Neural network specific hardware acceleration optimizations are disclosed, including an optimized multicast network and an optimized DRAM transfer unit to perform in constant or linear time. The multicast network is a set of switch nodes organized into layers and configured to operate as a Beneš network. Configuration data may be accessed by all switch nodes in the network. Each layer is configured to perform a Beneš network transformation of the -previous layer within a computer instruction. Since the computer instructions are pipelined, the entire network of switch nodes may be configured in constant or linear time. Similarly a DRAM transfer unit configured to access memory in strides organizes memory into banks indexed by prime or relatively prime number amounts. The index value is selected as not to cause memory address collisions. Upon receiving a memory specification, the DRAM transfer unit may calculate out strides thereby accessing an entire tile of a tensor in constant or linear time.

Some examples provide a method for automatic network assembly. The following instructions may be used to implement automatic network assembly in a modular infrastructure. Instructions to automatically connect a management port to a management network. Instructions to automatically connect link ports to form a scalable ring. Instructions to automatically connect each modular infrastructure management device to a bay management network port.

[Problem] Connection between a centralized control apparatus and a group of transfer apparatuses can be prevented from having a single point of failure. Traffic can be distributed among a plurality of paths. A bypass path is selected when a failure occurs in a switch cluster.

[Solution] Transfer apparatuses 61a to 61d perform communications for path control with a centralized control apparatus 73 that performs centralized control from the outside of a switch cluster including the group of transfer apparatuses, through a path similar to D-plane (main signal). A packet flow controller 87 serving as a separation unit that separates a packet for the inside of the cluster 61 and a packet for the outside of the cluster transmitted through the similar path from each other, and an internal route engine 85 that performs path control of obtaining a path for freely passing through a plurality of paths in the cluster are provided. The packet flow controller 87 separates a path control packet for the inside of the cluster, and the engine 85 performs, when a failure to communicate the path control packet for the inside thus separated occurs, path control of generating a path that bypasses a path with the failure.

Provided is a control apparatus that controls any one or all of a plurality of slave station apparatuses communicating with a terminal apparatus, a plurality of master station apparatuses that control the slave station apparatuses, and a transfer apparatus that transfers data transmitted and received between the master station apparatuses and the slave station apparatuses, the control apparatus including an information acquisition unit that acquires information regarding traffic of the data transmitted and received between the master station apparatuses and the slave station apparatuses, and a switching control unit that performs, on the basis of the information regarding the traffic acquired by the information acquisition unit, switching-control of an assignment relationship between the master station apparatus and the slave station apparatus and switching-control of a transfer path of data between the master station apparatus and the slave station apparatus.

Techniques and mechanisms for interconnecting network circuitry of an integrated circuit (IC) die and physical layer (PHY) circuits of the same IC die. In an embodiment, nodes of the network circuitry include first routers and processor cores, where the first routers are coupled to one another in an array configuration which includes rows and columns. First interconnects each extend to couple both to a corresponding one of the PHY circuits and to a corresponding one of the first routers. For each of one or more of the first interconnects, a respective one or more rows (or one or more columns) of the array configuration extend between the corresponding PHY and the corresponding router. In another embodiment, the network circuitry comprises network clusters which each include a different respective row of the array configuration.

In general, the present invention relates to a virtual platform in which one or more distributed virtual switches can be created for use in virtual networking. According to some aspects, the distributed virtual switch according to the invention provides the ability for virtual and physical machines to more readily, securely, and efficiently communicate with each other even if they are not located on the same physical host and/or in the same subnet or VLAN. According other aspects, the distributed virtual switches of the invention can support integration with traditional IP networks and support sophisticated IP technologies including NAT functionality, stateful firewalling, and notifying the IP network of workload migration. According to further aspects, the virtual platform of the invention creates one or more distributed virtual switches which may be allocated to a tenant, application, or other entity requiring isolation and/or independent configuration state. According to still further aspects, the virtual platform of the invention manages and/or uses VLAN or tunnels (e.g, GRE) to create a distributed virtual switch for a network while working with existing switches and routers in the network. The present invention finds utility in both enterprise networks, datacenters and other facilities.

Methods and systems for wireless packet switching include determining a schedule for transceivers in an enclosure. The schedule specifies which of the transceivers will act as a transmitter and which will act as a receiver. A beamforming direction for transmitting data from each transmitter to each corresponding receiver is determined. It is determined that an angle of the beamforming direction for at least one transmitter is lower than a minimum angle. Data is transmitted from a transmitter to the corresponding receiver by a wired connection, responsive to the determination that the angle of the beamforming direction is lower than a minimum angle.

A storage system is provided. The storage system includes a plurality of storage units, each of the plurality of storage units having storage memory for user data and a plurality of storage nodes, each of the plurality of storage nodes configured to have ownership of a portion of the user data. The storage system includes a first pathway, coupling the plurality of storage units such that each of the plurality of storage units can communicate with at least one other of the plurality of storage units via the first pathway without assistance from the plurality of storage nodes.

System and method for supporting node role attributes in a high performance computing environment. In accordance with an embodiment, a node role attribute can comprise a vendor defined subnet management attribute. When a subnet manager attempts to discover a high performance computing environment, such as an InfiniBand subnet, or a switch topology, identifying a topology is quite complex when subnet manager can only observe connectivity, without context behind the connectivity (the roles of the different nodes in the connectivity). However, when a subnet has a node role attribute enabled, the subnet manager can map the interconnect more effectively as it can discover not only the connectivity during the initial sweep, but it can also discover the role of each node discovered, thus leading to a more efficient interconnect discovery.

Methods to strengthen the cyber-security and privacy in a proposed deterministic Internet of Things (IoT) network are described. The proposed deterministic IoT consists of a network of simple deterministic packet switches under the control of a low-complexity ‘Software Defined Networking’ (SDN) control-plane. The network can transport ‘Deterministic Traffic Flows’ (DTFs), where each DTF has a source node, a destination node, a fixed path through the network, and a deterministic or guaranteed rate of transmission. The SDN control-plane can configure millions of distinct interference-free ‘Deterministic Virtual Networks’ (DVNs) into the IoT, where each DVN is a collection of interference-free DTFs. The SDN control-plane can configure each deterministic packet switch to store several deterministic periodic schedules, defined for a scheduling-frame which comprises F time-slots. The schedules of a network determine which DTFs are authorized to transmit data over each fiber-optic link of the network. These schedules also ensure that each DTF will receive a deterministic rate of transmission through every switch it traverses, with full immunity to congestion, interference and Denial-of-Service (DoS) attacks. Any unauthorized transmissions by a cyber-attacker can also be detected quickly, since the schedules also identify unauthorized transmissions. Each source node and destination node of a DTF, and optionally each switch in the network, can have a low-complexity private-key encryption/decryption unit. The SDN control-plane can configure the source and destination nodes of a DTF, and optionally the switches in the network, to encrypt and decrypt the packets of a DTF using these low-complexity encryption/decryption units. To strengthen security and privacy and to lower the energy use, the private keys can be very large, for example several thousands of bits. The SDN control-plane can configure each DTF to achieve a desired level of security well beyond what is possible with existing schemes such as AES, by using very long keys. The encryption/decryption units also use a new serial permutation unit the very low hardware cost, which allows for exceptional security and very-high throughputs in FPGA hardware.