Amazon Technologies, Inc.

Marvell Israel (misl) Ltd.

Other

Marvell Israle (m.i.s.l) Ltd.

Erez Izenberg

Nafea Bshara

Leah Shalev

Guy Nakibly

Christopher Joseph Pettey

Said Bshara

Ron Diamant

Mark Bradley Davis

Asif Khan

Georgy Machulsky

Robert Michael Johnson

Dimitry Melts

Avigdor Segal

Erez Amit

Jonathan Cohen

Ofer Frishman

Islam Mohamed Hatem Abdulfattah Mohamed Atta

Ofer Naaman

Kypros Constantinides

Islam Atta

Adi Habusha

Yaniv Shapira

Noam Katz

Uri Leder

Ruven Torok

Alon Pais

Benzi Denkberg

Oren Shafrir

Yuval Caduri

Guy M Benhaim

Curtis Karl Ohrt

Matthew Shawn Wilson

Alan Michael Judge

Zorik Machulsky

Ori Weber

Brian William Barrett

Itay Poleg

Anna Rom-Saksonov

Joshua Benjamin Levinson

Adiel Sarusi

Julien Ridoux

Jonathan Chocron

Nitzan Zisman

Dmitri Epshtein

Noam Attias

Oren Ben-Hayune

Arie Elias

Shaul Ben-Mayor

Erez Bashan

Amiram Lifshitz

Anthony Nicholas Liguori

Leonid Baryudin

Erez Sabbag

Gil Stoler

Barak Wasserstrom

Robert Klein

Se Wang Oh

Noga Smith

Dan Saad

Roi Ben Haim

Nofar Mann

Shadi Ammouri

Rotem Shaanan

Dror Fleischmann

Ruvel Torok

Dimitri Epshtein

Apparatus and methods are disclosed herein for remote, direct memory access (RDMA) technology that enables direct memory access from one host computer memory to another host computer memory over a physical or virtual computer network. In one example, RDMA functionality is provided by combining a host executing instructions for RDMA software applications with dedicated hardware accelerator, where the host has access to substantially all RDMA connection context (e.g., configuration, status, state machine, counters, etc.) and a hardware accelerator stores a portion of the context information for a set of currently active transactions. The hardware accelerator derives a transaction identifier from header information in received RDMA packets and performs a local RDMA operation using at least a portion of the received data, a destination address of the location RDMA operation being based at least in part on a memory address determined using the transaction identifier.

Transaction based remote direct memory access

Apparatus and methods are disclosed herein for remote, direct memory access (RDMA) technology that enables direct memory access from one host computer memory to another host computer memory over a physical or virtual computer network according to a number of different RDMA protocols. In one example, a method includes receiving remote direct memory access (RDMA) packets via a network adapter, deriving a protocol index identifying an RDMA protocol used to encode data for an RDMA transaction associated with the RDMA packets, applying the protocol index to a generate RDMA commands from header information in at least one of the received RDMA packets, and performing an RDMA operation using the RDMA commands.

Flexible remote direct memory access

A wake-up management circuit for implementation in circuit boards with multiple processors is disclosed. The wake-up management circuit utilizes hardware resources to generate a wake-up signal for a processor in a low power mode when traffic intended for the processor is received by the wake-up management circuit. The wake-up management circuit is configured to generate a wake-up signal for a specific processor when address information in a data packet received by the wake-up management circuit indicates that the data packet is intended for the specific processor (e.g., the specific processor is the destination for the data packet).

Wake-up management circuit for multiple processors

Techniques to virtualize a physical hardware clock are described. The techniques may include utilizing a coefficients storage to store timestamp coefficient entries. When a timestamp request is received from a client, a selection circuit selects a timestamp coefficient entry from the coefficients storage. A compute circuit then computes a timestamp according to the client's virtualized clock based on the selected timestamp coefficient entry and a counter value of a counter that is driven by a physical clock signal.

Physical hardware clock virtualization

A resource manager of a virtualized computing service indicates to a client that FPGA-enabled compute instances are supported at the service. From a set of virtualization hosts of the service, a particular host from which an FPGA is accessible is selected for the client based on an indication of computation objectives of the client. Configuration operations are performed to prepare the host for the application, and an FPGA-enabled compute instance is launched at the host for the client.

FPGA-enabled compute instances

Techniques for low-latency packet processing are disclosed. A network device receives a first set of write transactions including a first set of data segments corresponding to a first DMA descriptor from a host. The network device receives a second set of write transactions including a second set of data segments corresponding to a second DMA descriptor from the host. The network device detects that the first set of data segments have been written. In response to detecting that the first set of data segments have been written, and prior to completely writing the second set of data segments and to receiving a packet notifier from the host, the network device processes the first DMA descriptor.

Low-latency packet processing for network device

The following description is directed to a configurable logic platform. In one example, a configurable logic platform includes host logic and a reconfigurable logic region. The reconfigurable logic region can include logic blocks that are configurable to implement application logic. The host logic can be used for encapsulating the reconfigurable logic region. The host logic can include a host interface for communicating with a processor. The host logic can include a management function accessible via the host interface. The management function can be adapted to cause the reconfigurable logic region to be configured with the application logic in response to an authorized request from the host interface. The host logic can include a data path function accessible via the host interface. The data path function can include a layer for formatting data transfers between the host interface and the application logic.

Configurable logic platform

An integrated circuit for distributing processing tasks includes a pre-selector circuit and a scheduler circuit. The pre-selector circuit is configured to receive a processing task, determine a category of the processing task, and select, from a set of task distribution techniques and based at least in part on the category of the processing task, a task distribution technique for distributing the processing task to a group of processing units. The scheduler circuit is configured to implement the selected task distribution technique to select, from the group of processing units, a target processing unit for performing the processing task.

Processing unit selection mechanism

A packet processing technique can include receiving a packet, and parsing the packet based on a protocol field to generate a parse result vector. The parse result vector is used to select between forwarding the packet to a virtual machine executing on a host processing integrated circuit, forwarding the packet to a physical media access controller, multicasting the packet to multiple virtual machines executing on the host processing integrated circuit, and sending the packet to a hypervisor.

Configurable parser and a method for parsing information units

Systems and methods are provided to reduce the latency in accessing an input/output (I/O) hardware register by software executing on a central processing unit (CPU). The hardware register is located in a controller coupled to the CPU via an I/O bus. The CPU software can send a command to the controller for execution. The controller can execute the command and update the hardware register to indicate that the command has been executed. The controller can write contents of the hardware register to a specified address in a CPU memory that is assigned by the CPU software. The CPU software can read the specified address to determine that the command has been executed instead of reading the hardware register on the I/O bus.

Input/output (I/O) register reflected in CPU memory

A technique for remote direct memory access (RDMA) may include receiving a packet that was sent over a network, and determining the packet has metadata used for queue selection. The technique may also include selecting a queue based on the metadata, and writing the data of the packet to an application memory using the datapath associated with the selected queue. Amended metadata can be generated to indicate that the data has been written to the application memory, and the amended metadata can be stored in a software accessible buffer.

Hybrid remote direct memory access

A clock disciplining scheme uses a pulse per second (PPS) signal that is distributed throughout a network to coordinate timing. In determining the time, jitter can occur due to latency between detection of the PPS signal and a software interrupt generated there from. This jitter affects the accuracy of the clock disciplining process. To eliminate the jitter, extra hardware is used to capture when the PPS signal occurred relative to a hardware clock counter associated with the clock disciplining software. In one embodiment, the extra hardware can be a sampling logic, which captures a state of a hardware clock counter upon PPS detection. In another embodiment, the extra hardware can initiate a counter that calculates a delay by the clock disciplining software in reading the hardware clock counter. The disciplining software can then subtract the calculated delay from a hardware clock counter to obtain the original PPS signal.

Clock synchronization in a network using a distributed pulse signal

A first configurable address decoder can be coupled between a source node and a first interconnect fabric, and a second address decoder can be coupled between the first interconnect fabric and a second interconnect fabric. The first address decoder can be configured with a first address mapping table that can map a first set of address ranges to a first set of target nodes connected to the first interconnect fabric. The second address decoder can be configured with a second address mapping table that can map a second set of address ranges to a second set of target nodes connected to the second interconnect fabric. The second address decoder can be part of the first set of target nodes. The first address decoder and the second address decoder can be configured or re-configured to determine different routes for a transaction from the source node to a target node in the second set of target nodes via the first and second interconnect fabrics.

Configurable routing in a multi-chip system

A coherent connection and a non-coherent connection are provided between system-on-chips (SoCs). The coherent connection can be coupled to coherent interconnects on the SoCs, and the non-coherent connection can be coupled to non-coherent interconnects on the SoCs. An input/output (I/O) transaction from an I/O device on a first SoC that is targeted to a second SoC can be transmitted via the non-coherent connection, and a processor transaction from the first SoC that is targeted to the second SoC can be transmitted via the coherent connection.

Non-coherent and coherent connections in a multi-chip system

Growing community of inventors

Tel Aviv, Israel

Erez Izenberg