摘要:
A method and structure is provided for buffering data packets having a header and a remainder in a network processor system. The network processor system has a processor on a chip and at least one buffer on the chip. Each buffer on the chip is configured to buffer the header of the packets in a preselected order before execution in the processor, and the remainder of the packet is stored in an external buffer apart from the chip. The method comprises utilizing the header information to identify the location and extent of the remainder of the packet. The entire selected packet is stored in the external buffer when the buffer of the stored header of the given packet is full, and moving only the header of a selected packet stored in the external buffer to the buffer on the chip when the buffer on the chip has space therefor.
摘要:
A pipeline configuration is described for use in network traffic management for the hardware scheduling of events arranged in a hierarchical linkage. The configuration reduces costs by minimizing the use of external SRAM memory devices. This results in some external memory devices being shared by different types of control blocks, such as flow queue control blocks, frame control blocks and hierarchy control blocks. Both SRAM and DRAM memory devices are used, depending on the content of the control block (Read-Modify-Write or ‘read’ only) at enqueue and dequeue, or Read-Modify-Write solely at dequeue. The scheduler utilizes time-based calendars and weighted fair queueing calendars in the egress calendar design. Control blocks that are accessed infrequently are stored in DRAM memory while those accessed frequently are stored in SRAM.
摘要:
Systems and methods for distributing thread instructions in the pipeline of a multi-threading digital processor are disclosed. More particularly, hardware and software are disclosed for successively selecting threads in an ordered sequence for execution in the processor pipeline. If a thread to be selected cannot execute, then a complementary thread is selected for execution.
摘要:
Access arbiters are used to prioritize read and write access requests to individual memory banks in DRAM memory devices, particularly fast cycle DRAMs. This serves to optimize the memory bandwidth available for the read and the write operations by avoiding consecutive accesses to the same memory bank and by minimizing dead cycles. The arbiter first divides DRAM accesses into write accesses and read accesses. The access requests are divided into accesses per memory bank with a threshold limit imposed on the number of accesses to each memory bank. The write receive packets are rotated among the banks based on the write queue status. The status of the write queue for each memory bank may also be used for system flow control. The arbiter also typically includes the ability to determine access windows based on the status of the command queues, and to perform arbitration on each access window.
摘要:
Systems and methods for scheduling data packets in a network processor are disclosed. Embodiments provide a network processor that comprises a best-effort scheduler with a minimal calendar structure for addressing schedule control blocks. In one embodiment, a three-entry calendar structure provides for weighted best effort scheduling. Each of a plurality different flows has an associated schedule control block. Schedule control blocks are stored as linked lists in a last-in-first-out buffer. Each calendar entry is associated with a different linked list by storing in the calendar entry the address of the first-out schedule control block in the linked list. Each schedule control block has a counter and is assigned a weight according to the bandwidth priority of the flow to which the corresponding packet belongs. Each time a schedule control block is accessed from a last-in-first-out buffer storing the linked list, the scheduler generates a scheduling event and the counter of the schedule control block is incremented. When an incremented counter of a schedule control block equals its weight, the schedule control block is temporarily removed from further scheduling.
摘要:
A communication network used to link information handling systems together utilizes a switching network to transmit data among senders and receivers. Each individual packet of data is described and controlled by an FCB. The bandwidth associated with the storing and distribution of data is optimized by chaining the data packets in different types of queues, or operating without chaining outside a queue. When a frame is in an output queue, the third word contains an RFCBA for egress of the frame to a line port, and an MCID for ingress from an output queue to a switch port. The RFCBA and the MCID have multicast capabilities. The format does not require a third word when a frame is in an input queue.
摘要:
A system and method for computing a blind checksum includes a host Ethernet adapter (HEA) with a system for receiving a packet. The system determines whether or not the packet is in Internet protocol version four (IPv4). If the packet is not in IPv4, the system computes the checksum of the packet. If the packet is in IPv4, the system determines whether the packet is in transmission control protocol (TCP) or user datagram protocol (UDP). If the packet is not in either of TCP or UDP the system attaches a pseudo-header to the packet and computes the checksum of the packet based on the pseudo-header and the IPv4 standard.
摘要:
Systems and methods for implementing multi-frame control blocks in a network processor are disclosed. Embodiments include systems and methods to reduce long latency memory access to less expensive memory such as DRAM. As a network processor in a network receives packets of data, the network processor forms a frame control block for each packet. The frame control block contains a pointer to a memory location where the packet data is stored, and is thereby associated with the packet. The network processor associates a plurality of frame control blocks together in a table control block that is stored in a control store. Each table control block comprises a pointer to a memory location of a next table control block in a chain of table control blocks. Because frame control blocks are stored and accessed in table control blocks, less frequent memory accesses may be needed to keep up with the frame rate of packet transmission.
摘要:
Systems and methods for adaptively mapping system memory address bits into an instruction tag and an index into the cache are disclosed. More particularly, hardware and software are disclosed for observing collisions that occur for a given mapping of system memory bits into a tag and an index. Based on the observations, an optimal mapping may be determined that minimizes collisions.
摘要:
Systems and methods for scheduling data packets in a network processor are disclosed. Embodiments provide a network processor that comprises a best-effort scheduler with a minimal calendar structure for addressing schedule control blocks. In one embodiment, a four-entry calendar structure provides for rate-limited weighted best effort scheduling. Each of a plurality of different flows has associated schedule control blocks. Schedule control blocks are stored as linked lists in a last-in-first-out buffer. Each calendar entry is associated with a different linked list by storing in the calendar entry the address of the first-out schedule control block in the linked list. Each schedule control block has a counter and is assigned a rate limit according to the bandwidth priority of the flow to which the corresponding packet belongs. Each time a schedule control block is accessed from a last-in-first-out buffer storing the linked list, the scheduler generates a scheduling event and the counter of the schedule control block is incremented. When an incremented counter of a schedule control block equals its rate limit, the schedule control block is temporarily removed from further scheduling until a time interval concludes.