comprises the following steps of: configuring, according to the received configuration information of the first hard zone, the PCI-E switch to make computing nodes of the first hard zone only access the I/O equipment of the first hard zone; and establishing,according to the received configuration information of the first hard zone, a mapping relationship between the computing nodes of the first hard zone and the I/O equipment of the first hard zone to make an operating system indicate computing nodes executing I/O tasks to access the I/O equipment of the first hard zone according to the mapping relationship. The embodiment of the invention is used for the multi-path accessing of the I/O equipment.

Description

Method for multi-path access I/O device, I/O multi-path manager and system

The present invention relates to the field of communications, and in particular, to a method for multi-path access to an I/O device, an I/O multipath manager, and a system. Background technique

Currently, computing nodes in network communications, such as servers, directly access I/O (Input/Output) devices through PCI-E (Peripheral Component Interconnect-Express). Most of them are based on the cluster scenario, that is, each interface (port) of the default PCI-E switch is connected to different systems, and is not well considered based on NUMA (Non-Uniform Memory Access). System scenario. In a NUMA system, all compute nodes are divided into sets of nodes, and each set of nodes is electrically isolated from other sets of nodes. Such a set is called a hard partition (or a large node), and each hard partition It includes one or more computing nodes, each of which has an RC (Root Complex) that can be connected to PCI-E.

A NUMA system is composed of multiple computing nodes aggregated through a NUMA network. The hardware resources of these physical nodes are managed by an OS (Operating System) or a hypervisor. Take the NUMA system as an example. As shown in Figure 1, a network diagram showing only one large node (or hard partition), including NUMA office and network 1, hard partition 2 including 3 compute nodes 21~23, PCI- E switch 3, external I/O device 4, each compute node has at least 1 CPU (Central Processing Unit), an NC (Node Controller), located at ΙΟΗ (Input/Output Hub, The root complex (RC) in the input and output hub), the external I/O device 6 includes a Fibre Channel network card 31 and an Ethernet card 32. It is assumed that the computing node 21 is the master node, and assumes the tasks of primary partition initiation and resource management, and the computing nodes 22-24 are slave nodes. The three compute nodes form a 6-way system through a NUMA aggregation network (actually, this PCI-E switch is also connected to other hard-partitioned compute nodes in the NUMA system). In the system shown in FIG. 1, the compute node 21 is externally connected to the PCI-E link of the PCI-E switch. 01 is a solid line display, and the PCI-E links 02 and 03 of the computing nodes 22 to 23 connected to the PCI-E switch are dotted lines. This is because software (such as firmware, OS, I/O applications, etc.) can only see 10 devices such as Ethernet cards through the link of the master node connected to the PCI-E switch, and the software cannot be accessed through the slave node. I/O device. As shown in Figure 2, it is a schematic diagram of the resource information table of the system shown in Figure 1. The compute nodes 22~23 cannot directly see I/O devices such as Ethernet cards and Fibre Channel network cards. To access I/O devices such as Ethernet cards, To pass through the NUMA aggregation network through the NUMA link, and then access through the link of the compute node 21. This increases the delay. In the case of a large amount of traffic, the primary node itself may have a bottleneck of input and output, and the PCI-E links that are connected to other slave nodes have no effect, and the bandwidth utilization is low. Summary of the invention

Embodiments of the present invention provide a method for multipath access I/O device, an I/O multipath manager, and a system, which can effectively enable a PCI-E link between all computing nodes and a PCI-E switch, eliminating the Bottlenecks in case of latency and high traffic increase bandwidth utilization. The embodiment of the present invention adopts the following technical solutions: In one aspect, a method for accessing an I/O device by using a multi-path includes: configuring a PCI-E switch according to the received configuration information of the first hard partition, to isolate the a hard partition other than the first hard partition, the computing node of the first hard partition accessing only the I/O device of the first hard partition; establishing the location according to the received configuration information of the first hard partition a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition, so that the operating system instructs the computing node that performs the I/O task to access the first hard according to the mapping relationship Partitioned I/O devices. In one aspect, an I/O multipath manager includes: a PCI-E switch configuration module, configured to configure a peripheral component to quickly interconnect a standard PCI-E switch according to the received configuration information of the first hard partition, to isolate And other hard partitions other than the first hard partition, so that the computing node of the first hard partition accesses only the I/O device of the first hard partition; An I/O multipath configuration module, configured to establish, according to the received configuration information of the first hard partition, a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition And causing the operating system to instruct the computing node performing the I/O task to access the I/O device of the first hard partition according to the mapping relationship. In another aspect, a system for providing multi-path access to an I/O device includes:

An I/O multipath manager, configured to configure a PCI-E switch according to the received configuration information of the first hard partition, to isolate other hard partitions except the first hard partition, so that the first hard partition The computing node only accesses the I/O device of the first hard partition, and establishes the computing node of the first hard partition and the I/ of the first hard partition according to the received configuration information of the first hard partition. a mapping relationship between the devices, so that the operating system instructs the computing node that performs the I/O task to access the I/O device of the first hard partition according to the mapping relationship; the I/O multipath manager is located in the firmware Or an operating system; an aggregation network, configured to connect computing nodes in the system, so that the system controls the computing node through an operating system; at least two hard partitions, wherein each hard partition includes at least one computing node;

a PCI-E switch, configured to establish a connection between the computing node and the I/O device, so that the computing node accesses the computing node by using a PCI-E link established between itself and the PCI-E switch I/O device belonging to a hard partition;

An I/O device, configured to connect between the computing node and an external network; a storage device, configured to store firmware, an operating system, an I/O application. The method for multi-path access I/O device, the I/O multi-path manager and the system provided by the embodiments of the present invention enable the PCI-E link between all the computing nodes and the PCI-E switch to be valid, so that the slave node It can also access I/O devices through its own PCI-E link, eliminating bottlenecks in case of delay and high traffic, and improving bandwidth utilization. BRIEF DESCRIPTION OF THE DRAWINGS In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will The drawings used in the embodiments or the description of the prior art are described in a single manner. It is obvious that the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art do not pay Other drawings can also be obtained from these drawings on the premise of creative labor. 1 is a schematic diagram of a networking diagram of a prior art NUMA system; FIG. 2 is a schematic structural diagram of a system resource information table of a prior art NUMA system; FIG. 3 is a schematic flowchart of a method for multipath accessing an I/O device according to an embodiment of the present invention; FIG. 4 is a schematic flowchart of a method for accessing an I/O device for multi-path access according to an embodiment of the present invention; FIG. 5 is a schematic flowchart of a method for accessing an I/O device for multi-path access according to an embodiment of the present invention; FIG. 7 is a schematic diagram of the networking of the NUMA system of the multi-path access I/O device according to the embodiment of the present invention; FIG. 8 is a schematic diagram of the networking of the NUMA system of the multi-path access I/O device according to the embodiment of the present invention; FIG. 9 is a schematic structural diagram of a system resource information table of a NUMA system according to an embodiment of the present invention; FIG. 10 is a schematic diagram of a system resource information table of a NUMA system according to an embodiment of the present invention; FIG. 11 is a schematic structural diagram of a system resource information table of an SMP system according to an embodiment of the present invention. detailed description

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention. Rather than all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without departing from the inventive scope are the scope of the present invention. The method for the multi-path access I/O device provided by the embodiment of the present invention, as shown in FIG. 3, includes:

S101. Configure a PCI-E switch according to the received configuration information of the first hard partition to isolate other hard partitions except the first hard partition, so that the computing node of the first hard partition only accesses the I/ of the first hard partition. O equipment.

S102. Establish a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition according to the received configuration information of the first hard partition, so that the operating system instructs to perform I/O according to the mapping relationship. The compute node of the task accesses the I/O device. The method for multi-path accessing an I/O device provided by the embodiment of the present invention enables a PCI-E link between all computing nodes and a PCI-E switch to be valid by establishing a mapping relationship between the computing node and the I/O device. Therefore, the slave node can also access the I/O device through its own PCI-E link, thereby eliminating the bottleneck in the case of delay and high traffic, and improving the bandwidth utilization. Embodiment 2 A method for accessing an I/O device by using a multi-path according to an embodiment of the present invention, as shown in FIG. 4, includes:

S201. The system reset is completed, and the primary node of the first hard partition starts to run firmware, and uses firmware to initialize the processor, the memory, and the chipset in the primary pair node. S202. The master node invokes an I/O multipath manager, and receives configuration information for analyzing the first hard partition. Specifically, the configuration information generally comes from a system management module (not shown) that manages the entire NUMA system by running management software. After receiving the configuration information, the I/O multipath manager analyzes the number of RCs in the first hard partitioned computing node in the configuration information, analyzes the number of I/O devices of the first hard partition, and types of devices, and identifies PCI. -E A corresponding port of the switch with the first hard partitioned compute node and a corresponding port of the PCI-E switch with the first hard partitioned I/O device.

5203. The master node invokes the I/O multipath manager to configure the PCI-E switch to isolate other hard partitions except the first hard partition. Specifically, the corresponding port of the PCI-E switch and the first hard partitioned computing node and the corresponding port of the first hard partitioned I/O device in the PCI-E switch are configured as a virtual switch to isolate The other hard partitioned I/O devices and I/O accesses other than the first hard partition enable the first hard partitioned compute node to access only the first hard partitioned I/O device.

5204. The master node searches for an I/O device and boots the slave node to start initialization. Specifically, the master node scans the first node of the first hard partition according to the number of the first hard partitioned computing nodes and the number and type of the first hard partitioned I/O devices. A hard-partitioned I/O device bus that searches for valid I/O devices. After searching for I/O devices, each of the RCs in the master node and the searched I/O devices are assigned addresses and memory, and are booted after the scan is complete. Initialize from the node. S205. The master node invokes the I/O multipath manager to establish a mapping relationship between the compute node of the first hard partition and the I/O device of the first hard partition. Specifically, the master node invokes the I/O multipath manager to send the address of the first hard partition I/O device to the RC of each slave node through a pointer, and the pointer points to the address of the above I/O device, so that A mapping relationship is established between the compute node of the first hard partition and the I/O device. S206. The master node invokes an I/O multipath manager to form a system resource information table. Specifically, the master node invokes the I/O multipath manager to form a system resource information table and sends a pointer of the system resource information table to the operating system, where the system resource information table includes the first hard partition computing node and the I/O device. The mapping relationship between the processor information and the memory information and the PCI-E link available between the compute node and the PCI-E switch. S207. Receive an I/O task, and allocate hardware resources for the I/O task according to the system resource information table. Specifically, the operating system receives the I/O task into the I/O task queue, invokes the system resource information table through the pointer of the system resource information table, and then determines the processor that executes the current I/O task and allocates the memory according to the I/O. The type of task determines which I/O device to access.

S208. The operating system instructs the processor that performs the I/O task to access the I/O device through the shortest path according to the system resource information table. Specifically, the operating system selects one according to a mapping relationship between the first hard partitioned computing node and the I/O device in the system resource information table, and a PCI-E link available between the computing node and the PCI-E switch. Perform the shortest path of the current I/O task. In general, this path is the PCI-E link between the compute node and the PCI-E switch where the processor itself performs the current I/O task. When the primary node needs to exit the hard partition due to a failure or resource reallocation, etc., as shown in the figure

5, also includes:

S301. The system management module receives the master node to issue an exit request instruction. Normally, when the primary node of the first hard partition needs to exit the first hard partition due to a failure or resource reallocation, the primary node sends an exit request command to the system management software through the system management module.

5302. The system management module receives an exit response command from the master node, instructing the master node to exit the first hard partition.

5303. After receiving the exit request instruction, the system management module sends an instruction to upgrade to the new primary node to one of the slave nodes. After receiving the exit request instruction, the management module of the system selects one of the slave nodes according to the policy of the system, and sends an instruction to the new master node.

5304. The new master node receives hardware resource information and I/O tasks from the original master node. The hardware resource information includes processor information for performing an I/O task, information of an I/O device that needs to be accessed to perform an I/O task, memory information, and PCI-E link information that is required to perform an I/O task. 5305. The original primary node exits the first hard partition and waits for maintenance or reallocation.

5306. The new master node updates the system resource information table.

5307. The new primary node enables the I/O multipath manager to configure the PCI-E switch according to the updated system resource information table, and isolates other hard partitions other than the first hard partition, and the preparation method is the same as step 203 in FIG. It is exactly the same and will not be described again. The system in this embodiment may be a NUMA system or an SMP system, and the computing node may be a server. The method for multi-path accessing an I/O device provided by the embodiment of the present invention enables a PCI-E link between all computing nodes and a PCI-E switch to be valid by establishing a mapping relationship between the computing node and the I/O device. Therefore, the slave node can also access the I/O device through its own PCI-E link, thereby eliminating the bottleneck in the case of delay and high traffic, and improving the bandwidth utilization. Embodiment 3 The I/O multipath manager 10 provided by the embodiment of the present invention, as shown in FIG. 6, includes: a call function interface 101, where a master node in a compute node of a first hard partition is invoked through an operating system or firmware The I/O multipath manager. The hard partition resource analysis module 102 is configured to receive configuration information of the first hard partition, analyze the number and address of the RC in the first hard partition of the configuration information, and analyze the number of I/O devices in the first hard partition. And a type of the device, identifying a corresponding port of the PCI-E switch with the compute node of the first hard partition, and a corresponding port of the PCI-E switch with the I/O device of the first hard partition.

The PCI-E switch configuration module 103 is configured to configure a PCI-E switch according to the received configuration information of the first hard partition to isolate other hard partitions except the first hard partition, so that the first hard partition compute node only Access the I/O device of the first hard partition.

The I/O multipath configuration module 104 is configured to establish a mapping relationship between the first hard partitioned computing node and the first hard partitioned I/O device according to the received configuration information of the first hard partition, so that The operating system instructs the computing node performing the I/O task to access the I/O device of the first hard partition according to the mapping relationship. The I/O multipath manager provided by the embodiment of the present invention enables the PCI-E link between all the computing nodes and the PCI-E switch to be valid by establishing a mapping relationship between the computing node and the I/O device, thereby The slave node can also access the I/O device through its own PCI-E link, thereby eliminating the bottleneck in the case of delay and high traffic, and improving the bandwidth utilization. The fourth embodiment of the present invention provides a multi-path access I/O device. The NUMA system is taken as an example. As shown in FIG. 7, the method includes: an I/O multipath manager 10 as shown in FIG. The firmware 51 is configured to configure the PCI-E switch 3 according to the received configuration information of the first hard partition 2 to isolate other hard partitions except the first hard partition 2, so that the computing nodes 21-22 access only the first The I/O device 4 of the hard partition 2, and establishing a mapping relationship between the computing nodes 21~23 and the I/O device 4 according to the received configuration information of the first hard partition 2, so that the computing nodes 21~23 and the PCI- The PCI-E links 01~03 between the E switches 3 become valid. As shown in Figure 7, the PCI-E links 01~03 are all solid lines. Then, the I/O multipath manager 10 compares the mapping relationship between the compute nodes 21-23 and the I/O device 4, the processor information, the memory information, and the available PCI between the compute nodes 21-23 and the PCI-E switch 3. The E links 01~03 are associated to form a system resource information table. As shown in FIG. 9, the operating system 52 accesses the I/O device according to the shortest path according to the indication computer points 21~23 according to the system resource information table, which is generally the shortest. The path is the link between the compute node itself and the PCI-E switch that performs the I/O task.

NUMA aggregation network 1 is used to connect all compute nodes through NC aggregation and control all compute nodes through an operating system. At least one hard partition 2 (only the first hard partition 2 is drawn in FIG. 7, other hard partitions are not shown), including one master node 21 and two slave nodes 22, 23, of course, more slave nodes can be added. . Wherein, each computing node includes: a node controller NC, used for computing nodes and NUMA converged network connection; two CPUs for performing I/O tasks; one RC for I/O device scanning and compute node connection to PCI-E corresponding port. In this embodiment, the RC is located in an IOH (Input-Output Hub), and the RC can also be located in a CPU or a MUX (Multiplexer). The above computing node can be a server.

The PCI-E switch 3 is configured to establish a link between the computing nodes 21~23 of the first hard partition 2 and the I/O device 4 of the first hard partition, as shown in FIG. 7, the calculation of the first hard partition 2. The nodes 21~23 and the PCI-E switch link are both solid lines, and the compute nodes 21~23 can directly access the I/O device 4 of the first hard partition 2 through the own link 01-03. Of course, the PCI-E switch is also connected to other hard partitions, not shown in the figure.

The I/O device 4 includes a Fibre Channel (FC) network card 41 and an Ethernet card for connecting the computing nodes to the external network. The storage device 5 is for storing the firmware 51, the operating system 52, and the firmware 51 includes an I/O multipath manager 511. Another multi-path access I/O device system provided by this embodiment takes a NUMA system as an example. As shown in FIG. 8, the storage device 5 is configured to store firmware 51, an operating system 52, and an I/O multipath manager. 10 is located in the operating system 52, and the rest is exactly the same as the system shown in FIG. 4, and will not be described again. The system resource information table of the NUMA system is shown in FIG. 9. The system of the multi-path access I/O device provided by the embodiment of the present invention enables the PCI-E link between all the computing nodes and the PCI-E switch to be valid by establishing a mapping relationship between the computing node and the I/O device. Therefore, the slave node can also access the I/O device through its own PCI-E link, thereby eliminating the bottleneck in the case of delay and high traffic, and improving the bandwidth utilization. The fifth embodiment of the present invention provides a multi-path access I/O device. The SMP (Symmetric Multi-Process) system is used as an example. As shown in FIG. 10, the method includes: /O Multipath Manager 10, located in firmware 51, for receiving The configuration information of the second hard partition 2a is configured to configure the PCI-E switch 3 to isolate other hard partitions except the second hard partition, so that the computing node 2al~2a2 only accesses the I/O device 4 of the second hard partition 2a. And establishing a mapping relationship between the computing nodes 2al~2a3 and the I/O device 4 according to the received configuration information of the second hard partition 2a, so that the PCI-E chain between the nodes 2al~2a3 and the PCI-E switch 3 is calculated. The roads 01~03 are all valid. As shown in Figure 10, the PCI-E links 01~03 are all solid lines. Then, the I/O multipath manager 10 compares the mapping relationship between the compute nodes 2al~2a3 and the I/O device 4, the processor information, the memory information, and the available PCI between the compute nodes 2al~2a3 and the PCI-E switch 3. The E links 01~03 are associated to form a system resource information table. As shown in FIG. 11, the operating system 52 accesses the I/O device according to the shortest path according to the indication computer point 2al~2a3 according to the system resource information table, which is generally the shortest. The path is the link between the compute node itself and the PCI-E switch that performs the I/O task. Of course, the I/O multipath manager 10 can also be located in the operating system 52 (not shown in FIG. 10).

The SMP aggregation network 1 is used to directly interconnect the CPUs of all computing nodes, does not require an NC, and controls all computing nodes through an operating system. At least two hard partitions (only the second hard partition 2a is shown in FIG. 10, other hard partitions are not shown), including one master node 2al and two slave nodes 2a2, 2a3, and of course, more slave nodes can be added. . Each compute node includes: two CPUs for performing I/O tasks and direct interconnection between the nodes; and an RC for calculating the connection of the node to the corresponding port of the PCI-E. In this embodiment, the RC is located in the IOH, and the RC may also be located in the CPU or the MUX. The above computing node can be a server. The rest of the system provided by this embodiment is completely the same as the system shown in FIG. 7, and will not be described again. The system resource table of the system is shown in FIG. The system of the multi-path access I/O device provided by the embodiment of the present invention enables the PCI-E link between all the computing nodes and the PCI-E switch to be valid by establishing a mapping relationship between the computing node and the I/O device. Therefore, the slave node can also access the I/O device through its own PCI-E link, thereby eliminating the bottleneck in the case of delay and high traffic, and improving the bandwidth utilization. The above is only the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of changes or substitutions within the technical scope of the present invention. It should be covered by the scope of the present invention. Therefore, the scope of the invention should be determined by the scope of the appended claims.

Claims

Claim

A method for multi-path access to an input/output I/O device, comprising: configuring a peripheral component quick interconnect standard according to the received configuration information of the first hard partition

a PCI-E switch, to isolate other hard partitions except the first hard partition, so that the computing node of the first hard partition only accesses the I/O device of the first hard partition; according to the received The configuration information of the first hard partition establishes a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition, so that the operating system instructs to perform I/O according to the mapping relationship. The compute node of the task accesses the I/O device of the first hard partition.

2. The method according to claim 1, wherein the peripheral component is quickly interconnected with the standard PCI-E switch according to the received configuration information of the first hard partition to isolate the first hard partition. The external hard partition, before the computing node of the first hard partition accesses only the I/O device of the first hard partition, further includes: a master node in the first hard partition computing node, a processor, a memory, The chipset performs initialization; the master node receives configuration information of the first hard partition; and the master node boots the slave node for initialization.

The method according to claim 1 or 2, wherein the configuration information of the first hard partition includes: a quantity and an identifier of each computing node of the first hard partition; The number and type of the /O devices; the corresponding ports of the PCI-E switches and the respective computing nodes of the first hard partition; and the I/O devices of the PCI-E switch and the first hard partition Corresponding port.

The method according to claim 1 or 2, wherein the peripheral component is quickly interconnected with the standard PCI-E switch according to the received configuration information of the first hard partition to isolate the first hard partition An external hard partition, the computing node of the first hard partition accessing only the I/O device of the first hard partition includes: corresponding to each computing node of the first hard partition in the PCI-E switch a port and a corresponding port of the PCI-E switch and the I/O device of the first hard partition, configured as a virtual switch, for isolating other hard partitions except the first hard partition /O device and I/O access, such that the computing node of the first hard partition accesses only the I/O device of the first hard partition.

The method according to claim 1 or 2, wherein the establishing the first hard partition computing node and the first hard partition I according to the received configuration information of the first hard partition a mapping relationship between the /O devices, so that the operating system instructs the computing node that performs the I/O task to access the I/O device of the first hard partition according to the mapping relationship: the primary node is according to the first The number, the identifier, and the number and type of the I/O devices of the first hard partition, which are provided in the configuration information of the hard partition, are scanned one by one by the root component of the primary node. The first hard partitioned I/O device bus searches for a valid I/O device, and after searching for the I/O device, respectively allocates an address and a memory for the follower component and the searched I/O device in the master node; Transmitting the allocated address of the I/O device to the root component of the slave node, thereby establishing a mapping relationship between the compute node of the first hard partition and the I/O device of the first hard partition; System resource information table, including the first a mapping relationship between the partitioned compute node and the first hard partitioned I/O device, processor information and memory information, and available PCI between the compute node of the first hard partition and the PCI-E switch Information of the -E link; after receiving the I/O task, the operating system allocates hardware resources for the I/O task according to the system resource information table, where the hardware resources include performing the I/O task The processor, the I/O device and the memory information that are required to be accessed by the I/O task; the operating system indicates the execution according to the mapping relationship in the system resource information table. The processor of the I/O task selects one of the available PCI-E links to access the I/O device that needs to be accessed.

6. An I/O multipath manager, comprising:

a PCI-E switch configuration module, configured to configure a peripheral component to quickly interconnect a standard PCI-E switch according to the received configuration information of the first hard partition, to isolate other hard partitions except the first hard partition, so that The computing node of the first hard partition accesses only the I/O device of the first hard partition;

An I/O multipath configuration module, configured to establish, according to the received configuration information of the first hard partition, a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition And causing the operating system to instruct the computing node performing the I/O task to access the I/O device of the first hard partition according to the mapping relationship.

The I/O multipath management device according to claim 6, further comprising: a hard partition resource analysis module, configured to receive configuration information of the first hard partition, and analyze the configuration information. The number and identification of the middle root components of the first hard partitioned compute node, the number of I/O devices analyzing the first hard partition, and the type of the device, identifying the first hard partition in the PCI-E switch Corresponding port of the computing node and a corresponding port of the PCI-E switch and the I/O device of the first hard partition; calling a function interface, a primary node enabling node in the computing node of the first hard partition I/O multipath manager.

A system for multi-path accessing an I/O device, comprising: the I/O multipath manager according to any one of claims 6 to 7 for receiving the first The configuration information of the hard partition configures the PCI-E switch to isolate other hard partitions except the first hard partition, so that the computing node of the first hard partition only accesses the I/O device of the first hard partition Establishing a mapping relationship between the computing node of the first hard partition and the I/O device of the first hard partition according to the received configuration information of the first hard partition, so that the operating system according to the mapping The relationship indicates that the computing node performing the I/O task accesses the first Hard partitioned I/O device; the I/O multipath manager is located in firmware or an operating system; an aggregate network is used to connect computing nodes in the system, so that the system controls the computing node through an operating system; at least two a hard partition, wherein each hard partition includes at least one compute node;

An I/O device, configured to connect between the computing node and an external network; a storage device, configured to store firmware, an operating system, an I/O application.

The system according to claim 8, wherein the aggregation network comprises: a non-uniform memory access system aggregation network, and a symmetric multi-processing system aggregation network.

10. The system according to claim 8, wherein the computing node comprises: two central processors for the computing node to perform an I/O task; at least one root component, for the computing node Connection to PCI-E.

The system according to claim 10, wherein, when the aggregation network is a non-uniform memory access system aggregation network, the computing node further includes: a node controller, configured to be used by the computing node Connection to the non-uniform memory access system aggregation network and control of the compute nodes.