An implementation method and device of a data center lossless network

Abstract

The application relates to an implementation method and device of a data center lossless network, and relates to a data communication technology network. The method comprises the following steps: acquiring the queue cache percentage of each switch node message in real time; when there is a switch node higher than a preset threshold value, the message is switched from the switch node higher than the preset threshold to an idle switch node, and a corresponding index field is constructed and issued to a corresponding TOR node device; the TOR node device updates the corresponding index field in the shunt index of the message header, and forwards the message according to the shunt index; when the queue cache percentage of all switch nodes is higher than the preset threshold value, ECN / PFC processing is performed. The application reasonably plans the path, guarantees the network throughput, and makes the data flow avoid congestion.

Description

Technical Field

[0001] This application relates to data communication technology networks, specifically to a method and apparatus for implementing a lossless data center network. Background Technology

[0002] In today's internet landscape, numerous services require rapid and frequent responses to user requests, especially with the rise of machine learning and AI technologies, which have dramatically increased the demand for computing power. To meet the needs of increasingly complex neural networks and deep learning models, data centers have designed numerous distributed computing clusters. However, the communication latency of a large number of parallel programs significantly impacts the efficiency of the entire computing process.

[0003] To improve processing efficiency, Remote Direct Memory Access (RDMA) technology has been widely adopted. However, applying RDMA to Ethernet presents several challenges. Ethernet is inherently a "best-effort" network, lacking a robust packet loss protection mechanism and making it extremely sensitive to packet loss. Test data from relevant literature indicates that a packet loss rate exceeding 0.1% will cause a sharp drop in effective network throughput. Therefore, to ensure that RDMA throughput is unaffected, lossless Ethernet support is required.

[0004] Currently, in lossless Ethernet networks, PFC (Priority-based Flow Control) and ECN (Explicit Congestion Notification) technologies are commonly used for congestion control. PFC is a Layer 2 technology, while ECN is a network and transport layer technology. PFC and ECN are usually enabled simultaneously to ensure zero packet loss. This is typically achieved by adjusting the buffer thresholds of ECN and PFC, making ECN trigger faster than PFC, thus instructing the server to actively reduce its packet transmission rate. If this still doesn't solve the problem, PFC can be used to suspend packet transmission from upstream switches, sacrificing overall network throughput to reduce the likelihood of packet loss. The two technologies work together to achieve lossless network performance.

[0005] There are also some optimization ideas for threshold setting, judgment, and convergence, such as QCN (Quantized Congestion Notification) and DCQCN (Data Center Quantized Congestion Notification). Through algorithm optimization, attempts are made to more reasonably control the thresholds of the two buffers. Other methods, such as TIMELY, control from the perspective of RTT, or ExpressPass, control from the perspective of credit, which is equivalent to changing the triggering conditions. Fundamentally, these all achieve congestion control by reducing throughput, thus achieving lossless congestion control.

[0006] Currently, all these methods ultimately control congestion by reducing the sending or forwarding rate, which leads to a significant decrease in throughput. Analysis of the causes of congestion in data center networks reveals that, since most data center networks currently use a CLOS architecture, one of the most significant reasons for congestion in this architecture is its use of ECMP (Equal Cost Multipath) to build multiple equally priced links for load balancing. This involves setting a perturbation factor and using a hash algorithm to select links for forwarding, without considering whether the selected links themselves are congested. ECMP lacks a congestion-aware mechanism; it simply distributes traffic across different links for forwarding, which can potentially exacerbate congestion on already congested links. Summary of the Invention

[0007] This application provides a method and apparatus for implementing a lossless network in a data center, in order to solve the problem of controlling congestion by reducing throughput in related technologies.

[0008] Firstly, a method for implementing a lossless network in a data center is provided, including:

[0009] The system obtains the queue buffer percentage of packets for each switch node in real time. When there is a switch node with a higher than the preset threshold, the packet is switched from the switch node with the higher than the preset threshold to an idle switch node, and the corresponding index field is constructed and sent to the corresponding TOR node device.

[0010] The TOR node device updates the corresponding index field in the diversion index in the message header and forwards the message according to the diversion index;

[0011] When the queue buffer percentage of all switch nodes is higher than the preset threshold, ECN / PFC processing is performed.

[0012] In some embodiments, the switch nodes include fabric switch nodes in a data center network architecture and spine switch nodes;

[0013] The idle switch nodes are those with the smallest queue cache percentage within the same category under the data center network architecture.

[0014] In some embodiments, the traffic splitting index consists of three index fields: the ingress direction of the fabric switch, the egress direction of the fabric switch, and the index field of the spine switch.

[0015] The traffic splitting index field is sent directly or in a configured form to the corresponding TOR node device.

[0016] In some embodiments, the packet is switched from a switch node with a value above a preset threshold to an idle switch node, including:

[0017] Create a routing mapping table for each POD, including a mapping table from the fabric switch inlet to the TOR node device, a mapping table from the fabric switch outlet to the TOR node device, and a mapping table from the spine switch to the POD.

[0018] Switching of switch nodes is achieved by modifying the corresponding routing mapping table;

[0019] The corresponding index field is constructed according to the switched switch node. The index field is one of the index fields of the fabric switch ingress direction, the fabric switch egress direction, and the spine switch.

[0020] In some embodiments, if a switch node with a value higher than a preset threshold is a fabric switch node under a POD, modifying the corresponding routing mapping table includes:

[0021] When there is load in the ingress direction, modify the mapping table from the ingress of the fabric switch to the TOR node device; when there is no load in the ingress direction, modify the mapping table from the egress of the fabric switch to the TOR node device.

[0022] After modifying the routing mapping table, rewrite the routing index of the corresponding TOR node device under that POD.

[0023] If the switch node with a value exceeding the preset threshold is a Spine switch node, modify the corresponding routing mapping table as follows:

[0024] Obtain the index field of the Spine switch node with the lowest queue cache percentage, use the index field to find and modify the mapping table from the Spine switch to the POD, and update the traffic distribution index of all TOR node devices under that POD.

[0025] On the other hand, a system for implementing a lossless network in a data center is provided, comprising:

[0026] The controller is used to obtain the queue buffer percentage of each switch node in real time. When there is a switch node with a higher than the preset threshold, the packet is switched from the switch node with the higher than the preset threshold to an idle switch node, and the corresponding index field is constructed and sent to the corresponding TOR node device. It is also used to perform ECN / PFC processing when the queue buffer percentage of all switch nodes is higher than the preset threshold.

[0027] The message forwarding module is set up on each TOR node device and is used to update the corresponding index field in the diversion index of the message header and forward the message according to the diversion index.

[0028] In some embodiments, the switch nodes include fabric switch nodes and spine switch nodes in the data center network architecture; the idle switch nodes are the switch nodes with the smallest queue cache percentage in the same category under the data center network architecture.

[0029] In some embodiments, a shunting index register is also included in each TOR node device for storing the shunting index;

[0030] The traffic splitting index consists of three index fields: the ingress direction of the fabric switch, the egress direction of the fabric switch, and the index field of the spine switch.

[0031] The controller is also used to send the split index fields directly or in a configured form to the corresponding TOR node devices.

[0032] In some embodiments, the controller is configured with multiple routing mapping tables corresponding to each POD, including a mapping table from the fabric switch inlet to the TOR node device, a mapping table from the fabric switch outlet to the TOR node device, and a mapping table from the spine switch to the POD.

[0033] The controller switches the switch nodes by modifying the corresponding routing mapping table. The split index register constructs the corresponding index field according to the switched switch node. The index field is one of the index fields of the fabric switch inlet direction, the fabric switch outlet direction, and the spine switch.

[0034] The beneficial effects of the technical solution provided in this application include:

[0035] This application modifies the header of SR. While maintaining compatibility with existing standards, it monitors the queue buffer percentage of switch node packets to perform switch node switching and makes path planning. Based on the switched path, it updates the traffic splitting index in the packet header and forwards the packet according to the traffic splitting index. When the queue buffer percentage of all switch nodes is higher than a preset threshold, ECN / PFC processing is then performed. Through reasonable path planning, network throughput is guaranteed while data flow avoids congestion. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a data center topology diagram according to an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the standard G-SID message structure;

[0039] Figure 3 This is a schematic diagram of the NG-SID message structure in an embodiment of this application;

[0040] Figure 4 This is a schematic diagram of the split index register;

[0041] Figure 5 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] This application provides a method and apparatus for implementing a lossless network in a data center, which can solve the problem of congestion control by reducing throughput.

[0044] like Figure 1The diagram shows a data center topology. The dashed box represents a Point of Display (POD). It supports 26 PODs, fully connected by 8 Spine switches (represented by SPI in the diagram). The overall structure is a 5-level CLOS architecture, supporting over 30,000 servers. Expansion based on this architecture can be applied to scales of 10,000 to 100,000 servers.

[0045] This application provides an embodiment of a method for implementing a lossless network in a data center, comprising the following steps:

[0046] The system obtains the queue buffer percentage of each switch node in real time. When a switch node has a value higher than the preset threshold, the packet is switched from the switch node with the higher threshold to an idle switch node. The system then constructs the corresponding index field and sends it to the corresponding TOR (Top of Rack) node device.

[0047] After receiving the routing index field, the TOR node device updates the corresponding index field in the routing index of the packet header because a routing index has been added to the packet header in advance, and then forwards the packet according to the updated routing index.

[0048] When the queue buffer percentage of all switch nodes is higher than the preset threshold, ECN / PFC processing is performed.

[0049] In some embodiments, the index field can be directly sent to the corresponding TOR node device. After receiving the index field, the TOR node device converts it into local configuration and updates the packet header.

[0050] In other embodiments, after constructing the index field, the corresponding configuration is sent to the corresponding TOR node device. After receiving the configuration, the TOR node device updates the message header.

[0051] In the above steps, a routing index is added to the packet header, and a G-SID header compression scheme based on common prefix concatenation is used. The G-SID scheme draft is draft-cl-spring-generalizedsrv6-for-cmpr-02. SRv6 is a combination of SegmentRouting and IPv6 network technologies, possessing both flexible routing capabilities and SRv6's unique device-level programmability. In SDN (Software Defined Network) networking, the controller is responsible for orchestrating and distributing the segment list to achieve intelligent routing. The G-SID header compression scheme can solve the problems of excessive header overhead and low carrying efficiency in practical applications of SRv6 technology; it is compatible with current network address planning, achieving an effective balance between compression efficiency and processing performance.

[0052] like Figure 2 The image shows the standard G-SID message structure. This embodiment modifies the G-SID message structure (the modified one is denoted as NG-SID), making it compatible with the standard G-SID, easy to implement, and easy to expand, suitable for data centers of various sizes.

[0053] like Figure 3 The image shows an example of a G-SID message structure. In this example, three SIDs (Segment Identifiers) are defined, each SID being 16 bits. The three SIDs have the following meanings:

[0054] FSI: Fabric Switch Ingress, the direction of entry for the fabric switch.

[0055] FSE: Fabric Switch Engress, the direction of exit for the fabric switch.

[0056] SPI: Spine Switch, the core switch of the entire network.

[0057] Figure 1 In this context, FSI stands for ingress direction of the fabric switch, FSE stands for egress direction of the fabric switch, and the fabric switch is the core switch within the POD. TORI is the data ingress TOR node device, and TORE is the data egress TOR node device.

[0058] In the above steps, the switch nodes include fabric switch nodes (hereinafter referred to as FS nodes) inside the POD within the data center network architecture, and spine switch nodes (hereinafter referred to as SPI nodes) outside the POD. Idle switch nodes are those with the lowest queue buffer percentage within the same category within the data center network architecture.

[0059] like Figure 3 and Figure 4As shown, the aforementioned traffic splitting index consists of three index fields, each indexing one of the SIDs above. The three index fields are FSII (fabric switch ingress index), FSEI (fabric switch engress index), and SPII (spine switch index); FSII and FSEI are 2 bits each, and SPII is 3 bits. To ensure compatibility with standard G-SIDs, the index fields share one byte with the SI in the standard G-SID message structure. Each TOR node device stores the traffic splitting index, converts it to local configuration, and writes it into the header of the NG-SID message.

[0060] like Figure 4 As shown, a flow index register (FR-MAP) can be created on the TOR node device to store the flow index. In this embodiment, the flow index register has three fields, with the flow index being FSII+SPII+FSEI, totaling 8 bits. FSII has 2 bits, indexing 4 FSI nodes; FSEI has 2 bits, indexing 4 FSE nodes; and SPII has 3 bits, indexing 8 SPI nodes. Specifically, the controller sends the configuration information of the flow index register to the TOR node device. The TOR node device queries the values ​​of FSII+SPII+FSEI, selects a path according to the corresponding combination, and writes it into the header of the NG-SID message.

[0061] In some embodiments, the step of obtaining the real-time queue buffer percentage of each switch node can be implemented by monitoring the switch node queues. An FS-Quere (FS node queue) table is created, as shown in Table 1, recording the buffer percentage of each FS node. The size of Table 1 is POD-Number * FS-Number bytes (26 * 4 = 104 bytes in this implementation). The first column is the POD, and each row is the Queue of the FS node under one POD. The controller can obtain the buffer percentage of each FS node via the telemetry protocol and record it in Table 1.

[0062] Table 1

[0063] POD ID FS-1 FS-2 FS-3 FS-4 1 Quere Quere Quere Quere 2 Quere Quere Quere Quere 3 Quere Quere Quere Quere 4 Quere Quere Quere Quere ...... ...... ...... ...... ...... 26 Quere Quere Quere Quere

[0064] Similarly, create an SPI-Quere (SPI node queue) table, as shown in Table 2, to record the percentage of buffer for each SPI node. The size of Table 2 is SPI-Number bytes (8 bytes in this implementation case). The controller can obtain the percentage of buffer size for each SPI node through the telemetry protocol and record it in Table 2.

[0065] Table 2

[0066] SPI-1 SPI-2 SPI-3 SPI-4 SPI-5 SPI-6 SPI-7 SPI-8 Quere Quere Quere Quere Quere Quere Quere Quere

[0067] In the above steps, when there is a switch node with a value higher than the preset threshold, the system switches to an idle switch node and constructs the corresponding index field, specifically including:

[0068] Three routing mapping tables are created for each POD. The routing mapping tables include a mapping table from the fabric switch ingress (FSI) to the TOR node device, a mapping table from the fabric switch egress (FSE) to the TOR node device, and a mapping table from the spine switch (SPI) to the POD. Switch nodes are switched by modifying the corresponding routing mapping tables, and an index field is constructed according to the switched path. The index field is one of FSII, SPII, and FSEI.

[0069] Table 3 is an FSI-TOR-BITMAP table (a mapping table from the fabric switch ingress to the TOR node device) corresponding to a POD. Figure 1 The data center network supports 26 PODs, therefore Table 3 contains 26 tables, indexed by POD: FSI-TOR-BITMAP [POD index]. Table 3 is a bitmap table with a size of FSI-Number*TOR-number (4*120 bits in this embodiment). Each field uses 1 bit set to 1 to identify which FSI the TOR node device is using, and the other bits are 0.

[0070] Table 3

[0071]

[0072] Table 4 is an FSE-TOR-BITMAP table corresponding to a POD (a mapping table from the fabric switch egress to the TOR node device). Figure 1The data center network supports 26 PODs, therefore there are 26 Table 4 tables, indexed by POD: FSE-TOR-BITMAP[POD index]. The size of each table is FSE-Number*TOR-number (4*120 bits in this embodiment). Table 4 is also a bitmap table, where each field uses 1 bit set to 1 to indicate which FSE the TOR node device uses, and the other bits are 0.

[0073] Table 4

[0074]

[0075] Table 5 is a SPI-POD-BITMAP table. Figure 1 In the data center network, there is only one table 5. The SPI path switching is based on POD as the smallest granularity. The size of table 5 is SPI-Number*POD-Numble (8*26 bits in this embodiment). Each field uses 1 bit set to 1 to identify which SPI POD is using, and the other bits are 0.

[0076] Table 5

[0077]

[0078]

[0079] Specifically, based on the tables above, a threshold value is preset, such as... Figure 5 As shown, the controller can cache this preset threshold value. The control plane where the controller resides initiates the telemetry protocol to obtain the queue cache percentage of each switch node in the path in real time, and compares it with the preset threshold value. Specifically, the following scenarios apply:

[0080] The first method is as follows: When the queue buffer percentage of a certain FSI node under a certain POD is higher than the preset threshold, the index field FSII of the FSI node with the smallest buffer percentage under that POD is obtained so that it can be sent to the TOR node device in the future. Then, the routing mapping table is queried using that POD as the index.

[0081] Furthermore, to switch the FSI node path for the TOR node device: query the FSI-TOR-BITMAP table corresponding to the POD, find the field marked as 1 in the FSI node row corresponding to the buffer percentage exceeding the preset threshold and set it to 0 (FSI-TOR-BITMAP[POD][FS-old][TOR]=0). Then, in the FSI node row with the minimum buffer percentage, find the field in the same column and modify its marker to 1 (FSI-TOR-BITMAP[POD][FS-new][TOR]=1). If in Table 3, the buffer percentage of FSI1 exceeds the preset threshold, and the FSI node with the minimum buffer percentage is FSI3, then set the fields marked as 1 in the row where FSI1 is located, namely TOR-1 and TOR-5, to 0, and then set TOR-1 and TOR-5 in the row where FSI3 is located to 1.

[0082] If all TOR values ​​under FSI-TOR-BITMAP[POD][FS-old] are 0, it means there is no load in the inbound direction, and the congestion is definitely caused by the outbound direction. Continue to query the FSE-TOR-BITMAP table, find the field with the flag 1 in the FSE node row corresponding to the buffer percentage exceeding the preset threshold, and set it to 0; then find the field in the same column in the FSE node row with the minimum buffer percentage, modify the flag to 1, and switch the FS path for TOR.

[0083] It is understandable that at this point, it is impossible for all TOR values ​​to be 0 in the FSE-TOR-BITMAP table, because there is no load in the inlet direction. If there is also no load in the outlet direction, the FS is idle, and it is impossible for the buffer to exceed the threshold.

[0084] After the routing mapping table is modified, the routing index of the corresponding TOR node device under the POD can be rewritten by modifying the FR-MAP.

[0085] The second method: When the buffer percentage of a Spine switch node (SPI node) exceeds a preset threshold, retrieve the POD index field of the SPI node with the lowest buffer percentage. This information is then used to distribute the data to the TOR node devices. The corresponding POD is then located in the SPI-POD-BITMAP table, and its original value is set to 0. Finally, the corresponding POD in the row of the SPI node with the lowest buffer percentage is set to 1. Based on the retrieved POD index field, the traffic distribution index for all TOR node devices under that POD is updated.

[0086] If the buffer percentage of all FS nodes or all SPI nodes exceeds the preset threshold, then ECN / PFC control is used.

[0087] For compatibility reasons, the SI and FR-MAP fields of the NG-SID message are designed to be shared in the above steps. Since the G-SID message only uses the lower 2 bits of SI, with an offset range of 0-3, it's possible to determine whether the message is an NG-SID message by checking if the field value is greater than 3 before processing. The pseudocode for the process is as follows, where the variables SI, SL, FSI[], SPI[], FSE[], FSII, SPII, FSEI, and Prefix have the same meanings as above. Figure 1-4 They are consistent; DIP represents the destination IP address.

[0088]

[0089]

[0090] This application also provides an embodiment of a lossless network implementation system for a data center, which can be used to implement the above method. The implementation system includes a controller and a packet forwarding module, with the packet forwarding module respectively set in each TOR node device.

[0091] The controller is used to obtain the queue buffer percentage of packets from each switch node in real time. When there is a switch node with a higher than preset threshold, the packet is switched from the switch node with the higher than preset threshold to an idle switch node, and the corresponding index field is constructed and sent to the corresponding TOR node device. It is also used to perform ECN / PFC processing when the queue buffer percentage of all switch nodes is higher than the preset threshold.

[0092] The message forwarding module is used to update the corresponding index field in the diversion index of the message header and forward the message according to the diversion index.

[0093] Furthermore, the switch nodes include fabric switch nodes and spine switch nodes within the data center network architecture. The controller maintains an FS-Quere (FS node queue) table (Table 1) and an SPI-Quere (SPI node queue) table (Table 2). The FS-Quere table records the queue buffer percentage for each FS node, and the SPI-Quere table records the queue buffer percentage for each SPI node. The controller selects the switch node with the lowest queue buffer percentage within the same category within the data center network architecture for switching and constructs a traffic splitting index for the path to be switched.

[0094] The aforementioned system also includes a traffic splitting index register, located in each TOR node device, for storing traffic splitting indexes. The traffic splitting index consists of three index fields: FSII (Fabric Switch Inbound Direction Index), FSEI (Fabric Switch Outbound Direction Index), and SPII (Spine Switch Index). Each TOR node device stores the traffic splitting index and selects the corresponding path information to write into the packet header. The controller is used to directly or in a configured manner distribute the traffic splitting index fields to the corresponding TOR node devices.

[0095] The controller also has multiple routing mapping tables for each POD, including the FSI-TOR-BITMAP table, the FSE-TOR-BITMAP table, and the SPI-POD-BITMAP table. The controller switches the switch nodes by modifying the corresponding routing mapping table, and the routing index register modifies the routing index according to the switched path.

[0096] like Figure 5 As shown, the controller controls all nodes and can obtain the buffer percentage of each node in the path in real time. By comparing it with the preset percentage threshold of the cache, it realizes the switching of different nodes and sends the index field to the TOR node device. After receiving the index field, the TOR node device (including TORI and TORE) writes the FSII+SPII+FSEI index of the FR-MAP table, and updates the FSII+SPII+FSEI in the NG-SID header.

[0097] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0098] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for implementing a lossless data center network, characterized in that, include: The system continuously monitors the queue buffer percentage of packets at each switch node. When a switch node exceeds a preset threshold, packets are switched from that node to an idle switch node. These switch nodes include fabric switch nodes and spine switch nodes within the data center network architecture. The switching process involves: creating a routing mapping table for each Point of Destination (POD), including mapping tables from fabric switch ingress to TOR node devices, fabric switch egress to TOR node devices, and spine switch to POD; modifying the corresponding routing mapping tables to switch the switch nodes; and constructing corresponding index fields based on the switched switch nodes and sending them to the corresponding TOR node devices. The TOR node device updates the corresponding index field in the diversion index in the message header and forwards the message according to the diversion index; When the queue buffer percentage of all switch nodes is higher than the preset threshold, ECN / PFC processing is performed.

2. The method for implementing a lossless data center network as described in claim 1, characterized in that: The idle switch nodes are those with the smallest queue cache percentage within the same category under the data center network architecture.

3. The method for implementing a lossless data center network as described in claim 2, characterized in that: The traffic splitting index consists of three index fields: the ingress direction of the fabric switch, the egress direction of the fabric switch, and the index field of the spine switch. The traffic splitting index field is sent directly or in a configured form to the corresponding TOR node device.

4. The method for implementing a lossless data center network as described in claim 3, characterized in that, The message is switched from a switch node with a value higher than a preset threshold to an idle switch node, including: The index field is one of the index fields for the inlet direction of the fabric switch, the outlet direction of the fabric switch, and the spine switch.

5. The method for implementing a lossless data center network as described in claim 4, characterized in that, If a switch node exceeding the preset threshold is a fabric switch node under a POD, modify the corresponding routing mapping table as follows: When there is load in the ingress direction, modify the mapping table from the ingress of the fabric switch to the TOR node device; when there is no load in the ingress direction, modify the mapping table from the egress of the fabric switch to the TOR node device. After modifying the routing mapping table, rewrite the routing index of the corresponding TOR node device under that POD.

6. The method for implementing a lossless data center network as described in claim 4, characterized in that, If the switch node with a value exceeding the preset threshold is a Spine switch node, modify the corresponding routing mapping table as follows: Obtain the index field of the Spine switch node with the lowest queue cache percentage, use the index field to find and modify the mapping table from the Spine switch to the POD, and update the traffic distribution index of all TOR node devices under that POD.

7. A system for implementing a lossless network in a data center, characterized in that, include: The controller is used to obtain the queue buffer percentage of packets from each switch node in real time. When there is a switch node with a value higher than a preset threshold, the packets are switched from the switch node with the higher threshold to an idle switch node. The switch nodes include fabric switch nodes and spine switch nodes in the data center network architecture. The controller is configured with multiple routing mapping tables for each POD, including a mapping table from the fabric switch inlet to the TOR node device, a mapping table from the fabric switch outlet to the TOR node device, and a mapping table from the spine switch to the POD. The controller switches the switch nodes by modifying the corresponding routing mapping tables. The traffic splitting index register constructs the corresponding index field according to the switched switch node and sends it to the corresponding TOR node device. It is also used to perform ECN / PFC processing when the queue buffer percentage of all switch nodes is higher than a preset threshold. The message forwarding module is set up on each TOR node device and is used to update the corresponding index field in the diversion index of the message header and forward the message according to the diversion index.

8. The system for implementing a lossless data center network as described in claim 7, characterized in that: The idle switch nodes are those with the smallest queue cache percentage within the same category under the data center network architecture.

9. The system for implementing a lossless data center network as described in claim 8, characterized in that: It also includes a shunting index register set in each TOR node device to store the shunting index; The traffic splitting index consists of three index fields: the ingress direction of the fabric switch, the egress direction of the fabric switch, and the index field of the spine switch. The controller is also used to send the split index fields directly or in a configured form to the corresponding TOR node devices.

10. The system for implementing a lossless data center network as described in claim 9, characterized in that: The index field is one of the index fields for the inlet direction of the fabric switch, the outlet direction of the fabric switch, and the spine switch.

Images