Mellanox Technologies, Ltd.

Mellanox Technologies Tlv Ltd.

Richard Leigh Graham

Daniel Marcovitch

Gil Bloch

Noam Bloch

Lion Levi

Yossef Itigin

Michael Kagan

Ariel Shahar

Diego Crupnicoff

Eitan Zahavi

Roee Moyal

Roman Nudelman

Ortal Ben Moshe

Yong Qin

Yaniv Blumenfeld

Todd Wilde

Igor Voks

Ofer Hayut

Dotan David Levi

Liran Liss

Benny Koren

Shlomo Raikin

Larry Robert Dennison

Shahaf Shuler

Oded Wertheim

Peter Paneah

Eyal Srebro

Dror Bohrer

Aviad Levy

Eliel Peretz

Ana Gainaru

Sean Pieper

Disclosed are apparatuses, systems, and techniques that improve efficiency and decrease latency of remote direct memory access (RDMA) operations. The techniques include but are not limited to unified RDMA operations that are recognizable by various communicating devices, such as network controllers and target memory devices, as requests to establish, set, and/or update arrival indicators in the target memory devices responsive to arrival of one or more portions of the data being communicated.

Remote direct memory access operations with integrated data arrival indication

Devices, methods, and systems are provided. In one example, a device is described to include a device interface that receives data from at least one data source; a data shuffle unit that collects the data received from the at least one data source, receives a descriptor that describes a data shuffle operation to perform on the data received from the at least one data source, performs the data shuffle operation on the collected data to produce shuffled data, and provides the shuffled data to at least one data target.

Data shuffle offload

A network device includes a first interface, a second interface, and circuitry. The first interface is configured to communicate at least with a memory. The second interface is configured to communicate over a network with a peer network device. The circuitry is configured to receive a request to transfer data over the network between the memory and the peer network device in accordance with (i) a pattern of offsets to be accessed in the memory and (ii) a memory key representing a memory space to be accessed using the pattern, and to transfer the data in accordance with the request.

Patterned memory-network data transfer

A method includes providing a plurality of processes interconnected by a network, each of the plurality of processes being configured to hold a block of data destined for others of the plurality of processes. A set of data for all-to-all data exchange is received from one or more of the processes. The set of data is configured as a plurality of blocks of data in a matrix as matrix data, the matrix being distributed among the plurality of processes. The matrix data is transposed by changing the position of selected blocks of data of the plurality of blocks of data relative to the other blocks of data of the plurality of the blocks of data, without changing the structure of each of the blocks of data. The transposed matrix data is over the network and is then received, repacked, and conveyed to destination processes.

Collective communication system and methods

A Direct Memory Access (DMA) device includes an interface and a DMA engine. The interface is configured to communicate with a first memory and with a second memory. The DMA engine is configured to (i) receive a request to transfer data between the first memory and the second memory in accordance with a pattern of offsets to be accessed in the first memory or in the second memory, and (ii) transfer the data in accordance with the request.

Patterned direct memory access (DMA)

A network device includes a first interface, a second interface and circuitry. The first interface is configured to communicate at least with a first memory. The second interface is configured to communicate over a network with a peer network device coupled to a second memory. The circuitry is configured to (i) receive a request to transfer data over the network between the first memory and the second memory in accordance with a pattern of offsets to be accessed in the first memory or in the second memory, and (ii) transfer the data in accordance with the request.

Patterned remote direct memory access (RDMA)

A method for collective communications includes invoking a collective operation over a group of computing processes in which the processes concurrently transmit and receive data to and from other processes in the group via a communication medium. Messages are composed for transmission by source processes including metadata indicating how the data to be transmitted by the source processes in the collective operation are to be handled by destination processes that are to receive the data and also including in at least some of the messages the data to be transmitted by one or more of the source processes to one or more of the destination processes. The composed messages are transmitted concurrently from the source processes to the destination processes in the group over the communication medium. The data are processed by the destination processes in response to the metadata included in the messages received by the destination processes.

Single-step collective operations

A method in which a plurality of process are configured to hold a block of data destined for other processes, with data repacking circuitry including receiving circuitry configured to receive at least one block of data from a source process of the plurality of processes, the repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes, and sending circuitry configured to send the repacked data to the at least one destination process of the plurality of processes, receiving a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes, and transposing the data by each of the plurality of processes sending matrix data from the process to the repacking circuitry, and the repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes.

Systems, methods, and devices for performing computing operations are provided. In one example, a system is described to include an endpoint belonging to a collective that is organized as a hierarchical tree. The collective includes one or more application groups that are connected to leaf nodes of the hierarchical tree, where each application-level process in the one or more application groups initiate processing of data based on an order of arrival known for at least one other application-level process joining the one or more application groups.

Message passing interface (MPI) collectives using multi-allgather

A network device in a communication network includes a controller and processing circuitry. The controller is configured to manage execution of an operation whose execution depends on inputs from a group of one or more work-request initiators. The processing circuitry is configured to read one or more values, which are set by the work-request initiators in one or more memory locations that are accessible to the work-request initiators and to the network device, and to trigger execution of the operation in response to verifying that the one or more values read from the one or more memory locations indicate that the work-request initiators in the group have provided the respective inputs.

Memory-based synchronization of distributed operations

An apparatus includes one or more ports for connecting to a communication network, processing circuitry and a message aggregation circuit (MAC). The processing circuitry is configured to communicate messages over the communication network via the one or more ports. The MAC is configured to receive messages, which originate in one or more source processes and are destined to one or more destination processes, to aggregate two or more of the messages that share a common destination into an aggregated message, and to send the aggregated message using the processing circuitry over the communication network.

Small message aggregation

A method including providing a SHARP tree including a plurality of data receiving processes and at least one aggregation node, designating a data movement command, providing a plurality of data input vectors to each of the plurality of data receiving processes, respectively, the plurality of data receiving processes each passing on the respective received data input vector to the at least one aggregation node, and the at least one aggregation node carrying out the data movement command on the received plurality of data input vectors. Related apparatus and methods are also provided.

Growing community of inventors

Knoxville, TN, United States of America

Richard Leigh Graham