Nicira, Inc.

Hewlett Packard Enterprise Development LP

Hewlett-packard Development Company, L.P.

Apple Inc.

Vmware, Inc.

Mohan Parthasarathy

Jayant Jain

Anirban Sengupta

Xinhua Hong

Rick Lund

Serge Maskalik

Allwyn M Sequeira

Soumitra Chatterjee

Brijesh Nambiar

Prasad Palkar

Chinmay Ghosh

Sunil Vishwanathpur Lakshminarasimha

Raju Koganty

Ramsundar Janakiraman

Mashood Abdulla Kodavanji

Shankar Subramaniam

Steven Alexander

Giri Gopalan

Stuart David Cheshire

Min-Yi Shen

Sriram Narasimhan

Suhas Shetty

Viswesh Ananthakrishnan

Terrin D Eager

Mashood Abdulla K

Anjali Anant Kanak

Chandrashekhara Anandamurthy

Jayalakshmi P

Subbarao V R N

Unum Sarfraz

A novel method for stateful packet classification that uses hardware resources for performing stateless lookups and software resources for performing stateful connection flow handshaking is provided. To classify an incoming packet from a network, some embodiments perform stateless look up operations for the incoming packet in hardware and forward the result of the stateless look up to the software. The software in turn uses the result of the stateless look up to perform the stateful connection flow handshaking and to determine the result of the stateful packet classification.

Use of stateless marking to speed up stateful firewall rule processing

Some embodiments provide a novel method for load balancing data messages that are sent by a source compute node (SCN) to one or more different groups of destination compute nodes (DCNs). In some embodiments, the method deploys a load balancer in the source compute node's egress datapath. This load balancer receives each data message sent from the source compute node, and determines whether the data message is addressed to one of the DCN groups for which the load balancer spreads the data traffic to balance the load across (e.g., data traffic directed to) the DCNs in the group. When the received data message is not addressed to one of the load balanced DCN groups, the load balancer forwards the received data message to its addressed destination. On the other hand, when the received data message is addressed to one of load balancer's DCN groups, the load balancer identifies a DCN in the addressed DCN group that should receive the data message, and directs the data message to the identified DCN. To direct the data message to the identified DCN, the load balancer in some embodiments changes the destination address (e.g., the destination IP address, destination port, destination MAC address, etc.) in the data message from the address of the identified DCN group to the address (e.g., the destination IP address) of the identified DCN.

Inline load balancing

A novel algorithm for packet classification that is based on a novel search structure for packet classification rules is provided. Addresses from all the containers are merged and maintained in a single Trie. Each entry in the Trie has additional information that can be traced back to the container from where the address originated. This information is used to keep the Trie in sync with the containers when the container definition dynamically changes.

Efficient packet classification for dynamic containers

A method is provided that uses a consistent hashing technique to dispatch incoming packets in a stable system prior to adding of a node. The method uses a hash table and assigns hash buckets in the table to each network node. A set of fields in each incoming packet is hashed and is used to identify the corresponding hash bucket. The packets are then dispatched to the network nodes based on the nodes' hash buckets. During an observation period, the method identifies the ongoing sessions by creating a bit vector table that is used to identify the old and new sessions during a re-dispatching period. The method uses the consistent hashing method and the probabilistic method dispatch the incoming packets such that each packet that belongs to an old session is dispatched to the same old node that has been processing the other packets of the session.

Consistent hashing for network traffic dispatching

Some embodiments provide an elastic architecture for providing a service in a computing system. To perform a service on the data messages, the service architecture uses a service node (SN) group that includes one primary service node (PSN) and zero or more secondary service nodes (SSNs). The service can be performed on a data message by either the PSN or one of the SSN. However, in addition to performing the service, the PSN also performs a load balancing operation that assesses the load on each service node (i.e., on the PSN or each SSN), and based on this assessment, has the data messages distributed to the service node(s) in its SN group. Based on the assessed load, the PSN in some embodiments also has one or more SSNs added to or removed from its SN group. To add or remove an SSN to or from the service node group, the PSN in some embodiments directs a set of controllers to add (e.g., instantiate or allocate) or remove the SSN to or from the SN group. Also, to assess the load on the service nodes, the PSN in some embodiments receives message load data from the controller set, which collects such data from each service node. In other embodiments, the PSN receives such load data directly from the SSNs.

Method and apparatus for providing a service with a plurality of service nodes

In an example, a neural network program corresponding to a neural network model is received. The neural network program includes matrices, vectors, and matrix-vector multiplication (MVM) operations. A computation graph corresponding to the neural network model is generated. The computation graph includes a plurality of nodes, each node representing a MVM operation, a matrix, or a vector. Further, a class model corresponding to the neural network model is populated with a data structure pointing to the computation graph. The computation graph is traversed based on the class model. Based on the traversal, the plurality of MVM operations are assigned to MVM units of a neural network accelerator. Each MVM unit can perform a MVM operation. Based on assignment of the plurality of MVM operations, an executable file is generated for execution by the neural network accelerator.

Generation of executable files corresponding to neural network models

Systems and methods are provided for managing false positives in a network anomaly detection system. The methods may include receiving a plurality of anomaly reports; extracting fields, and values for the fields, from each of the anomaly reports; grouping the anomaly reports into a plurality of groups according to association rule learning, wherein each group is defined by a respective rule; for each group, creating a cluster based on common values for the fields; and marking each cluster as a possible false positive anomaly cluster.

Managing false positives in a network anomaly detection system

Disclosed is a method, system, and computer readable medium to manage (and possibly replace) cycles in graphs for a computer device. The method includes detecting a compound operation including a first tensor, the compound operation resulting from source code represented in a first graph structure as part of a compilation process from source code to binary executable code. To address a detected cycle, an instance of a proxy class may be created to store a pointer to a proxy instance of the first tensor based on the detection. In some examples, using the instance of the proxy class facilitates implementation of a level of indirection to replace a cyclical portion of the graph structure with an acyclical portion such that the second graph structure indicates assignment of a result of the compound operation to the proxy instance of the first tensor. Optimization may reduce a total number of indirection replacements.

Avoiding cycles in neural networks

Example implementations relate to assigning dependent matrix-vector multiplication (MVM) operations to consecutive crossbars of a dot product engine (DPE). A method can comprise grouping a first MVM operation of a computation graph with a second MVM operation of the computation graph where the first MVM operation is dependent on a result of the second MVM operation, assigning a first crossbar of a DPE to an operand of the first MVM operation, and assigning a second crossbar of the DPE to an operand of the second MVM operation, wherein the first and second crossbars are consecutive.

Assigning dependent matrix-vector multiplication operations to consecutive crossbars of a dot product engine

Some embodiments provide novel inline switches that distribute data messages from source compute nodes (SCNs) to different groups of destination service compute nodes (DSCNs). In some embodiments, the inline switches are deployed in the source compute nodes datapaths (e.g., egress datapath). The inline switches in some embodiments are service switches that (1) receive data messages from the SCNs, (2) identify service nodes in a service-node cluster for processing the data messages based on service policies that the switches implement, and (3) use tunnels to send the received data messages to their identified service nodes. Alternatively, or conjunctively, the inline service switches of some embodiments (1) identify service-nodes cluster for processing the data messages based on service policies that the switches implement, and (2) use tunnels to send the received data messages to the identified service-node clusters. The service-node clusters can perform the same service or can perform different services in some embodiments. This tunnel-based approach for distributing data messages to service nodes/clusters is advantageous for seamlessly implementing in a datacenter a cloud-based XaaS model (where XaaS stands for X as a service, and X stands for anything), in which any number of services are provided by service providers in the cloud.

Tunnel-enabled elastic service model

Repeating patterns are identified in a matrix. Based on the identification of the repeating patterns, instructions are generated, which are executable by processing cores of a dot product engine to allocate analog multiplication crossbars of the dot product engine to perform multiplication of the matrix with a vector.

Crossbar allocation for matrix-vector multiplications

According to examples, an apparatus may include a processor and a non-transitory computer readable medium having instructions that when executed by the processor, may cause the processor to partition a matrix of elements into a plurality of sub-matrices of elements. Each sub-matrix of the plurality of sub-matrices may include elements from a set of columns of the matrix of elements that includes a nonzero element. The processor may also assign elements of the plurality of sub-matrices to a plurality of crossbar devices to maximize a number of nonzero elements of the matrix of elements assigned to the crossbar devices.

Growing community of inventors

Cupertino, CA, United States of America

Mohan Parthasarathy