32 results about "InfiniBand" patented technology

A coherent, intelligent, packet-switched memory fabric enables predictive, cache-coherent access across distributed compute, accelerator, and memory resources using a Memory-Fabric Transaction Layer Protocol (MF-TLP). MF-TLP defines routable packet formats for read, write, vectorized, atomic, reduction, collective, and predictive-prefetch transactions executed by memory-centric network interface controllers (MC-NICs). Each MC-NIC performs packet parsing, address translation, coherence management, and near-memory arithmetic or tensor operations while coordinating with MF-TLP-aware switches providing hierarchical directory control, multi-path routing, and in-network aggregation. Vectorized and multimodal packets encode multiple addresses or tensor offsets to reduce scatter / gather overhead, and programmable caching and quality-of-service modules manage tiered memory and tenant fairness. MF-TLP supports extension headers for predictive prefetch, collective coordination, and tenant governance, operating across hierarchical leaf-spine topologies using Ultra-Ethernet Transport, InfiniBand, or CXL fabrics. The system delivers scalable, low-latency, memory-centric orchestration for large-language-model training, multimodal AI, and data-intensive analytics.

The invention relates to an 800G optical module InfiniBand networking test method, which comprises the following steps of: respectively connecting two ports of a switch with network cards of two servers by using an 800G optical module, and configuring the network cards into an InfiniBand mode; the two servers are used as a server side and a client side respectively, the server side and the client side are configured through the network card, an actual flow report of the 800G optical module is read, and an 800G optical module InfiniBand networking test environment is constructed; and according to the 800G optical module InfiniBand networking test environment, respectively carrying out a stability test, a function availability test and a robustness test on the 800G optical module. According to the method disclosed by the invention, the current network application can be simulated, the performance of the optical module in the server intercommunication service can be verified, and the function availability of the module can be evaluated; the possible operation of the current network on the server and the switch can be simulated, and the robustness of the module is evaluated; different use environments of the existing network can be simulated, and the use stability of the module is evaluated.

Technologies for spreading a burst of data across multiple network paths in remote direct memory access (RDMA) over converged Ethernet (ROCE) and InfiniBand are described. A RDMA adapter receives, from a requestor device over a local interface, a request to send data of a transport flow directed to a target device over a network interface, and one or more parameters being related to a multipath selection by the network controller. The RDMA adapter sends a first burst of data of the transport flow via a first network path to the target device. The RDMA adapter identifies, using the one or more parameters, a second network path to the target device. The RDMA adapter sends the second burst of data to the target device on the second network path.

The application relates to a network management method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: detecting whether the initial connection state of a to-be-probed connection link in an infinite bandwidth network is in a normal state; the to-be-probed connection link comprises a first connection end and a second connection end; if the initial connection state of the to-be-probed connection link is in the normal state, a fault detection instruction is sent from the first connection end of the to-be-probed connection link to the second connection end of the to-be-probed connection link; the fault detection instruction comprises management data reports corresponding to a subnet management interface in the infinite bandwidth network; the to-be-probed connection link is subjected to fault detection according to the fault detection instruction, a fault detection result is generated, and the infinite bandwidth network is subjected to network management based on the fault detection result. The method can improve the security of the IB network.

This invention discloses a communication middleware and method based on InfiniBand. In one specific embodiment, the communication middleware includes a master node, multiple computing nodes and interface nodes, an information matching and parameter exchange module, a communication link establishment module, and a data transmission module. Each computing node includes a receiver and a sender. The information matching and parameter exchange module is used to perform information matching and parameter exchange between the sender and receiver using a publish-subscribe model. The communication link establishment module is used to establish a link between the sender and receiver using a two-way handshake. The data transmission module is used to transmit data using a packet header and payload. This embodiment combines the InfiniBand-based communication middleware with a general distributed software framework, effectively solving current communication application problems and improving the stability and reliability of the middleware.

Embodiments of the present application provide a task resource allocation method and apparatus, an electronic device, and a storage medium. The method comprises: acquiring a topology structure corresponding to a device node of a task to be executed, and reconstructing a secondary resource tree in the topology structure, wherein the secondary resource tree comprises switching unit (SU) nodes and InfiniBand device child nodes under the SU nodes; in response to issuing of a task instance to which an InfiniBand device resource is to be allocated, determining idle InfiniBand device child nodes in the secondary resource tree as candidate child nodes for the task instance; determining, from among the candidate child nodes, a target child node for allocating the task instance to the same SU node; and allocating the task instance to the target child node, and executing the task instance on the target child node. In the present application, allocating instances of the same task to device nodes within the same SU reduces cross-SU communication, thereby decreasing the data exchange amount of a switch, and improving resource management efficiency and task execution efficiency.

The application discloses an Infiniband protocol conversion method and system, and relates to the technical field of computer network communication, and comprises the following steps: receiving an SMP data packet, initializing a hop count counter and a hop count pointer; performing routing decision based on the values of the hop count counter and the hop count pointer to forward the SMP data packet; and sending the SMP data packet to an Ethernet transceiver by an IB link layer subnet management agent node to convert the SMP data packet into an Ethernet data packet for forwarding. The application realizes lossless conversion between Infiniband and Ethernet protocols and supports long-distance RDMA communication across data centers.

The application discloses an InfiniBand transport layer hardware accelerator based on FPGA, which comprises a work request management module, a QP context record module, a protocol header generation module, a network message assembly module, a reply processing module and a completion queue element generation module, wherein the work request management module is used for completing instruction caching and distribution; the QP context record module is used for storing and maintaining the running state information of multiple QP channels; the protocol header generation module is used for scheduling multiple QPs and generating corresponding protocol packet headers; the network message assembly module is used for carrying out data splicing and assembling network messages; the reply processing module is used for completing message classification processing, retransmission triggering and flow control credit release; and the completion queue element generation module is used for confirming instruction completion states; and the reset control module is used for receiving global reset and single QP reset instructions and controlling the orderly execution of the reset process. The application unloads the whole-process logic such as instruction receiving and scheduling, protocol packet header generation, network message assembly, reply processing, state reporting and reset control to FPGA hardware execution, so that the deep closed loop of the data plane and the control plane is realized.

The invention provides a topology grouping method and device for a multi-track InfiniBand network. The method comprises the following steps: acquiring link information and port information of each device of the multi-track InfiniBand network; grouping the servers based on the link information and the port information of the servers to obtain a plurality of server groups; constructing a corresponding super computing unit based on each server group and the Leaf switch connected with the server group; based on the link information and the port information of the Leaf switch of each super computing unit, constructing a Spine switch group; and constructing a Core switch group based on the link information and the port information of the Spine switch of each Spine switch group. According to the method provided by the invention, through topological grouping of the multi-track InfiniBand network, efficient utilization of cluster resources and stable output of computing performance can be guaranteed.

The application discloses a large-bit-width InfiniBand network encoding and decoding layer processing system, which comprises a transmitter module, a transmission channel synchronization module, a scrambling module, an alignment module, a descrambling module, a receiving channel synchronization module and a receiver module; each module comprises two data paths of 8b10b encoding and 64b66b encoding, so as to be selected according to actual rate requirements, the bit width of each channel data path is 64b under 8b10b encoding, and the bit width of each channel data path is 256b under 64b66b encoding, and the 8b10b encoding data path bypasses the scrambling module and the descrambling module. The large-bit-width InfiniBand network encoding and decoding layer processing system has full-duplex data transmission mode, has 8b10b and 64b66b physical layer data encoding and decoding functions, supports large-bit-width multi-channel IB protocols, realizes channel alignment and inter-channel skew elimination technology, and can provide high-bandwidth and low-delay data transmission services in combination with IB upper layer design.

The invention discloses an RDMA-based phased array radio astronomical real-time data stream transmission processing method and system, and the method comprises the steps: S1, data receiving preparation: creating a queue pair of an original data packet type at a receiving end server through the RDMA, and constructing an annular receiving buffer area; s2, flow rule configuration and data capture: creating a flow filtering rule by using an InfiniBand flow control interface, directly writing a matched data flow into an annular receiving buffer area, and bypassing an operating system kernel protocol stack; s3, real-time signal processing: performing fast Fourier transform and spectrum accumulation processing on the extracted time domain signal to generate spectrum data; and S4, data persistent storage: transmitting the processed spectrum data to a host memory, and writing the spectrum data into a disk file in batches through a buffer mechanism. According to the invention, ultra-high-speed data transmission is realized by using zero copy of the RDMA and a kernel bypass mechanism, the end-to-end delay is reduced to a microsecond level, and the throughput, real-time performance and reliability of phased array radio astronomical observation data are remarkably improved.

Systems and methods herein are for one or more processing units of a subnet manager (SM) to communicate configuration information with at least one subnet management agent (SMA) that is associated with at least one switch and with a host machine, the configuration information to enable the at least one switch to configure a forwarding table based in part on a mapping of at least one virtual network address to physical network addresses of two or more physical ports of the host machine, and the configuration information to enable the host machine to communicate with other host machines using the at least one switch and the at least one virtual network address.

Systems and methods for using InfiniBand routing algorithms for Ethernet fabrics in a high performance computing environment. The method can provide, at a computer comprising one or more microprocessors, a plurality of switches, a plurality of hosts, a topology provider (TP) module, a routing engine (RE) module, and a switch initializer (SI) module. The method can perform a discovery sweep, by the TP, of the plurality of hosts and the plurality of switches and assigns an address to each of the plurality of hosts and the plurality of switches. The method can calculate, by the routing engine, a routing map, based upon a routing scheme, for the plurality of hosts and the plurality of switches, the routing map comprising a plurality of forwarding tables. The method can configure, each of the plurality of switches with a forwarding table of the plurality of forwarding tables calculated by the routing engine.

This disclosure is directed towards systems and methods that improve bandwidth shifting operations of an ADPLL without losing a lock of the ADPLL and having the benefit of being able to change the bandwidth an unlimited amount of times. Indeed, a processor may transmit amplification parameters to the ADPLL to implement a bandwidth shift. The shift may occur in response to a enable signal, such as a gear trigger control signal (gear_retime signal) or a enable signal generated to cause alignment of the shifting with a clock signal (e.g., enable signal generated by AND logic gates). These systems and methods described herein many enable multiple bandwidth changing operations to occur without compromising the complexity and footprint of the system.

The present invention discloses an InfiniBand network card virtualization networking system, comprising: a driver software module located in a host InfiniBand network card driver, configured to forward configuration information of a virtual network card from an upper-layer interface to a network card hardware function module, and further configured to receive and parse InfiniBand virtualization networking messages and generate corresponding response messages; a network card hardware function module disposed in the InfiniBand network card, configured to correctly forward messages received and sent by each virtual network card according to the virtual network card configuration information issued by the driver software module; and a virtual QP mapping relationship module configured to implement a one-to-one mapping relationship between a logical QP and a virtual QP number, thereby ensuring the correctness and efficiency of the virtualization networking function. The InfiniBand network card virtualization networking system of the present invention can reduce the network card hardware resource overhead used for virtualization functions, shorten the debugging and development cycle, and thus reduce the cost of the network card.

The present disclosure provides an InfiniBand network data cache processing system and method, belonging to the field of network communication technology. The solution uses FPGAs to implement data processing and caching, significantly reducing the overhead associated with hardware equipment. Furthermore, due to features such as field programmability, the data processing mechanism can be flexibly modified. Compared to software implementations, hardware implementations offer faster processing speeds. Furthermore, the solution utilizes a parallel structure design of multiple random access memories for cross-clock domain data processing, and multiple large-capacity storage buffers for parallel storage, storing different parts of the same data packet in different large-capacity storage buffers. These designs accelerate the write and read speeds of the cache system, increase throughput, and speed up the transmission speed of the FPGA data cache processing system, preventing the FPGA data cache processing system from losing data due to congestion.

The invention discloses an Infiniband protocol conversion method and system, and relates to the technical field of computer network communication, and the method comprises the steps: receiving an SMP data packet, and initializing a hop count counter and a hop count pointer; performing routing decision based on the values of the hop count counter and the hop count pointer so as to forward the SMP data packet; and the IB link layer subnet management agent node sends the SMP data packet to an Ethernet transceiver so as to convert the SMP data packet into an Ethernet data packet for forwarding. Lossless conversion between an InfiniBand protocol and an Ethernet protocol is realized, and long-distance RDMA (Remote Direct Memory Access) communication across data centers is supported.

The invention discloses an InfiniBand network card virtualized networking system, which comprises a driving software module, a network card hardware function module, a network card hardware function module, an InfiniBand network card hardware function module and an InfiniBand network card hardware function module, wherein the driving software module is positioned in a host InfiniBand network card driving program, is used for forwarding configuration information of an upper layer interface on a virtual network card to the network card hardware function module, and is also used for receiving and analyzing an InfiniBand virtualized networking message and generating a corresponding response message; the network card hardware function module is arranged in an InfiniBand network card and is used for correctly forwarding messages transmitted and received by each virtual network card according to the virtual network card configuration information issued by the driving software module; and the virtual QP mapping relation module is used for realizing a one-to-one mapping relation from a logic QP to a virtual QP number, and ensuring the correctness and high efficiency of a virtualized networking function. According to the InfiniBand network card virtualization networking system provided by the invention, the hardware resource overhead of the network card for a virtualization function can be reduced, and the debugging and development period is shortened, so that the cost of the network card is reduced.

The invention relates to a mercuric chloride artificial intelligence all-in-one machine and an AI large model training method thereof, and relates to the technical field of artificial intelligence, the mercuric chloride artificial intelligence all-in-one machine comprises a plurality of high-performance calculation units based on mercuric chloride chips, each calculation unit is configured with an HBM2E memory, a multi-layer storage structure is realized by adopting a 3D stacking technology, and an ECC error correction function is integrated. The data storage module is used for storing and processing a large-scale AI model training data set; and the high-speed internet module is used for constructing a communication link between the computing units based on an InfiniBand protocol. The method has the advantages that by adopting the high-performance computing unit based on the mercuric chloride chip and the large memory technology, the memory bottleneck problem of a traditional GPU cluster during large-scale AI model training processing is effectively solved, and the model training speed and efficiency are remarkably improved; the application of the energy efficiency optimization system comprises dynamic voltage frequency adjustment and a double-circulation liquid cooling heat dissipation technology, the energy consumption is greatly reduced, and the operation cost and environment pressure of the data center are reduced.

Approaches presented herein provide for the real-time backhauling of data streams direct to memory. In at least one embodiment, a host device can initiate a Remote Direct Memory Access (RDMA) transmission with a remote device for sensor data streams. Initialization data for the RDMA transmission can be sent from the host to the remote device, such as over an existing physical Ethernet or InfiniBand (IB) connection. The host device can enter a Ready-to-Receive (RTR) state and send instructions to the remote device to enter a Ready-to-Send (RTS) state. The RDMA transmission can be initialized between the host device and the remote device to transfer streaming data. Using a remote data manager, the streaming data can be sent over the existing Ethernet connection as a RDMA transmission and written directly to allocated memory. The system may be scaled to include higher throughput sensors, additional sensors, and multiple remote transmission devices.

The invention provides a loading command processing method and system based on a remote direct memory access protocol. The system converts a plurality of graphics processor loading commands into a plurality of source end read commands at a source end, combines the source end read commands in the same queue pair and transmits source end write data to a target end buffer, then decomposes the source end write data at a target end, generates a plurality of read commands and reads back data, and combines the read back data into write data. A target end work queue item is generated after processing, then the work queue item is read, writing data is transmitted to the source end, and the source end receives the target end writing data, then decomposes the target end writing data into multiple pieces of reading data and transmits the multiple pieces of reading data to the source end graphics processor. According to the method and the system provided by the invention, a plurality of merged read commands or data can be transmitted on the Ethernet in a writing manner while an infinite bandwidth protocol is met, so that the bandwidth utilization rate can be remarkably improved, the resource occupation of a sending queue is reduced, and the orderliness of system operation is kept.

Systems and methods for using InfiniBand routing algorithms for Ethernet fabrics in a high performance computing environment. The method can provide, at a computer comprising one or more microprocessors, a plurality of switches, a plurality of hosts, a topology provider (TP) module, a routing engine (RE) module, and a switch initializer (SI) module. The method can perform a discovery sweep, by the TP, of the plurality of hosts and the plurality of switches and assigns an address to each of the plurality of hosts and the plurality of switches. The method can calculate, by the routing engine, a routing map, based upon a routing scheme, for the plurality of hosts and the plurality of switches, the routing map comprising a plurality of forwarding tables. The method can configure, each of the plurality of switches with a forwarding table of the plurality of forwarding tables calculated by the routing engine.

A method of routing for direct interconnect networks which generates multiple edge-disjoint paths between all source and destination nodes, while providing deadlock-free routing. The method requires limited use of virtual layers, which makes it suitable for less complex and more energy efficient hardware devices, and inter-layer edge-disjointedness is preferably integrated, which improves the throughput capacity of multipath traffic patterns. The method allows for distributed route calculation, wherein each node is responsible for generation of its own routes without the need for a centralized algorithm controller or elected master (with global knowledge), as it is, for instance, with InfiniBand. The method also involves a method of mapping arbitrary topologies from any dimension to bi-dimensional coordinate system. In addition, the method employs a lower number of turn restrictions than other prior solutions because it uses six direction types instead of only four as used in Up* / Down*. This increase in the number of directions translates into a more diversified set of routes, and consequently, better routing solutions.

This disclosure is directed towards systems and methods that improve bandwidth shifting operations of an ADPLL without losing a lock of the ADPLL and having the benefit of being able to change the bandwidth an unlimited amount of times. Indeed, a processor may transmit amplification parameters to the ADPLL to implement a bandwidth shift. The shift may occur in response to a enable signal, such as a gear trigger control signal (gear_retime signal) or a enable signal generated to cause alignment of the shifting with a clock signal (e.g., enable signal generated by AND logic gates). These systems and methods described herein many enable multiple bandwidth changing operations to occur without compromising the complexity and footprint of the system.

Approaches presented herein provide for the real-time backhauling of data streams direct to memory. In at least one embodiment, a host device can initiate a Remote Direct Memory Access (RDMA) transmission with a remote device for sensor data streams. Initialization data for the RDMA transmission can be sent from the host to the remote device, such as over an existing physical Ethernet or InfiniBand (IB) connection. The host device can enter a Ready-to-Receive (RTR) state and send instructions to the remote device to enter a Ready-to-Send (RTS) state. The RDMA transmission can be initialized between the host device and the remote device to transfer streaming data. Using a remote data manager, the streaming data can be sent over the existing Ethernet connection as a RDMA transmission and written directly to allocated memory. The system may be scaled to include higher throughput sensors, additional sensors, and multiple remote transmission devices.

The present invention provides a load command processing method and system based on the Remote Direct Memory Access (RDMA) protocol. The system converts multiple graphics processor load commands into multiple source-side read commands at the source end, merges source-side read commands in the same queue pair, and transmits source-side write data to a target-side buffer. The target end then decomposes the source-side write data to generate multiple read commands and read back the data. The read-back data is merged into write data, and after processing, a target-side work queue item is generated. The work queue item is then read and the write data is transmitted to the source end. After receiving the target-side write data, the source end decomposes it into multiple read data and transmits them to the source-side graphics processor. The method and system provided by the present invention can transmit multiple merged read commands or data in a write manner over Ethernet while meeting the InfiniBand protocol, significantly improving bandwidth utilization, reducing the resource usage of the send queue, and maintaining the orderliness of system operations.

The invention relates to a computing node processing method and device, equipment and a storage medium. The method comprises the following steps: under the condition that a network detection event aiming at a computing node is monitored, performing single-machine-dimension network detection and online-dimension infinite bandwidth IB network detection on the computing node, and under the condition that the single-machine-dimension network detection is passed and the online-dimension IB network detection is not passed, according to configuration information of the computing node, determining whether the computing node passes the network detection according to the configuration information of the computing node; and single-machine operation detection is performed on the computing nodes, online operation parameters of the computing nodes are optimized, and online processing is performed on the computing nodes after the single-machine operation test is passed. By adopting the method, the online efficiency of the computational nodes can be improved.

This invention relates to the field of artificial intelligence security technology, specifically disclosing an AI security flaw detection system based on deep analysis and adversarial testing. The system includes a data preprocessing module, a deep analysis module, an adversarial testing module, a result evaluation module, a report generation module, and a central control module. The data preprocessing module communicates with the deep analysis module via a high-speed data bus. The deep analysis module communicates bidirectionally with the adversarial testing module via dual gigabit Ethernet. The adversarial testing module is connected to the result evaluation module via an InfiniBand network. The result evaluation module communicates with the report generation module via an HTTPS secure connection. The central control module is connected to each module via a CANFD control bus and a TSN time-sensitive data bus. Therefore, this invention, through the collaborative approach of deep analysis and targeted adversarial testing, can improve the accuracy of detecting security flaws in AI models and reduce the false negative rate. It is applicable to fields such as finance and healthcare, providing strong assurance for the reliable operation of AI systems.

The invention relates to network transmission of destination-initiated data streams. The method presented herein provides real-time backhaul of a data stream directly to memory. In at least one embodiment, a host device may initiate a remote direct memory access (RDMA) transfer of a sensor data stream with a remote device. Initialization data of the RDMA transmission may be sent from a host device to a remote device, such as over an existing physical Ethernet or InfiniBand (IB) connection. A host device may enter a'ready to receive (RTR) 'state and send an instruction to a remote device to enter a'ready to send (RTS)' state. RDMA transmissions may be initialized between a host device and a remote device to transport streaming data. Using a remote data manager, streaming data can be sent in the form of RDMA transmissions over existing Ethernet connections and written directly to allocated memories. The system may be extended to include a higher throughput sensor, additional sensors, and a plurality of remote transmission devices.