Nvidia Corporation

Benjamin Klenk

Larry Robert Dennison

Nan Jiang

Gregory Michael Thorson

Alan Lynn Davis

Aggregation of small payloads from multiple packets may improve bandwidth efficiency of a network, particularly a high-performance compute cluster with thousands of network endpoints and distributed data. Aggregation is context-based and a packet header is reduced because the common components that are shared by the aggregated messages are included once within the header. Execution contexts are explicitly created and destroyed by application programs. Each participating endpoint stores context-specific properties until the context is destroyed, so that the properties are not included in the header. Aggregation may be performed at different hierarchical levels by switches and/or endpoints.

In-network message aggregation for efficient small message transport

A network device configured to perform scalable, in-network computations is described. The network device is configured to process pull requests and/or push requests from a plurality of endpoints connected to the network. A collective communication primitive from a particular endpoint can be received at a network device. The collective communication primitive is associated with a multicast region of a shared global address space and is mapped to a plurality of participating endpoints. The network device is configured to perform an in-network computation based on information received from the participating endpoints before forwarding a response to the collective communication primitive back to one or more of the participating endpoints. An injection policy comprising the issuing of credits enables each endpoint to limit the amount of collective communication primitives injected into the network simultaneously to reduce network congestion caused by increased network traffic due to the multicast capability of the network devices.

Injection limiting and wave synchronization for scalable in-network computation

A network device configured to perform scalable, in-network computations is described. The network device is configured to process pull requests and/or push requests from a plurality of endpoints connected to the network. A collective communication primitive from a particular endpoint can be received at a network device. The collective communication primitive is associated with a multicast region of a shared global address space and is mapped to a plurality of participating endpoints. The network device is configured to perform an in-network computation based on information received from the participating endpoints before forwarding a response to the collective communication primitive back to one or more of the participating endpoints. The endpoints can inject pull requests (e.g., load commands) and/or push requests (e.g., store commands) into the network. A multicast capability enables tasks, such as a reduction operation, to be offloaded to hardware in the network device.

Scalable in-network computation for massively-parallel shared-memory processors

A technique for performing data parallel training of a neural network model is disclosed that incorporates batch normalization techniques using partial populations to generate normalization parameters. The technique involves processing, by each processor of a plurality of processors in parallel, a first portion of a sub-batch of training samples allocated to the processor to generate activations for the first portion of the sub-batch. Each processor analyzes the activations and transmits statistical measures for the first portion to an additional processor that reduces the statistical measures from multiple processors to generate normalization parameters for a partial population of the training samples that includes the first portion from each of the plurality of processors. The normalization parameters are then transmitted back to each of the processors to normalize the activations for both the first portion and a second portion of the sub-batch of training samples allocated to each processor.

Distributed batch normalization using partial populations

A technique utilizing speculative execution and rollback for performing data parallel training of a neural network model is disclosed. Activations for a layer of the neural network model are normalized during a speculative normalization operation using estimated normalization parameters associated with a partial population of a set of training data allocated to a particular processor. Normalization parameters associated with the total population of the set of training data are generated by a distributed reduce operation in parallel with the speculative normalization operation. An optional rollback operation can revert the activations to a pre-normalization state if the estimated normalization parameters for the partial population are subsequently determined to be inaccurate compared to the normalization parameters for the population of the set of training data distributed across a plurality of processors.

Growing community of inventors

San Jose, CA, United States of America

Benjamin Klenk