Broadcom Corporation

Brad Matthews

Puneet Agarwal

Bruce Hui Kwan

Mohan V Kalkunte

Yongbum Kim

Michael David Johas Teener

Howard M Frazier

A system and method to monitor network congestion is provided. The system includes a plurality of ingress, egress ports, and a plurality of queues coupled to the ingress and egress ports and configured to store incoming and outgoing packets. The system also includes a monitoring unit configured to monitor at least one attribute of packets in at least one queue when a start condition occurs, stop monitoring the attribute when an end condition occurs, determine a flow that caused the start condition based on the monitored attribute, and report the monitored attribute and the flow.

Identifying flows causing undesirable network events

A system for transferring data includes an egress node including an egress port, and an ingress node configured to receive a data segment of a data packet destined for the egress port. The data packet is associated with a packet priority level. The ingress node is configured to receive an egress statement vector from the egress node indicating whether the egress port is or is not flow controlled for data associated with the packet priority level. The ingress node is configured to determine whether the egress port is available to receive the data segment from the ingress node before other data segments of the data packet are received at the ingress node based on the egress statement vector.

System for performing distributed data cut-through

While transmitting a first Ethernet frame from the first buffer onto an Ethernet link, a first Ethernet device may stop transmitting the first frame prior to completing transmission of the frame. The first Ethernet device may then transmit a second frame from a second buffer onto the Ethernet link. The first Ethernet device may resume transmission of the first frame from the first buffer onto the Ethernet link. A second Ethernet device may receive, via the Ethernet link, a first portion of a first Ethernet frame and store the first portion of the first Ethernet frame in a first buffer. The second Ethernet device may then receive, via the Ethernet link, a second Ethernet frame and store the second Ethernet frame in a second buffer. The second Ethernet device may then receive, via the Ethernet link, a second portion of the first Ethernet frame and append it to the contents of the first buffer.

Packet preemption for low latency

Methods and apparatus for improving hash-based load balancing with per-hop seeding are disclosed. The methods and apparatus described herein provide a set of techniques that enable nodes to perform differing mathematical transformations when selecting a destination link. The techniques include manipulation of seeds, hash configuration mode randomization at a per node basis, per node/microflow basis or per microflow basis, seed index generation, and member selection. A node can utilize any, or all, of the techniques presented in this disclosure simultaneously to improve traffic distribution and avoid path starvation with a degree of determinism.

Hash-based load balancing with per-hop seeding

Various methods and systems are provided for oversubscription buffer management. In one embodiment, among others, a method for oversubscription control determines a utilization level of an oversubscription buffer that is common to a plurality of ingress ports and initiates adjustment of an ingress packet rate of the oversubscription buffer in response to the utilization level. In another embodiment, a method determines an occupancy level of a virtual oversubscription buffer associated with an oversubscription buffer and initiates adjustment of an ingress packet rate in response to the occupancy level. In another embodiment, a rack switch includes an oversubscription buffer configured to receive packets from a plurality of ingress ports and provide the received packets for processing by the rack switch and a packet flow control configured to monitor an occupancy level of the oversubscription buffer and to initiate adjustment of an ingress packet rate in response to the occupancy level.

Oversubscription buffer management

A system for multicast switching for distributed devices may include an ingress node including an ingress memory and an egress node including an egress memory, where the ingress node is communicatively coupled to the egress node. The ingress node may be operable to receive a portion of a multicast frame over an ingress port, bypass the ingress memory and provide the portion to the egress node when the portion satisfies an ingress criteria, otherwise receive and store the entire frame in the ingress memory before providing the frame to the egress node. The egress node may be operable to receive the portion from the ingress node, bypass the egress memory for the portion and provide the portion to the first egress port when an egress criteria is satisfied, otherwise receive and store the entire multicast frame in the egress memory before providing the multicast frame to an egress port.

Multicast switching for distributed devices

A data segment of a data packet destined for an egress port of an egress node may be received at a first ingress node. An egress statement vector and an ingress statement vector may be identified at the first ingress node. A determination may be made, based on the egress statement vector and ingress statement vector, whether the first ingress node is authorized to transfer the data segment to the egress port before the other data segments of the data packet are received at the first ingress node. The data segment may be transferred to the egress port before the other data segments of the data packet are received at the first ingress node when the determination indicates the first ingress node is authorized. The data segment may be stored in a buffer of the first ingress node when the determination indicates the first ingress node is not authorized.

A system transfers data. The system includes an ingress node transferring data at a determined bandwidth. The ingress node includes a buffer and operates based on a monitored node parameter. The system includes a controller in communication with the ingress node. The controller is configured to allocate, based on the monitored node parameter, an amount of the determined bandwidth for directly transferring data to bypass the buffer of the ingress node.

System for performing data cut-through

Disclosed are various embodiments that provide an architecture of memory buffers for a network component configured to process packets. A network component may receive a packet, the packet being associated with a control structure and packet data, an input port set and an output port set. The network component determines one of a plurality of control structure memory partitions for writing the control structure, the one of the plurality of control structure memory partitions being determined based at least upon the input port set and the output port set; and determines one of a plurality of packet data memory partitions for writing the packet data, the one of the plurality of packet data memory partitions being determined independently of the input port set.

Scalable egress partitioned shared memory architecture

Various example embodiments are disclosed. According to an example embodiment, an apparatus may include a switch fabric. The switch fabric may be configured to assign packets to either a first flow set or a second flow set based on fields included in the packets. The switch fabric may also be configured to send a first packet from the first flow set to a first flow set destination via a first path. The switch fabric may also be configured to determine, based at least in part on delays of the first path and a second path, whether sending a second packet from the first flow set to the first flow set destination via a second path will result in the second packet reaching the first flow set destination after the first packet reaches the first flow set destination, the second packet having been received by the router after the first packet. The switch fabric may also be configured to send the second packet to the first flow set destination via the second path based at least in part on the determining that sending the second packet from the first flow set to the first flow set destination via a second path will result in the second packet reaching the first flow set destination after the first packet reaches the first flow set destination.

Network traffic management

The systems and methods described herein allow for the scaling of output-buffered switches by decoupling the data path from the control path. Some embodiment of the invention include a switch with a memory management unit (MMU), in which the MMU enqueues data packets to an egress queue at a rate that is less than the maximum ingress rate of the switch. Other embodiments include switches that employ pre-enqueue work queues, with an arbiter that selects a data packet for forwarding from one of the pre-enqueue work queues to an egress queue.

Scaling output-buffered switches

Disclosed are various embodiments for improving hash table utilization. A key corresponding to a data item to be inserted into a hash table can be transformed to improve the entropy of the key space and the resultant hash codes that can generated. Transformation data can be inserted into the key in various ways, which can result in a greater degree of variance in the resultant hash code calculated based upon the transformed key.

Hash table organization

Processing techniques in a network switch help reduce latency in the delivery of data packets to a recipient. The processing techniques include speculative flow status messaging, for example. The speculative flow status messaging may alert an egress tile or output port of an incoming packet before the incoming packet is fully received. The processing techniques may also include implementing a separate accelerated credit pool which provides controlled push capability for the ingress tile or input port to send packets to the egress tile or output port without waiting for a bandwidth credit from the egress tile or output port.

Reducing store and forward delay in distributed systems

A system and method for achieving lossless packet delivery in packet rate oversubscribed systems. Link-level packet rate control can be effected through the transmission of packet rate control messages to link partners of an oversubscribed system. The transmission of packet rate control messages can be triggered upon a determination that a packet arrival rate over a set of ingress ports exceeds a packet processing rate of a packet processing unit bound to the set of ingress ports. In one embodiment, the packet processing rate is artificially reduced due to a reduction in power consumption in the oversubscribed system.

System and method for achieving lossless packet delivery in packet rate oversubscribed systems

Methods and apparatus for dynamic load balancing using virtual link credit accounting are disclosed. An example method includes receiving, at a network device, a data packet to be communicated using an aggregation group, the aggregation group including a plurality of virtual links having a common destination. The example method further includes determining a hash value based on the packet and determining an assigned virtual link of the plurality of virtual links based on the hash value. The example method also includes reducing a number of available transmission credits for the aggregation group and reducing a number of available transmission credits for the assigned virtual link. The example method still further includes communicating the packet to another network device using the assigned virtual link.

Growing community of inventors

San Jose, CA, United States of America

Brad Matthews