Hewlett Packard Enterprise Development LP

Hewlett-packard Company

Brian Peter L'Ecuyer

Charles Tuffli

Giuseppe Scaglione

Jonathan Michael Seely

Mark Shillingburg

Eric Pelletier

Steven Lahr

A network infrastructure device (e.g., network switch), that integrates solid-state drive (SSD) storage, using Non-volatile Memory Express (NVMe) data transfer protocol, for use by remote application hosts is provided. High availability configurations of network switches using direct rate control (RC) feedback for a plurality of submission queues mapped to SSD storage is provided. NVMe over fabric (NVMe-oF) is an implementation of NVMe protocol over a network fabric. Access to SSDs over network fabrics may be controlled using a direct RC feedback signal between an egress queue congestion accounting (associated with a single egress output) and a source node receiving input/output commands from remote hosts for the integrated SSD devices. In some implementations, direct RC feedback signals use hardware based signals. In some implementations, direct RC feedback signals are implemented in the hardware logic (silicon chip logic) within an internal switch fabric of the network switch.

Mapping NVMe-over-fabric packets using virtual output queues

Non-volatile memory express (NVMe) is a data transfer protocol used to enable high-speed data transfer between a host computer system and a solid-state drive (SSD). NVMe may be implemented over network fabrics and referred to as NVMe over fabrics (NVMe-oF). Access to SSD storage over network fabrics via NVMe-oF allows software defined storage to scale to allow access to a number of NVMe devices and extend distances between devices within a datacenter over which NVMe devices may be accessed. A network device is provided to automatically detect, prioritize, and route NVMe network packets in a network that includes multiple data communication protocols. For example, the network device may obtain network packets, analyze network packets to identify packet type and protocol, and redirect the network packets based on the analysis and detection. Thus, a processing priority may be provided for NVMe packets to assist in lossless communication implementations for storage across a network.

Separate routing of NVMe-over-fabric packets and non-NVMe packets

A network switch is provided. The network switch may incorporate resources for use as network attached resources for remote devices. Resources may include SSD storage for use as network attached storage for remote devices. The network switch may also include one or more client applications configured to detect an intended access path between a remote device and a non-volatile memory express (NVMe) storage device. The intended access path may be based on one of many different NVMe over fabric (NVMe-oF) protocols. The network switch (via the one or more client applications) automatically configures parameters to provide a network connectivity path between the remote device and the NVMe storage device. Communication includes at least one of a virtual local area network (VLAN) and a communication tunnel and/or other form of dedicated communication path to facilitate remotely accessible storage capacity provided from the network switch to the remote device.

Automatic system provisioning for NVME-over-fabric storage

One embodiment can provide a method and system for implementing flow control. During operation, a switch identifies a command from a host to access a storage device coupled to the switch. The switch queues the command in a command queue corresponding to the host. In response to determining that an amount of data pending transmission to the host from the storage device is below a predetermined threshold, the switch removes a command from the command queue and forwards the removed command to the storage device.

System and method for regulating NVMe-oF command requests and data flow across a network with mismatched rates

Examples include sampling a transmit rate of an egress port queue on a destination node, determining a utilization percentage of the egress port queue based on the transmit rate and a total rate capacity of the egress port, and determining a backlog percentage of the egress port queue. Examples also include determining an output rate for a virtual output queue on a source node based on the utilization percentage and the backlog percentage. The virtual output queue is for the egress port queue.

Output rates for virtual output queses

An example of a system may include a processing resource and a controller including a memory resource storing instructions executable by the processing resource to determine a rate of traffic communication at each of a plurality of ingresses participating in a communication of a packet flow context, determine a rate of traffic communication at each of a plurality of egresses participating in the communication of the packet flow context, determine a target packet admission rate applicable to each of the plurality of ingresses from the rate of traffic communication at each of the plurality of ingresses and the rate of traffic communication at each of the plurality of egresses, and communicate the target packet admission rate to an ingress of the plurality of ingresses.

Packet admission

Data is read from a first-in-first-out (FIFO) queue. A first condition flag is generated which indicates whether a read transaction of a first transaction size may be performed. When a write address for the FIFO queue is greater than a read address for the FIFO queue, the first condition flag is set to true when the read address plus the first transaction size is less than or equal to the write address. When the write address for the FIFO queue is less than the read address for the FIFO queue, the first condition flag is set to true when the read address plus the first transaction size is less than the write address plus a maximum depth of the FIFO queue. A first read transaction of the first transaction size from the FIFO queue is performed only when the first condition flag is true.

Growing community of inventors

Elk Grove, CA, United States of America

Brian Peter L'Ecuyer