CN117729165A - Acceleration card expansion system of Ethernet interface and communication method thereof - Google Patents

Abstract

The invention discloses an acceleration card expansion system of Ethernet interface and a communication method thereof, relating to the field of data centers, and the technical scheme is as follows: the system comprises a server and an Ethernet expansion frame; a DHCP client module is deployed on the server; the Ethernet expansion frame comprises an Ethernet switching unit, a management control unit and a plurality of acceleration cards; the Ethernet exchange unit is used for completing data exchange between the server and the acceleration card and/or between different acceleration cards according to the online connection table so as to establish a communication link between the server and the acceleration card and/or between the acceleration card and the acceleration card; the plurality of accelerator cards are all provided with DHCP client modules; the management control unit is provided with a DHCP server module and an online connection table management module, the DHCP server module stores IP addresses and DHCP option information, and after receiving the information of the request IP addresses sent by the DHCP client module, the management control unit sends the IP addresses to the DHCP client module. The method is characterized by high data exchange speed, low complexity of the acceleration card and rich management and control functions.

Description

Acceleration card expansion system of Ethernet interface and communication method thereof

Technical Field

The invention relates to the field of data centers, in particular to an acceleration card expansion system of an Ethernet interface and a communication method thereof.

Background

From the common network card, the display card, the DPU and the GPU with the acceleration function, and even the AI acceleration card, the variety is more and more, and the functions are more and more enriched. The earliest network cards and display cards served as the connection devices between the server and the external devices, and later as the functions become more and more powerful, some work of the server can be allocated to these cards to be borne, and these cards become various acceleration cards. The interface between the accelerator card and the server is also a PCI/PCIE interface, as is common with the accelerator card. In some scenarios, multiple accelerator cards may be required to meet business requirements without affecting the normal operation of the server. The common server generally has only a limited number of PCIE interfaces, so that no more accelerator cards can be inserted, and in order to expand PCIE interfaces, an expansion frame based on PCIE interfaces is introduced.

The expansion problem of the acceleration card is solved based on the PCIE interface expansion frame, so that the parallel computation based on the acceleration card on a large scale is possible. The expansion frame based on the PCIE interface has the following advantages:

the expansion capability is large: 2 PCIEx16 can be extended to 16 PCIEx16 interfaces;

The expansion frame can be independently deployed, an independent heat dissipation system and an independent power supply system are configured, the problem that a connected server is blocked in a heat dissipation channel and is delayed due to insufficient power supply caused by the addition of an acceleration card is not considered, and the problems of heat dissipation and power supply are effectively solved by using the expansion frame;

however, the expansion box based on PCIE interface also has some problems as follows:

1. the problem of interface between the server and the acceleration card; the server and the acceleration card are connected by adopting PCIE extension lines, the PCIE extension lines have large line loss, and the interface card based on PCIE is not easy to be contacted and is easy to reduce speed.

2. A north-oriented performance bottleneck problem caused by a communication mechanism; the traffic of the expansion frame based on the PCIE interface can only be the traffic between the server and the acceleration card, and the limited uplink port is easy to generate a communication performance bottleneck, for example, 2 PCIEX16 uplink expansion ports are expanded to 16 PCIEX16 interfaces, the convergence ratio of 1:8 exists, and the performance bottleneck can occur when a plurality of PCIE acceleration cards simultaneously run for communication.

3. The communication mechanism causes the problem that the acceleration cards cannot be mutually visited; the PCIE interface protocol only can access the master device and the slave device mutually, and each PCIE accelerator card cannot access each other, and when communications are needed between the accelerator cards, transfer must be performed through a CPU of the server.

Management function is lost; the expansion frame is used as independent equipment, but no independent management system exists, the starting and image upgrading of the accelerator card are solved by installing a driver in the server system, and the supervision functions of the power supply and the fan of the expansion frame are lost.

Disclosure of Invention

Aiming at the problems existing in the prior art, the first object of the invention is to provide an acceleration card expansion system of an Ethernet interface, which is characterized by high data exchange speed, low complexity of the acceleration card and rich management control functions.

In order to achieve the above purpose, the invention adopts the following technical scheme: an acceleration card expansion system of Ethernet interface comprises a server and an Ethernet expansion frame; the server is provided with a network card and a DHCP client module; the Ethernet expansion frame is provided with a plurality of Ethernet communication interfaces, and the server is interconnected with the Ethernet communication interfaces of the Ethernet expansion frame through a network card; the Ethernet expansion frame also comprises an Ethernet switching unit, a management control unit and a plurality of acceleration cards; the Ethernet switching unit is connected with the Ethernet communication interface and also connected with a plurality of acceleration cards respectively, and completes data exchange between the server and the acceleration cards and/or between different acceleration cards according to an online connection table so as to establish a communication link between the server and the acceleration cards and/or between the acceleration cards, wherein two ends of the communication link are divided into a communication server and a communication client according to interaction roles; a plurality of accelerator cards are all deployed with DHCP client modules; the management control unit is connected with the Ethernet switching unit and connected with the accelerator card and/or the server through the Ethernet switching unit, and is provided with a DHCP server module and an online connection table management module, wherein the online connection table management module can change an online connection table; the DHCP server module stores the IP address and the DHCP option information, and sends the IP address to the DHCP client module after receiving the information of the request IP address sent by the DHCP client module; the online connection table comprises communication link information of the accelerator card and the server, namely communication link information of a communication server and a communication client in each communication link in the system; the communication link information comprises an online condition and a connection condition, the connection condition comprises a communication server IP address, a communication client IP address, a mapping relation and a connection state, and the online condition can be set to be online state information or offline state information according to the power-on condition of the communication server or the communication client.

Further, the system also comprises a BMC unit, wherein the BMC unit is connected with the acceleration card and is used for carrying out-of-band management on the acceleration card.

Another object of the present invention is to provide a method for updating communication connection of an acceleration card extension system of an ethernet interface, which is characterized by high data exchange speed, low complexity of the acceleration card, and rich management and control functions.

In order to achieve the above purpose, the invention adopts the following technical scheme: the communication connection updating method of the acceleration card expansion system of the Ethernet interface is applied to the acceleration card expansion system and is characterized by comprising the following steps:

powering up the server and/or the accelerator card;

the server and/or the acceleration card obtains an IP address from the management control unit through the Ethernet switching unit by utilizing a DHCP protocol;

the online connection table management module of the management control unit manually or automatically changes the online connection table according to the IP address information returned by the DHCP server module;

and the Ethernet switching unit forwards the data according to the changed online connection table so as to complete data switching.

Further, the server and/or the accelerator card acquire the IP address from the management control unit through the ethernet switching unit by using the DHCP protocol, which specifically includes the following steps:

The DHCP client of the server and/or the accelerator card sends a DHCP_discover data message for requesting the IP address to the Ethernet switching unit;

after receiving the dhcp_discover data message, the ethernet switching unit encapsulates the input port information of the dhcp_discover data message as extension header information into an eth header of the dhcp_discover data message to form a dhcp_discover data message with the extension header information, and broadcasts the dhcp_discover data message with the extension header information at each port except the input port;

after receiving the DHCP_discover data message with the extension header information, the DHCP server module analyzes the data message to obtain input port information, inquires a DHCP configuration file according to the input port information, encapsulates the obtained IP address and DHCP option information into a DHCP_Offer data message, and sends the DHCP_Offer data message to the Ethernet switching unit;

and after receiving the DHCP_Offer data message, the Ethernet switching unit removes the extension header information and forwards the extension header information to the DHCP client module.

Further, the automatic change online connection table specifically comprises the following steps:

taking the powered accelerator card or the server as one end of a communication link, and respectively taking the powered accelerator card or the server and other accelerator cards and/or servers in the system as the other ends of two ends of the communication link to form the communication link; the online connection table management module takes the IP addresses and the mapping relation thereof corresponding to the two ends of the communication link and the online state information of the powered accelerator card or the server as the communication link information to be input into the online connection table according to the IP address information returned by the DHCP server module;

The management control unit sends a connection starting message to the corresponding communication server and communication client according to the input communication link information;

after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

after receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection sign through the online connection table management module, so that the Ethernet switching unit can complete data switching according to the changed online connection table.

Further, the manual change of the online connection table specifically includes the following steps:

the online connection table management module receives communication server information and communication client information of a communication link input from the outside, and inputs the communication server IP address and the communication client IP address and the mapping relation thereof into the online connection table according to the IP address information returned by the DHCP server module;

the management control unit sends a connection starting message to the corresponding communication server and communication client according to the input communication link information;

after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

After receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection sign through the online connection table management module, so that the Ethernet switching unit can complete data switching according to the changed online connection table.

Another object of the present invention is to provide a method for monitoring communication connection of an acceleration card extension system of an ethernet interface, which is characterized by high data exchange speed, low complexity of the acceleration card, and rich management and control functions.

In order to achieve the above purpose, the invention adopts the following technical scheme: a communication connection monitoring method of an acceleration card expansion system of an Ethernet interface is applied to the acceleration card expansion system and comprises the following steps:

the management control unit sequentially sends monitoring messages to the communication service terminals with the online state information in the communication links according to the sequence of the communication links on the online connection table, wherein the monitoring messages comprise the IP information of the communication service terminals and the IP information of the communication clients in the communication links which are currently monitored;

the server receives and analyzes the monitoring message, matches the obtained IP information of the communication server and the IP information of the communication client with the local connection IP information, returns a monitoring return message containing matching success information to the management control unit if the matching is successful, otherwise returns a monitoring return message containing matching unsuccessful information;

The management control unit analyzes the message of the monitoring return message, and changes the connection condition on the online connection table according to the monitoring return message, wherein,

if the monitoring return message contains the matching unsuccessful information, the online connection table management module deletes the current monitored communication link information in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark, and if the monitoring return message contains the matching successful message, the processing is not performed;

if the management control unit does not receive the monitoring return message within the preset time, judging that the communication service end in the current monitored communication link is not on-line, and updating the on-line condition of the communication service end in the on-line connection table into off-line state information.

Another object of the present invention is to provide a method for monitoring communication connection of an acceleration card extension system of an ethernet interface, which is characterized by high data exchange speed, low complexity of the acceleration card, and rich management and control functions.

In order to achieve the above purpose, the invention adopts the following technical scheme: a communication connection monitoring method of an acceleration card expansion system of an Ethernet interface is applied to the acceleration card expansion system, and comprises the following steps:

The management control unit sequentially sends monitoring messages to communication clients with online state information in the communication links according to the sequence of the communication links on the online connection table;

the communication client receives the monitoring message and sends a monitoring return message to the management control unit; if the management control unit does not receive the monitoring return message within the preset time, judging that the communication client which is currently monitored is not online, updating the online condition of the communication client in the online connection table into offline state information, and sending connection dismantling information to the communication servers in all communication links where the communication client is located;

the communication server receives the connection demolition information, deletes the information corresponding to the current monitored communication link in the local connection IP information, and returns a connection demolition response to the management control unit;

and after receiving the connection dismantling response, the management control unit deletes the information of the communication link currently monitored in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark.

Another object of the present invention is to provide a communication scheduling method of an acceleration card extension system of an ethernet interface, which is characterized by high data exchange speed, low complexity of the acceleration card, and rich management control functions.

In order to achieve the above purpose, the invention adopts the following technical scheme: the communication scheduling method of the acceleration card extension system of the Ethernet interface is applied to the acceleration card extension system, and an upper layer application is arranged in a server and can initiate calculation; the communication scheduling method comprises the following steps:

the server creates virtual equipment according to the type of the accelerator card;

the virtual equipment establishes connection with one or more acceleration cards according to the online connection table;

the upper layer application initiates a calculation request and delivers the calculation request to the virtual equipment; the calculation request comprises data to be processed and a calculation command;

after receiving the calculation request, the virtual equipment selects an acceleration card according to a preset scheduling strategy, and sends data to be processed and a calculation command to the selected acceleration card for processing through established connection;

the acceleration card receives the data to be processed and the calculation command, processes the data to be processed and the calculation command to obtain a processing result, and returns the processing result to the virtual equipment;

and the virtual equipment receives the processing result and sends the processing result to the upper application.

Further, the scheduling policy includes a load balancing mode and a primary and backup mode, wherein,

the load balancing mode is to uniformly distribute the flow of the data to be processed and the calculation command to each acceleration card;

The main and standby modes are to select one acceleration card as a main receiver, and other acceleration cards as standby receivers, and when the main receiver has a problem, the standby receiver is switched to receive data to be processed and calculation commands and process the data.

The invention has the beneficial effects that:

1. the interface between the server and the expansion frame does not adopt a PCIE interface, but adopts a more common and loose Ethernet interface;

2. switching PCIE switching chips in the expansion frame into Ethernet chips with higher switching rate;

3. adding a CPU as a management CPU of the expansion frame;

4. adding a BMC chip as a processor with out-of-band management;

5. the accelerator card may not be limited to supporting PCIE interfaces, thereby reducing the complexity of the accelerator card.

Compared with the traditional PCIE expansion frame, the acceleration card expansion system of the Ethernet interface has the advantages of high data exchange speed, low complexity of the acceleration card, rich management control functions and the like.

Drawings

Fig. 1 is a schematic diagram of the accelerator card expansion system of embodiment 1;

fig. 2 is a schematic diagram of the structure of the module for showing the DHCP protocol processing according to embodiment 1;

fig. 3 is a flowchart of a communication connection establishment method of embodiment 2;

FIG. 4 is a timing diagram of the request and allocation of IP addresses of embodiment 2;

FIG. 5 is a flow chart of an automatic change online connection table of embodiment 2;

FIG. 6 is a flow chart of a manual change online connection table of example 2

FIG. 7 is a schematic diagram of an online connection table in one embodiment;

FIG. 8 is a flow chart of a communication connection monitoring method of embodiment 3;

FIG. 9 is a schematic diagram of an online connection table in another embodiment;

fig. 10 is a flowchart of a communication connection monitoring method of embodiment 4;

fig. 11 is a flowchart of a communication scheduling method of embodiment 5.

Description of the embodiments

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures. It should be noted that the words "front", "back", "left", "right", "upper" and "lower" used in the following description refer to directions in the drawings, and the words "bottom" and "top", "inner" and "outer" refer to directions toward or away from, respectively, the geometric center of a particular component.

Example 1: an accelerating card expanding system of Ethernet interface, as shown in figures 1 and 2, comprises a server and an Ethernet expanding frame; wherein,

the server is provided with a network card and a DHCP client module; the Ethernet expansion frame is provided with a plurality of Ethernet communication interfaces, and the server is interconnected with the Ethernet communication interfaces of the Ethernet expansion frame through a network card;

specifically, the network card of the server may be a network card supporting 40G or 100G network communication, and the ethernet communication interface of the ethernet extension frame is used to transmit data between the extension frame and the external device, and may be an ethernet optical interface supporting 40G or 100G communication. The network card and the Ethernet communication interface are in communication connection through a transmission medium.

The Ethernet expansion frame also comprises an Ethernet switching unit, a management control unit and a plurality of acceleration cards; wherein,

the Ethernet exchange unit is connected with the Ethernet communication interface and also connected with the acceleration cards respectively, and completes data exchange between the server and the acceleration cards and/or between different acceleration cards according to an online connection table so as to establish a communication link between the server and the acceleration cards and/or between the acceleration cards, and two ends of the communication link are divided into a communication server and a communication client according to interaction roles.

In particular, the ethernet switching unit may be an ethernet switching chip, and the connection interface for implementing the connection between the ethernet switching unit and the accelerator card may be an ethernet optical interface or an ethernet electrical interface supporting 10G or 25G or 40G communication. The communication server may be a server, and at this time, the communication client may be an acceleration card connected to the server; the communication server may be an accelerator card, and in this case, the communication client may be another accelerator card connected to the accelerator card, or may be a server connected to the accelerator card. The ethernet switch unit may be a SOC chip.

And the plurality of accelerator cards are all provided with DHCP client modules.

The management control unit is connected with the Ethernet switching unit and connected with the accelerator card and/or the server through the Ethernet switching unit, and is provided with a DHCP server module and an online connection table management module, and the online connection table management module can change an online connection table.

Specifically, the management control unit may be a SOC chip. In a specific embodiment, the ethernet switching unit and the management control unit are built into one SOC chip. In another specific embodiment, the ethernet switching unit and the management control unit are located in different SOC chips, respectively. At this time, the ethernet switching unit needs to encapsulate source port information of the message content for the message sent to the management control unit, so as to ensure that the management control unit can know which port of the ethernet switching unit the message comes from when receiving the message. The management control unit is also connected with the Ethernet communication interface and transmits data with the external device through the Ethernet communication interface.

The DHCP server module stores the IP address and the DHCP option information, and sends the IP address to the DHCP client module after receiving the request information sent by the DHCP client module. The DHCP server module also sends DHCP option information to the DHCP client module.

The online connection table contains the communication link information of the acceleration card and the server, namely the communication link information of the communication server and the communication client in each communication link in the system; the communication link information comprises an online condition and a connection condition, the connection condition comprises a communication server IP address, a communication client IP address, a mapping relation and a connection state, and the online condition can be set to be online state information or offline state information according to the power-on condition of the communication server or the communication client.

It will be appreciated by those skilled in the art that the management control unit can also be responsible for accelerating the image download and management of the card, accelerating the traffic monitoring and bandwidth management of the card.

The system also comprises a BMC unit, wherein the BMC unit is connected with the acceleration card and is used for carrying out-of-band management on the acceleration card.

Specifically, devices such as a fan, a power supply and a sensor are also connected to each acceleration card, a BMC unit is added as a processor for out-of-band management, and the management of the devices such as the fan, the power supply and the sensor is responsible for out-of-band management such as power supply management and state monitoring of the acceleration card. Techniques for BMC as an out-of-band management processor are well known in the art. Meanwhile, it will be understood by those skilled in the art that the BMC unit may also serve as a standby processor for managing the control unit, providing functions such as image upgrade of the accelerator card. The BMC unit also interfaces with the Ethernet and communicates data with external devices via the Ethernet interface.

The connection interface between the BMC and the accelerator card can be I2C, RS, RS485 and other interfaces.

The accelerating card expansion system designs an accelerating card expansion frame based on an Ethernet interface, avoids the problems of PCIE extension line loss and poor contact of PCIE interface cards in the traditional PCIE expansion frame, has good system reliability based on an Ethernet switching technology, and can realize the mutual access among the accelerating cards. PCIE switching chips in the expansion frame are switched into Ethernet chips with higher switching rate, and the upper limit of the data switching rate is high. And the BMC chip is added to serve as a processor for out-of-band management, so that the management function of the system is expanded. Meanwhile, the acceleration card does not need to support a PCIE interface, so that the acceleration card with lower complexity can be used, and the calculation bottleneck of the acceleration card can be broken through.

Example 2: a communication connection establishment method of an acceleration card expansion system of an Ethernet interface is applied to the acceleration card expansion system of the embodiment 1, as shown in FIG. 3, and comprises the following steps:

s1.1, powering on a server and/or an acceleration card;

s1.2, a server and/or an acceleration card acquire an IP address from a management control unit through an Ethernet switching unit by utilizing a DHCP protocol;

s1.3, an online connection table management module of the management control unit manually or automatically changes an online connection table according to IP address information returned by the DHCP server module;

s1.4, the Ethernet switching unit forwards data according to the online connection table so as to complete data switching.

The steps S1.1, S1.2 and S1.3 are sequentially executed, and no requirement of sequence is executed between the steps S1.4 and the steps S1.1, S1.2 and S1.3, namely, whether a server and an acceleration card are powered on or not and an IP is acquired online or not, whether an online connection table is changed or not, the Ethernet exchange unit forwards data according to the online connection table and completes data exchange.

The server and/or the accelerator card in step S1.2 obtains the IP address from the management control unit through the ethernet switching unit by using the DHCP protocol, and specifically includes the following steps:

The DHCP client of the server and/or the accelerator card sends a DHCP_discover data message for requesting the IP address to the Ethernet switching unit;

after receiving the dhcp_discover data message, the ethernet switching unit encapsulates the input port information of the dhcp_discover data message as extension header information into an eth header of the dhcp_discover data message to form a dhcp_discover data message with the extension header information, and broadcasts the dhcp_discover data message with the extension header information at each port except the input port;

after receiving the DHCP_discover data message with the extension header information, the DHCP server module analyzes the data message to obtain input port information, inquires a DHCP configuration file according to the input port information, packages the obtained IP address and DHCP option information into a DHCP_Offer data message with the extension header information, and sends the DHCP_Offer data message to the Ethernet switching unit; the extension header information is input port information;

and after receiving the DHCP_Offer data message with the extension header information, the Ethernet switching unit removes the extension header information and forwards the extension header information to the DHCP client module.

In this scheme, as shown in fig. 4, the process of requesting and allocating an IP address is specifically as follows:

Dhcp discovery:

when the server or accelerator card is started and an IP address is required, their DHCP client module sends a DHCP Discover data message to the ethernet switching unit. The ethernet switching unit encapsulates the input port information of the dhcp_discover data packet as extension header information into an eth header of the dhcp_discover data packet to form a dhcp_discover data packet with the extension header information, and then performs flood (broadcast) on each port, where the data packet is received by all devices in the network except the device requesting the IP address, including a management control unit in the extension frame.

Dhcp provides:

after receiving the DHCP_discover data message with the extension header information, the DHCP server module of the management control unit obtains the input port information by analyzing the data message, inquires a DHCP configuration file according to the input port information, encapsulates the obtained IP address and DHCP option information into a DHCP_Offer data message with the extension header information, and sends the DHCP_Offer data message to the Ethernet switching unit. The dhcp_offer data packet includes DHCP option information such as an IP address, a subnet mask, an address lease, and the like. When the data message passes through the ethernet switching unit, the ethernet switching unit strips the extension header information and forwards the extension header information to the DHCP client module requesting the IP address.

Dhcp request:

after receiving the dhcp_offer data message, the DHCP client module of the server or accelerator card requesting the IP address returns a dhcp_request data message in a broadcast manner through the ethernet switch chip. The data message contains information such as the MAC address of the client, the IP address in the lease received, the DHCP server module address that provided the lease, etc.

Dhcp acknowledgement:

after receiving the dhcp_request of the DHCP client, the DHCP server module of the management control unit confirms the allocation of the IP address, that is, confirms the allocation of the IP address to the DHCP client and broadcasts a dhcp_ack message back to the DHCP client through the ethernet switch chip, indicating that the selection of the DHCP client has been accepted, and telling the DHCP client that the DHCP client can use the IP address provided by the dhcp_ack message. And putting legal lease and other configuration information of the IP address into the DHCP_ack message to other DHCP client modules.

5. Address usage:

after receiving the dhcp_ack message, the DHCP client module of the acceleration card or the server requesting the IP address sends three ARP resolution requests for the IP address to other devices in the network through the ethernet switch chip to perform conflict monitoring, so as to query whether other machines use the IP address on the network. Meanwhile, the ethernet switch unit will maintain MAC tables for each port.

The ethernet switch unit will keep a MAC table of the respective ports, a so-called MAC learning procedure. Since the ethernet switch unit is responsible for forwarding data packets from a source port to one or more destination ports, in order to do this efficiently, the ethernet switch unit needs to know the MAC address of each device (e.g., server, accelerator card, etc.) connected to its port. The specific steps of the MAC learning process are as follows:

when the ethernet switch unit receives a data message, it checks the source MAC address of the data message. The ethernet switch unit looks up this source MAC address in an internal MAC address table (also called forwarding table or MAC table). If this address is not already in the table, the ethernet switch unit will add the source MAC address to the MAC table along with the port number from which the data packet was received.

The ethernet switch unit will also look at the destination MAC address of the data packet. It will then look up this target MAC address in the MAC table. If a matching entry is found, the ethernet switch unit forwards the data packet to the port associated with the destination MAC address. If no matching entry is found, the ethernet switch unit will typically broadcast the data message to all ports (except the source port).

The ethernet switching unit will periodically update its MAC table to maintain the accuracy of the information. If a MAC address does not appear in the source address of any data message for a period of time, the ethernet switch may assume that the device has been removed from the network and then delete the address from the MAC table.

The MAC learning process enables the Ethernet switch unit to effectively forward the data message from the source port to the target port without broadcasting the data message to all ports, thereby improving network efficiency.

In summary, in this process, the DHCP server module of the management control unit and the DHCP client module of the server or the accelerator card interact through four steps (discovery, provision, request, confirmation), and finally, dynamic allocation of IP addresses is achieved. Meanwhile, the Ethernet switching unit is responsible for transmitting and processing the data message in the whole process.

Further, the automatic change online connection table, as shown in fig. 5, specifically includes the following steps:

s1.3.1, using the powered accelerator card or server as one end of a communication link, and using the powered accelerator card or server and other accelerator cards and/or servers in the system as the other ends of the two ends of the communication link respectively to form a communication link; the online connection table management module takes the IP addresses and the mapping relation thereof corresponding to the two ends of the communication link and the online state information of the powered accelerator card or the server as the communication link information to be input into the online connection table according to the IP address information returned by the DHCP server module;

S1.3.2, the management control unit sends a connection start message to the corresponding communication server and communication client according to the input communication link information;

s1.3.3, after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

s1.3.4, after receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection flag through the online connection table management module, so that the ethernet switching unit can complete data switching according to the changed online connection table.

Further, the manual change on-line connection table, as shown in fig. 6, specifically includes the following steps:

s1.3.5 the online connection table management module receives communication server information and communication client information of a communication link input from the outside, and the online connection table management module inputs the IP address of the communication server and the IP address of the communication client into the online connection table according to the IP address information returned by the DHCP server module;

s1.3.6, the management control unit sends a connection start message to the corresponding communication server and communication client according to the input communication link information;

S1.3.7, after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

s1.3.8, after receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection flag through the online connection table management module, so that the ethernet switching unit can complete data switching according to the changed online connection table.

In a specific embodiment, the format of the online connection table is shown in fig. 7, where each row of the table contains communication link information of one communication link, and in each piece of communication link information, an IP address and an online condition of the communication server, an IP address and an online condition of the communication client, and a connection condition of the communication link are displayed. The mapping relation between the IP address of the communication server and the IP address of the communication client can be displayed through the IP address of the communication server and the IP address of the corresponding communication client in each row. The online status in the table contains online status information "online" indicating that the server and/or accelerator card is in a powered-on state and offline status information "offline" indicating that the server and/or accelerator card is in a non-powered-on state. The connection status of the communication link in the table includes a connection-in flag "connected" flag indicating that the communication link is in a connected state, and a disconnection flag "disconnected" flag indicating that the communication link is in a disconnected state. Only when the ethernet switching unit receives a message sent by a communication service end in an "online" power-on state, a communication link where the communication service end is located is in a "connected" connection state, and the power-on state of a corresponding communication client is "online", the ethernet switching unit forwards the message to an IP address of the corresponding communication client. And the data forwarding from the communication client to the communication server is the same. Therefore, of the two communication links illustrated in the figure, the communication link with the sequence number 1 can normally transmit data, and the communication link with the sequence number 2 cannot normally transmit data.

Example 3: a communication connection monitoring method of an acceleration card expansion system of an Ethernet interface is applied to the acceleration card expansion system of the embodiment 1, as shown in FIG. 8, and comprises the following steps:

s2.1, the management control unit sequentially sends monitoring messages to the communication service end with the online state information in the communication link according to the sequence of the communication link on the online connection table, wherein the monitoring messages comprise the IP information of the communication service end and the IP information of the communication client end in the communication link which is currently monitored;

s2.2, the server receives and analyzes the monitoring message, matches the obtained IP information of the communication server and the IP information of the communication client with the local connection IP information, returns a monitoring return message containing matching success information to the management control unit if the matching is successful, and returns a monitoring return message containing matching unsuccessful information if the matching is not successful;

s2.3, the management control unit analyzes the monitoring return message, and changes the connection condition on the online connection table according to the monitoring return message, wherein,

if the monitoring return message contains the matching unsuccessful information, the online connection table management module deletes the current monitored communication link information in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark, and if the monitoring return message contains the matching successful message, the processing is not performed;

If the management control unit does not receive the monitoring return message within the preset time, judging that the communication service end in the current monitored communication link is not on-line, and updating the on-line condition of the communication service end in the on-line connection table into off-line state information.

In a specific embodiment, the format of the online connection table is shown in fig. 9, where each row of the table contains communication link information of one communication link, and in each piece of communication link information, an IP address and an online condition of the communication server, an IP address and an online condition of the communication client, a connection condition of the communication link, and a last check time are displayed. The management control unit reads the last inspection time and executes the monitoring method of the present embodiment after a preset timing time.

Example 4: a communication connection monitoring method of an acceleration card expansion system of an Ethernet interface is applied to the acceleration card expansion system of the embodiment 1, as shown in FIG. 10, and comprises the following steps:

s3.1, the management control unit sequentially sends monitoring messages to communication clients with online state information in the communication links according to the sequence of the communication links on the online connection table;

s3.2, the communication client receives the monitoring message and sends a monitoring return message to the management control unit; if the management control unit does not receive the monitoring return message within the preset time, judging that the communication client which is currently monitored is not online, updating the online condition of the communication client in the online connection table into offline state information, and sending connection dismantling information to the communication servers in all communication links where the communication client is located;

S3.3, the communication server receives connection demolition information, deletes information corresponding to the currently monitored communication link in the local connection IP information, and returns a connection demolition response to the management control unit;

and S3.4, after receiving the connection dismantling response, the management control unit deletes the information of the communication link currently monitored in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark.

In a specific embodiment, the format of the online connection table is shown in fig. 10, where each row of the table contains communication link information of one communication link, and in each piece of communication link information, an IP address and an online condition of the communication server, an IP address and an online condition of the communication client, a connection condition of the communication link, and a last check time are displayed. The management control unit reads the last inspection time and executes the monitoring method of the present embodiment after a preset timing time.

Example 5: a communication scheduling method of an acceleration card extension system of an Ethernet interface is applied to the acceleration card extension system of the embodiment 1, and an upper layer application is arranged in a server, wherein the upper layer application can initiate calculation; as shown in fig. 11, the communication scheduling method includes the steps of:

S4.1, the server creates virtual equipment according to the type of the accelerator card;

s4.2, the virtual equipment establishes connection with one or more acceleration cards according to the online connection table;

s4.3, the upper layer application initiates a calculation request and delivers the calculation request to the virtual equipment; the calculation request comprises data to be processed and a calculation command;

s4.4, after the virtual equipment receives the calculation request, selecting an acceleration card according to a preset scheduling strategy, and sending data to be processed and a calculation command to the selected acceleration card for processing through established connection;

s4.5, after receiving the data to be processed and the calculation command, the accelerator card processes the data to be processed and the calculation command to obtain a processing result, and returns the processing result to the virtual equipment;

and S4.6, after receiving the processing result, the virtual equipment sends the processing result to the upper layer application.

Further, the scheduling policy includes a load balancing mode and a primary and backup mode, wherein,

the load balancing mode is to uniformly distribute the flow of the data to be processed and the calculation command to each acceleration card;

the main and standby modes are to select one acceleration card as a main receiver, and other acceleration cards as standby receivers, and when the main receiver has a problem, the standby receiver is switched to receive data to be processed and calculation commands and process the data.

Specifically, the server creates a virtual device based on the accelerator card type. For example, if the server needs to perform privacy calculations, it may create a privacy computing device. This virtual device is in fact a software entity that represents the connection between the server and the accelerator cards.

The virtual device automatically chooses to establish a connection with multiple accelerator cards. This procedure refers to the connection establishment procedure in embodiment 2. Briefly, a virtual device will automatically establish a connection with multiple accelerator cards based on information in an online connection table.

An upper layer application (e.g., a program running on a server) may initiate a computing task. This task requires processing some data and may require executing some specific commands. The upper layer application will send these data and commands to the virtual device.

After receiving the data and the command, the virtual device starts task allocation. This process is performed according to a certain scheduling policy. For example, if the scheduling policy is load balancing, the virtual device may decide which accelerator cards to assign tasks based on the current load situation of each accelerator card, and the status of each connection.

In some embodiments, the server may collect some physical characteristic data of the accelerator card, such as the total bandwidth of the interface, link traffic statistics, etc. Then, based on these data, the virtual device calculates the remaining bandwidth of each accelerator card, and performs task allocation based on the remaining bandwidth. For example, the server is connected with three accelerator cards, the total bandwidths between the accelerator cards and the server are respectively 10Gbps,20Gbps and 30Gbps, the existing flows between the accelerator cards and the server are respectively 5Gbps,4Gbps and 6Gbps, and the residual bandwidths between the accelerator cards and the server are respectively 5Gbps,16 Gbps and 24 Gbps. At this time, if the load balancing mode is adopted, the traffic is distributed to the accelerator cards in a balanced manner according to the ratio of 5:16:24, and if the primary/secondary mode is adopted, one connection with larger residual bandwidth is selected as a primary traffic receiver, and the rest accelerator cards are used as standby receivers. The flow is the data quantity transmitted in unit time, and the transmitted data mainly comprises data to be processed and calculation commands.

The virtual device will send the data to the selected accelerator card for processing over the established connection. This process involves the operations of encoding and decoding the data to ensure that the data can be properly received and processed by the accelerator card.

After the accelerator card receives the data, it starts processing. After the processing is completed, the accelerator card returns the result to the virtual device.

After receiving the result, the virtual device will transmit the result to the upper layer application. Thus, the upper layer application can acquire the processing result to finish one calculation task.

In this process, the virtual device needs to interact frequently with the online connection table. The online connection table provides the virtual device with detailed information of all available connections in the current system, including the status of each connection, the online status of each accelerator card, etc. This information is very important to the virtual devices, which can help the virtual devices make more reasonable scheduling decisions. At the same time, the operation of the virtual device may also affect the online connection table, for example, when a new connection is established or an old connection is removed, the online connection table management module needs to change the relevant information in the online connection table.

It should be noted that the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same. While the invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents substituted for elements thereof without departing from the scope of the invention, which is to be encompassed by the appended claims.

Claims (10)

1. An acceleration card expansion system of an ethernet interface, characterized in that: the system comprises a server and an Ethernet expansion frame; wherein,

the server is provided with a network card and a DHCP client module;

the Ethernet expansion frame is provided with a plurality of Ethernet communication interfaces, and the server is interconnected with the Ethernet communication interfaces of the Ethernet expansion frame through a network card; the Ethernet expansion frame also comprises an Ethernet switching unit, a management control unit and a plurality of acceleration cards; wherein,

the Ethernet exchange unit is connected with the Ethernet communication interface and also connected with a plurality of acceleration cards respectively, and completes data exchange between the server and the acceleration cards and/or between different acceleration cards according to an online connection table so as to establish a communication link between the server and the acceleration cards and/or between the acceleration cards, wherein two ends of the communication link are divided into a communication server and a communication client according to interaction roles;

a plurality of accelerator cards are all deployed with DHCP client modules;

the management control unit is connected with the Ethernet switching unit and connected with the accelerator card and/or the server through the Ethernet switching unit, and is provided with a DHCP server module and an online connection table management module, wherein the online connection table management module can change an online connection table;

The DHCP server module stores the IP address and the DHCP option information, and sends the IP address to the DHCP client module after receiving the information of the request IP address sent by the DHCP client module;

the online connection table comprises communication link information of the accelerator card and the server, namely communication link information of a communication server and a communication client in each communication link in the system; the communication link information comprises an online condition and a connection condition, the connection condition comprises a communication server IP address, a communication client IP address, a mapping relation and a connection state, and the online condition can be set to be online state information or offline state information according to the power-on condition of the communication server or the communication client.

2. An accelerator card extension system for an ethernet interface as defined in claim 1, wherein: the system also comprises a BMC unit, wherein the BMC unit is connected with the acceleration card and is used for carrying out-of-band management on the acceleration card.

3. A method for updating communication connection of an acceleration card extension system of an ethernet interface, applied to the acceleration card extension system of claim 1 or 2, comprising the steps of:

powering up the server and/or the accelerator card;

The server and/or the acceleration card obtains an IP address from the management control unit through the Ethernet switching unit by utilizing a DHCP protocol;

the online connection table management module of the management control unit manually or automatically changes the online connection table according to the IP address information returned by the DHCP server module;

and the Ethernet switching unit forwards the data according to the changed online connection table so as to complete data switching.

4. A method of updating a communication connection as claimed in claim 3, wherein: the server and/or the acceleration card acquire the IP address from the management control unit through the Ethernet switching unit by utilizing the DHCP protocol, and the method specifically comprises the following steps:

the DHCP client of the server and/or the accelerator card sends a DHCP_discover data message for requesting the IP address to the Ethernet switching unit;

after receiving the dhcp_discover data message, the ethernet switching unit encapsulates the input port information of the dhcp_discover data message as extension header information into an eth header of the dhcp_discover data message to form a dhcp_discover data message with the extension header information, and broadcasts the dhcp_discover data message with the extension header information at each port except the input port;

after receiving the DHCP_discover data message with the extension header information, the DHCP server module analyzes the data message to obtain input port information, inquires a DHCP configuration file according to the input port information, encapsulates the obtained IP address and DHCP option information into a DHCP_Offer data message, and sends the DHCP_Offer data message to the Ethernet switching unit;

And after receiving the DHCP_Offer data message, the Ethernet switching unit removes the extension header information and forwards the extension header information to the DHCP client module.

5. The communication connection updating method according to claim 4, wherein: the automatic change online connection table specifically comprises the following steps:

taking the powered accelerator card or the server as one end of a communication link, and respectively taking the powered accelerator card or the server and other accelerator cards and/or servers in the system as the other ends of two ends of the communication link to form the communication link; the online connection table management module takes the IP addresses and the mapping relation thereof corresponding to the two ends of the communication link and the online state information of the powered accelerator card or the server as the communication link information to be input into the online connection table according to the IP address information returned by the DHCP server module;

the management control unit sends a connection starting message to the corresponding communication server and communication client according to the input communication link information;

after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

after receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection sign through the online connection table management module, so that the Ethernet switching unit can complete data switching according to the changed online connection table.

6. The communication connection updating method according to claim 4, wherein: the manual change online connection table specifically comprises the following steps:

the online connection table management module receives communication server information and communication client information of a communication link input from the outside, and inputs the communication server IP address and the communication client IP address and the mapping relation thereof into the online connection table according to the IP address information returned by the DHCP server module;

the management control unit sends a connection starting message to the corresponding communication server and communication client according to the input communication link information;

after receiving the connection start message, the communication server and the communication client return a connection establishment response message to the management control unit;

after receiving the connection establishment response message, the management control unit changes the connection state of the communication link on the online connection table into a connection sign through the online connection table management module, so that the Ethernet switching unit can complete data switching according to the changed online connection table.

7. A method for monitoring communication connection of an acceleration card extension system of an ethernet interface, applied to the acceleration card extension system of claim 1 or 2, comprising the following steps:

The management control unit sequentially sends monitoring messages to the communication service terminals with the online state information in the communication links according to the sequence of the communication links on the online connection table, wherein the monitoring messages comprise the IP information of the communication service terminals and the IP information of the communication clients in the communication links which are currently monitored;

the server receives and analyzes the monitoring message, matches the obtained IP information of the communication server and the IP information of the communication client with the local connection IP information, returns a monitoring return message containing matching success information to the management control unit if the matching is successful, otherwise returns a monitoring return message containing matching unsuccessful information;

the management control unit analyzes the monitoring return message, and changes the connection condition on the online connection table according to the monitoring return message, wherein,

if the monitoring return message contains the matching unsuccessful information, the online connection table management module deletes the current monitored communication link information in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark, and if the monitoring return message contains the matching successful message, the processing is not performed;

if the management control unit does not receive the monitoring return message within the preset time, judging that the communication service end in the current monitored communication link is not on-line, and updating the on-line condition of the communication service end in the on-line connection table into off-line state information.

8. A method for monitoring communication connection of an acceleration card extension system of an ethernet interface, applied to the acceleration card extension system of claim 1 or 2, comprising the following steps:

the management control unit sequentially sends monitoring messages to communication clients with online state information in the communication links according to the sequence of the communication links on the online connection table;

the communication client receives the monitoring message and sends a monitoring return message to the management control unit; if the management control unit does not receive the monitoring return message within the preset time, judging that the communication client which is currently monitored is not online, updating the online condition of the communication client in the online connection table into offline state information, and sending connection dismantling information to the communication servers in all communication links where the communication client is located;

the communication server receives the connection demolition information, deletes the information corresponding to the current monitored communication link in the local connection IP information, and returns a connection demolition response to the management control unit;

and after receiving the connection dismantling response, the management control unit deletes the information of the communication link currently monitored in the online connection table or changes the connection state of the communication link on the online connection table into a non-connection mark.

9. A communication scheduling method of an acceleration card extension system of an ethernet interface, applied to the acceleration card extension system of claim 1 or 2, characterized in that: the server is provided with an upper layer application, and the upper layer application can initiate calculation; the communication scheduling method comprises the following steps:

the server creates virtual equipment according to the type of the accelerator card;

the virtual equipment establishes connection with one or more acceleration cards according to the online connection table;

the upper layer application initiates a calculation request and delivers the calculation request to the virtual equipment; the calculation request comprises data to be processed and a calculation command;

after receiving the calculation request, the virtual equipment selects an acceleration card according to a preset scheduling strategy, and sends data to be processed and a calculation command to the selected acceleration card for processing through established connection;

the acceleration card receives the data to be processed and the calculation command, processes the data to be processed and the calculation command to obtain a processing result, and returns the processing result to the virtual equipment;

and the virtual equipment receives the processing result and sends the processing result to the upper application.

10. The communication scheduling method according to claim 9, wherein: the scheduling policy includes a load balancing mode and a primary and backup mode, wherein,

The load balancing mode is to uniformly distribute the flow of the data to be processed and the calculation command to each acceleration card;

the main and standby modes are to select one acceleration card as a main receiver, and other acceleration cards as standby receivers, and when the main receiver has a problem, the standby receiver is switched to receive data to be processed and calculation commands and process the data.