Mellanox Technologies, Ltd.

Samsung Electronics Co., Ltd.

Mellanox Technologies Tlv Ltd.

Omer Shabtai

Matty Kadosh

Yuval Shpigelman

Yamin Friedman

Alan Lo

Marian Pritsak

Yonatan Piasetzky

Rotem Levinson

Yotam Platner

Gil Levy

Liron Mula

Aviv Kfir

Idan Burstein

Nathan Henri Levy

Ishay Goldin

Yonatan Piasetsky

Ran Vitek

Alexandra Dana

Khalid Manaa

Netanel Stein

Charlie Mbariky

Or Davidi

Tal Heller

Alon Intrater

Yaron Ukrainitz

David Tsidkiahu

Otniel Van Handel

Zuf Pilosof

Technologies for optimizing the spreading of traffic across multiple local output ports while considering both local load and end-to-end (E2E) load are described. One device has multiple outgoing ports and a network adapter that determines, for a first flow of packets, a first end-to-end (E2E) congestion rate of at least some of the outgoing ports. The network adapter determines a port state of at least some of the outgoing ports. The network adapter receives a first packet associated with the first flow of packets. The network adapter determines, using a first desired rate for the first flow, the first E2E congestion rates, and the port states, i) a first time at which the first packet is to be transmitted and ii) a first outgoing port on which the first packet is to be transmitted. The first packet is sent on the first outgoing port at the first time.

Combined congestion control and load balancing

Techniques and apparatuses enabling high accuracy video object detection using reduced system resource requirements (e.g., reduced computational load, shallower neural network designs, etc.) are described. For example, a search domain of an object detection scheme (e.g., a target object class, a target object size, a target object rotation angle, etc.) may be separated into subdomains (e.g., such as subdomains of object classes, subdomains of object sizes, subdomains object rotation angles, etc.). Specialized, subdomain-level object detection/segmentation tasks may then be separated across sequential video frames. As such, different subdomain-level processing techniques (e.g., via specialized neural networks) may be implemented across different frames of a video sequence. Moreover, redundancy information of consecutive video frames may be leveraged, such that specialized object detection tasks combined with visual object tracking across consecutive frames may enable more efficient (e.g., more accurate, less computationally intensive, etc.) full domain object detection and object segmentation schemes.

Fast object detection in video via scale separation

A networking device and system are described, among other things. An illustrative system is disclosed to include a congestion controller that manages traffic across a network fabric using receiver-based packet scheduling and a networking device that employs the congestion controller for data flows qualified as a large data flow but bypasses the congestion controller for data flows qualified as a small data flow. For example, the networking device may receive information describing a data flow directed toward a processing network; determine, based on the information describing the data flow, a size of the data flow; determine the size of the data flow is below a predetermined flow threshold; and in response to determining that the size of the data flow is below a predetermined threshold, bypass the congestion controller.

Dynamic fabric reaction for optimized collective communication

The described techniques provide for an automated process of hardware design using main applications (e.g., use cases) and a modeling heuristic of a compiler. Hardware may be optimized and designed based on the knowledge of the compiler for the hardware and firmware information. For instance, a user may define constraints (e.g., area constraints, power constraints, performance constraints, accuracy degradation constraints, etc.) and the changes of compilation heuristics and hardware parameters may be optimized to efficiently achieve the user defined constraints. Accordingly, hardware configuration parameters may be optimized based on the neural network's compilation process (e.g., actual compiler constraints) and optimization of power, performance, and area (PPA) constraints (e.g., user defined constraints). Specific neural processor (SNP) hardware may thus be designed based on the optimized hardware configuration parameters (e.g., via modeling heuristics of its compiler and main applications informed via user defined constraints or PPA constraints).

Hardware architecture determination based on a neural network and a network compilation process

A network adapter includes a port and one or more circuits. The port communicates packets over a network in which switches forward packets in accordance with tuples of the packets. The one or more circuits are to hold a user-programmable scheme specifying assignments of the packets of a given flow destined to a peer node to sub-flows having respective different tuples, assign first packets of the given flow to one or more of the sub-flows in accordance with the user-programmable scheme, by setting respective tuples of the first packets, transmit the first packets to the peer node via the port, monitor notifications received from the network, the notifications being indicative of respective states of the sub-flows, based on the notifications and on the user-programmable scheme determine an assignment of second packets of the given flow to the sub-flows, and transmit the second packets to the peer node via the port.

Flexible per-flow multipath managed by sender-side network adapter

Devices, methods, and systems are provided. In one example, a method is described to include measuring a cardinality of actual data flows at a flow-processing resource, determining that the cardinality of the actual data flows triggers a congestion control action, and, in response to determining that the cardinality of the actual data flows triggers the congestion control action, implementing the congestion control action with respect to the flow-processing resource.

Cardinality-based traffic control

A network device includes processing circuitry and a plurality of ports. The ports connect to a communication network. The processing circuitry is configured to receive, via an input port, data packets and probe packets that are addressed to a common output port, to store the data packets in a first queue and the probe packets in a second queue, both the first queue and the second queue are served by the output port, to produce telemetry data indicative of a state of the network device, based on a processing path that the data packets traverse within the network device, to schedule transmission of the data packets from the first queue at a first priority, and schedule transmission of the probe packets from the second queue at a second priority higher than the first priority, and to modify the scheduled probe packets so as to carry the telemetry data.

Low-latency delivery of in-band telemetry data

A method including providing: a switching device including a main mapping unit configured to provide a main mapping which maps virtual addresses to direct addresses; management logic configured to store a connection tracking table stored in memory and configured for storing a plurality of connection mappings each including a virtual-to-direct mapping from a virtual address to a direct address; and a migrated connection table stored in memory and configured for storing a plurality of migrated connection mappings each including a virtual-to-migrated-direct mapping from a virtual address to a migrated direct address. Upon receiving an indication that a direct address has changed: when the indication that a direct address has changed includes an indication that a direct address stored in one or more connection mappings in the connection tracking table is not active, copying the one or more connection mappings from the connection tracking table to the migrated connection table, when the indication that a direct address has changed includes an indication that a new direct address has become active, creating one or more connection mappings in the migrated connection table, and altering the main mapping in accordance with the direct address change.

Switching device with migrated connection table

Growing community of inventors

Tel Aviv, Israel

Omer Shabtai