Marvell Israel (misl) Ltd.

Marvellworld Trade Ltd.

Shira Turgeman

Gil Levy

Rami Zemach

Aviran Kadosh

Sharon Ulman

Amir Bishara

Amir Roitshtein

Evgeny Shumsky

A method for processing network packets in a network device is described. A network packet is stored in a transient buffer as the network packet is being received at an ingress port of the network device. After at least a first portion of the network packet has been received and before the entire network packet has been received: the first portion is processed to identify an egress port of the network device from which the network packet is to be transmitted; a congestion state of the egress port is determined; and the network packet is selectively transferred from the transient buffer to the identified egress port for transmission from the network device or a different action is performed on the network packet, based on the congestion state.

Methods and network device for performing cut-through

A packet received via a network is stored in a packet memory. Information in a header of the packet is analyzed to determine multiple ports to which the packet is to be forwarded. The packet is read from the packet memory in connection with transmitting the packet via a first port among the multiple ports. Packet data corresponding to the packet is stored in a packet cache different than the packet memory. In connection with forwarding one or more instances of the packet to ports among the multiple ports, the packet data is read, based on information included in a packet descriptor associated with the packet, from the packet cache. Ones of the one or more instances of the packet are forwarded, using the data corresponding to the packet read from the packet cache and without further accessing the packet memory, to respective one or more of the multiple ports.

Network device architecture using cache for multicast packets

In a method for managing memory space in a network device two or more respective allocation requests from two or more processing cores among a plurality of processing cores sharing a memory space are received at a memory management device during a first single clock cycle of the memory management device, the two or more allocation requests requesting to allocate, to the two or more processing cores, respective buffers in the shared memory space. In response to receiving the two or more allocation requests, the memory management device allocates to the two or more processing cores, respective two or more buffers in the shared memory space. Additionally, the management device, during a second single clock cycle of the memory management device, transmits respective allocation responses to each of the two or more processing cores, wherein each allocation response includes an indication of a respective allocated buffer.

Scalable dynamic memory management in a network device

A packet received via a network is stored in a packet memory. Information in a header of the packet is analyzed to determine one or more ports to which the packet is to be forwarded. The packet is read from the packet memory. It is ascertained whether the packet is of a multicast-type. In response to determining that the packet is the multicast-type, packet data that is i) read from the packet memory and ii) corresponds to the packet, is stored in a packet cache; the packet data is read from the packet cache, and multiple instances of the packet are forwarded to appropriate ports without further accessing the packet memory.

In a method for processing network packets, a network packet is received at an ingress port. After at least a first portion of the network packet has been received, but before the network packet has been completely received, a first data structure representing the network packet is generated based on the first portion, the first data structure is processed, and the network packet begins to be selectively forwarded to a first one or more egress ports, or selectively not forwarded to any egress port, responsively to processing the first data structure. A second data structure representing the network packet is generated and, after the network packet has been completely received, the second data structure is processed, and the network packet is selectively forwarded to a second one or more egress ports, different from the first one or more egress ports, responsively to processing the second data structure.

Cut-through processing for slow and fast ports

In a method for processing network packets, a network packet is received at an ingress port. Before the network packet has been completely received at the ingress port, a first data structure representing the network packet is generated based on a received first portion of the network packet, and the first data structure is processed at a packet processor of the network device. Processing the first data structure includes making a forwarding decision for the network packet. A second data structure representing the network packet is generated and, after the network packet has been completely received at the ingress port, at least one or more non-forwarding operations are performed with respect to the network packet using at least the second data structure.

Packets received via ports coupled to network links are processed to determine target ports to which the packets are to be forwarded. Appropriate control paths in a network device are selected for processing multicast packets from among a plurality of different control paths having respective processing latencies, the different control paths providing alternative processing paths for processing control information for multicast packets. Multicast packets are further processed using the selected control paths.

Growing community of inventors

Nirit, Israel

Shira Turgeman