摘要:
Sequences of items may be maintained using ordered locks. These items may correspond to anything, but using ordered locks to maintain sequences of packets may be particularly useful. One implementation uses a locking request, acceptance, and release protocol. One implementation associates instructions with locking requests such that when a lock is acquired, the locking mechanism executes or causes to be executed the associated instructions as an acceptance request of the lock is implied by the association of instructions (or may be explicitly requested). In some applications, the ordering of the entire sequence of packets is not required to be preserved, but rather only among certain sub-sequences of the entire sequence of items, which can be accomplished by converting an initial root ordered lock (maintaining the sequence of the entire stream of items) to various other locks (each maintaining a sequence of different sub-streams of items).
摘要:
The present invention provides a multithreaded processor, such as a network processor, that fetches instructions in a pipeline stage based on feedback signals from later stages. The multithreaded processor comprises a pipeline with an instruction unit in the early stage and an instruction queue, a thread interleaver, and an execution pipeline in the later stages. Feedback signals from the later stages cause the instruction unit to block fetching, immediately fetch, raise priority, or lower priority for a particular thread. The instruction queue generates a queue signal, on a per thread basis, responsive to a thread queue condition, etc., the thread interleaver generates an interleaver signal responsive to a thread condition, etc., and the execution pipeline generates an execution signal responsive to an execution stall, etc.
摘要:
The present invention provides a network multithreaded processor, such as a network processor, including a thread interleaver that implements fine-grained thread decisions to avoid underutilization of instruction execution resources in spite of large communication latencies. In an upper pipeline, an instruction unit determines an instruction fetch sequence responsive to an instruction queue depth on a per thread basis. In a lower pipeline, a thread interleaver determines a thread interleave sequence responsive to thread conditions including thread latency conditions. The thread interleaver selects threads using a two-level round robin arbitration. Thread latency signals are active responsive to thread latencies such as thread stalls, cache misses, and interlocks. During the subsequent one or more clock cycles, the thread is ineligible for arbitration. In one embodiment, other thread conditions affect selection decisions such as local priority, global stalls, and late stalls.
摘要:
In one embodiment, cells of a same packet are sent among multiple paths within a packet switching device. Each of these cells is associated with a same drop value for use in determining whether to drop or forward the cell at multiple positions within a packet switching fabric of a packet switching device in light of a current congestion measurement. In one embodiment, the drop value is calculated at each of these multiple positions based on fields of the cell that are packet variant, but not cell variant, so a same drop value is calculated by each cell of a packet. In one embodiment, at least one of these fields provides entropy (e.g., a timestamp of the packet) such that a produced drop value has, or approximately has, an equal probability of being any value within a predetermined range for fairness purposes.
摘要:
Disclosed are, inter alia, methods, apparatus, data structures, computer-readable media, and mechanisms, for identifying matches to a series of regular expressions, with the series of regular expressions including a first regular expression followed by a second regular expression, which avoids the potential overlap of characters used in matching the first and second regular expressions, while allowing individual deterministic finite automata (DFAs) to be used, whether standalone or as a merged DFA, which decreases the number of states required to represent the series of regular expressions. This potential overlap of characters can be avoided by adding marking states in a merged DFA as “divergent” in order to mask (e.g., ignore) a matching of the second regular expression for the potential overlap, or by using another DFA corresponding to the second regular expression for use during this divergent period.
摘要:
Methods and apparatus are disclosed using a random indication to map items to paths and to recirculate or delay the sending of a particular item when a destination over its mapped path is unreachable, including, but not limited to the context of sending of packets across multiple paths in a packet switching system. In one implementation, a set of items is buffered, with the set of items including a first and second sets of items. The items in the first set of items are forwarded over a set of paths in a first configuration. The set of paths is reconfigured into a second configuration, and the items in the second set of items are forwarded over the set of paths in the second configuration. In one implementation, a recirculation buffer is used to hold items not immediately sent. In one implementation, the paths are reconfigured in a random fashion.
摘要:
Methods and apparatus are disclosed for an adaptive rate control mechanism reactive to flow control messages in a packet switching system and other communications and computer systems. Typically, a multiplicative increase and exponential decrease technique is used to throttle traffic. Backpressure feedback is used to calculate the initial rate at which to allow traffic after backpressure is deasserted. This reduces the probability of underrun of buffers (e.g., too little traffic being carried). The adjustment to the initial rate is made by measuring the time between the XON and XOFF in factor periods. Then a target XON time is subtracted. If the result is positive (i.e., the measured XON time was too long), the rate is multiplicatively increased (e.g., by a factor of two to the difference). If the result is negative (i.e., the measured XON time was too short), the rate is exponentially decreased (e.g., by the square root).
摘要:
Methods and apparatus are disclosed for communicating time and latency sensitive information in a system, such as, but not limited to a computer or communications system. A first block of data is identified and transmitted. A check code is partially determined based on this first data. While the first data is being transmitted, the time-sensitive data (e.g., flow control, other control information, etc.) is identified. This identified time-sensitive data is then contiguously transmitted after the first data. The determination of the check code is completed based on the time-sensitive data, and the check code is contiguously transmitted after the time-sensitive data. One implementation receives the first data, the time-sensitive data, and the check code. If error correction is being used and is needed, the time-sensitive data is first corrected based on the check code, and then subsequently, the first data is corrected. In this manner, the latency of the availability of this time-sensitive data may be reduced.
摘要:
Methods and apparatus are disclosed for distributing flow control information in a packet switching system. In one packet switching system, flow control information is collected in a data structure in the first stage switching elements. Each of these switching elements transmit data from the flow control data structure as small messages or in fields included in packets being sent across multiple statically allocated paths. Flow control information is received by next stage elements, which are programmed to forward only flow control information received from a limited number of components or over a limited number of paths. The first stage switching elements may also periodically or occasionally delay sending flow control information or send a dummy message or information to accommodate bandwidth transmission differences between components of the packet switching system, including to accommodate bandwidth variations caused by plesiochronous timing across the network.
摘要:
Methods and apparatus are disclosed for using barrier phases to limit the disorder of packets which may be used in a computer or communications system. In one packet switching system, source nodes include an indication of their current barrier state in sent packets. For each barrier state, a predetermined range of sequence numbers may be used or a predetermined number of packets may be sent by a source node. The source, destination, and switching nodes are systematically switched between barrier phases, which is typically performed continuously in response to the flow of barrier request and barrier acknowledgement packets or signals. Each source node broadcasts to all forward connected nodes a barrier request to change to a next barrier state. After a switching node has received a barrier request on all incoming links, the switching node propagates the barrier request. Upon receiving barrier requests over all links, each destination stage relays an acknowledgement message to all connected source elements, which then send a barrier acknowledgement in much the same way, and each source element changes its barrier state causing the sequence number or counting space to be reset, and newly sent packets to indicate the new barrier state. Upon receiving all its barrier acknowledgement messages, each destination stage changes its barrier state, and then the destination can manipulate (e.g., resequence, reassemble, send, place in an output queue, etc.) packets marked with the previous barrier state as it knows that every packet from the previous barrier state has been received. This transition of barrier phases and limiting the number of packets sent per barrier phases may be used to limit the range of the sequence number space and the size of outgoing, resequencing, and reassembling buffers, as well providing a packet time-out mechanism which may be especially useful when non-continuous sequence numbers or time-stamps are included in packets for resequencing and/or reassembly purposes.