Abstract: The present disclosure provides a prediction method and an apparatus for a faulty GPU, an electronic device and a storage medium. The method includes: acquiring parameter information of each GPU in a plurality of GPUs to obtain a parameter information set; inputting the parameter information set into a plurality of pre-trained prediction models to obtain a prediction result corresponding to each prediction model; and determining a faulty GPU from the plurality of GPUs according to the prediction result.

Abstract: The disclosure relates to a method, apparatus, system, medium and electronic device for graph neural network generation. The method includes: obtaining a subgraph structure, the subgraph structure being configured to reflect a graph structure of a corresponding subgraph, and the subgraph comprising a plurality of nodes and edges between the plurality of nodes; obtaining, based on the subgraph structure and according to a predetermined priority, node features of the plurality of nodes and edge features of the edges from a plurality of memories; the predetermined priority being obtained by sorting the plurality of memories in accordance with memory size in an ascending order; fusing, based on the subgraph structure, the node features of the plurality of nodes and the edge features of the edges to obtain subgraph data; and training, based on the subgraph data, the graph neural network.

Abstract: Embodiments of the disclosure provide a method and apparatus for running a service, and an electronic device. An embodiment of the method includes: determining a target deployment manner of a graphics processing unit (GPU) according to performance data of each service in a service set, where the deployment manner includes: dividing the GPU into sub-GPUs of a respective size and determining a service configured to be run by each sub-GPU; and switching, for the service in the service set, running of the service from a sub-GPU indicated by a current deployment manner to a sub-GPU indicated by the target deployment manner. According to the embodiment, waste of the GPU can be reduced by running a plurality of services on the GPU.

Abstract: The present disclosure provides a prediction method and an apparatus for a faulty GPU, an electronic device and a storage medium. The method includes: acquiring parameter information of each GPU in a plurality of GPUs to obtain a parameter information set; inputting the parameter information set into a plurality of pre-trained prediction models to obtain a prediction result corresponding to each prediction model; and determining a faulty GPU from the plurality of GPUs according to the prediction result.

Abstract: Embodiments of the disclosure provide a method and apparatus for running a service, and an electronic device. An embodiment of the method includes: determining a target deployment manner of a graphics processing unit (GPU) according to performance data of each service in a service set, where the deployment manner includes: dividing the GPU into sub-GPUs of a respective size and determining a service configured to be run by each sub-GPU; and switching, for the service in the service set, running of the service from a sub-GPU indicated by a current deployment manner to a sub-GPU indicated by the target deployment manner. According to the embodiment, waste of the GPU can be reduced by running a plurality of services on the GPU.

Abstract: Techniques are described for balancing traffic load for networks configured in multi-rooted tree topologies, in the presence of link failures. Maximum flows (through minimum cuts) are calculated for subgraphs that incorporate effective link capacities on links between source and destination nodes. Effective link capacities may be determined that take into account link failures, as well as sharing of current available link capacities by multiple nodes. Traffic is balanced while simultaneously fully utilizing available link capacities, even available link capacities on partially failed links (e.g., partially failed Link Aggregation Groups (LAGs)).

Abstract: In some cases, a network monitoring system may determine an operating or health condition of a node or connection link in a network (e.g., a datacenter network) by preparing an encapsulated data packet according to a tunneling protocol. Depending on a result of routing the encapsulated data packet, the network monitoring system determines whether the node or connection link is functioning normally or is experiencing an issue such as overloading or malfunctioning.

Abstract: Techniques are described for balancing traffic load for networks configured in multi-rooted tree topologies, in the presence of link failures. Maximum flows (through minimum cuts) are calculated for subgraphs that incorporate effective link capacities on links between source and destination nodes. Effective link capacities may be determined that take into account link failures, as well as sharing of current available link capacities by multiple nodes. Traffic is balanced while simultaneously fully utilizing available link capacities, even available link capacities on partially failed links (e.g., partially failed Link Aggregation Groups (LAGs)).

Abstract: Techniques are described for balancing traffic load for networks configured in multi-rooted tree topologies, in the presence of link failures. Maximum flows (through minimum cuts) are calculated for subgraphs that incorporate effective link capacities on links between source and destination nodes. Effective link capacities may be determined that take into account link failures, as well as sharing of current available link capacities by multiple nodes. Traffic is balanced while simultaneously fully utilizing available link capacities, even available link capacities on partially failed links (e.g., partially failed Link Aggregation Groups (LAGs)).

Abstract: Disclosed are systems and methods for network architecture that is a server-centric network architectural design.

Abstract: Securely storing assets in a cloud computer storage service. Preparation to store assets for a user may comprise determining a location to store the assets, generating a write access signature, and sending the determined location and the signature to the user. A request is received from the user to store the assets. Such a request includes the assets, the location, and the signature. In response to receiving the request, a determination is made regarding the write access signature whether the request should be honored. When the request is honored, the assets are stored in the determined location and the write access signature is invalidated. Upon subsequent user requests to access the assets, short-term read access signatures are generated and provided to the user. The assets are written once per generated write access signature and reads are to be performed close in time to when short-term read access signatures are requested.

Abstract: Disclosed are systems and methods for network architecture that is a server-centric network architectural design.

Abstract: This application describes routing packets from a source server to a plurality of ports of a switch. The switch is programmed by the control server and is used to direct incoming data packets to one or more ports of the switch in a manner that reduces congestion of incoming data packets to a destination server. Further, the control server queries congestion information from the switch, and then sends congestion notification back to the source server to either increase or decrease the amount of data being sent to the destination server.

Abstract: A virtual data center allocation architecture with bandwidth guarantees that provides for the creation of multiple virtual data centers from a single physical infrastructure. The virtual data center allocation is accomplished in three steps. First, clusters are created from the servers in the physical infrastructure. Second, a bipartite graph is built to map the virtual machines to the servers located in a particular cluster and finally a path is calculated between two virtual machines. The virtual data centers may be dynamically expanded or contracted based on changing bandwidth guarantees.

Abstract: Disclosed are systems and methods for network architecture that is a server-centric network architectural design.

Abstract: Disclosed are systems and methods for network architecture that is a server-centric network architectural design.

Abstract: Techniques are described for balancing traffic load for networks configured in multi-rooted tree topologies, in the presence of link failures. Maximum flows (through minimum cuts) are calculated for subgraphs that incorporate effective link capacities on links between source and destination nodes. Effective link capacities may be determined that take into account link failures, as well as sharing of current available link capacities by multiple nodes. Traffic is balanced while simultaneously fully utilizing available link capacities, even available link capacities on partially failed links (e.g., partially failed Link Aggregation Groups (LAGs)).

Abstract: Disclosed are systems and methods for network architecture that is a server-centric network architectural design.

Abstract: A system, method or computer readable medium to provide efficient congestion notification is described herein. In various embodiments, a packet is received at an intermediate node of one or more data center networks. A current queue length at the intermediate node is determined. A threshold value for the current queue length is tuned by dynamically computing an upper bound and a lower bound based at least in part on the network. The packet is marked to indicate possible congestion in the one or more data center networks when the current queue length exceeds the threshold value. In some embodiments, the packet is marked when it is being de-queued. In a further embodiment, Flexible Initial Packet Size (FIPS) may be utilized to improve the efficiency of the tuning.

Abstract: In some cases, a network monitoring system may determine an operating or health condition of a node or connection link in a network (e.g., a datacenter network) by preparing an encapsulated data packet according to a tunneling protocol. Depending on a result of routing the encapsulated data packet, the network monitoring system determines whether the node or connection link is functioning normally or is experiencing an issue such as overloading or malfunctioning.