Sun Microsystems, Inc.

William C Van Loo

Satyanarayana Nishtala

Zahir Ebrahim

Kevin B Normoyle

Louis F Coffin, Iii

Paul N Loewenstein

Charles E Narad

Leslie Dean Kohn

John E Watkins

Sun-den Chen

David Labuda

Kurt Michels

Louis F Coffin Iii

Sue K Lee

Hugh Chang

Raymond Ng

A method and apparatus for packet-switched flow control of transaction requests in uniprocessor and multiprocessor computer systems that maximizes system resource utilization and throughput, and minimizes system latency. The computer system comprises one or more master interfaces, one or more slave interfaces, and an interconnect system controller which provides dedicated transaction request queues for each master interface and controls the forwarding of transactions to each slave interface. The master interface keeps track of the number of requests in the dedicated queue in the system controller, and the system controller keeps track of the number of requests in each slave interface queue. Both the master interface, and system controller a piori know the maximum capacity of the queue immediately downstream from it, and does not issue more transaction requests than what the downstream queue can accommodate.

Fast forwarding slave requests in a packet-switched computer system by transmitting request to slave in advance to avoid arbitration delay when system controller validates request

A system for maintaining reliable packet distribution in a ring network with support for strongly ordered, nonidempotent commands. Each consumer node on the network maintains a record of the sequence of packets that have passed through that node, and the state of each of the packets at the time it passed through, including a record of the last known good packet and its sequence number. When a producer node detects an error condition in an acknowledgment for a packet, resends all packets beginning with the last known good packet. Each consumer node is able to process or reject the resent packets, including packets that may already have been processed, which it is aware of due to the packet and state records for all packets. Strong ordering is thus supported, since the sending order of packets can be maintained for processing due to the resends, and nonidempotent commands are supported due to the consumer nodes' ability to determine whether they have already processed a given packet, and to send an acknowledge-done reply if that is the case.

System for dynamic ordering support in a ringlet serial interconnect

A system for maintaining reliable packet distribution in a ring network with support for strongly ordered, nonidempotent commands. Each consumer node on the network maintains a record of the sequence of packets that have passed through that node, and the state of each of the packets at the time it passed through, including a record of the last known good packet and its sequence number. When a producer node detects an error condition in an acknowledgement for a packet, resends all packets beginning with the last known good packet. Each consumer node is able to process or reject the resent packets, including packets that may already have been processed, which it is aware of due to the packet and state records for all packets.

System for maintaining strongly sequentially ordered packet flow in a ring network system with busy and failed nodes

A data transfer control system for a computer system having a packet-switched data bus, for controlling the transfer of data words from a device having one bus width to a device having a different bus width. The first bus may be the master bus, and the second bus is a bus of a functional unit, such as a memory or other device, connected to the computer system and coupled to the master bus. When the second bus is smaller than the first bus, the data transfer control system accommodates this by transferring only partial words per clock cycle to the second bus, effectively delaying the transference of data to a rate that the second bus can handle. The transfer rate factor is substantially equal to the ratio of the second bus width to the first bus width. In an alternative embodiment, the data transfer control system transfers the data words at substantially the full rate available, given the computer system's processor or clock speed and the width of the master bus, and the words are buffered before transference to the second bus. The speed of transference to the buffer is then not limited by the size of the second bus, but only by the size of the buffer, which preferably has an input bus at least as wide as that of the master bus. The system accommodates functional units of different sizes all to be couple to a master bus of a given size, where the different sizes may be larger or smaller than the master bus size, and accordingly the functional units can be built with optimal bus sizes for their particular functions without having to conform to a perhaps inappropriate, inefficient or expensive or wasteful master bus size.

System for multisized bus coupling in a packet-switched computer system

A method and system for packet-switched flow control of transaction requests that maximizes resource utilization and throughput, and minimizes latency. A system controller provides dedicated transaction request queues and controls the forwarding of transactions from a processor to a slave. The transaction requests are automatically forwarded to an intended slave on the same address bus as the system controller immediately. The system controller determines whether the proper criteria are met for that slave to receive such a request, such as the slave's request receive queue is not full and that global ordering requirements are met. If so, then on a separately provided line, the system controller validates the request for immediate reception by the slave.

Method and apparatus for fast-forwarding slave requests in a packet-switched computer system

A system for maintaining reliable packet distribution in a ring network with support for strongly ordered, nonidempotent commands. Each consumer node on the network maintains a record of the sequence of packets that have passed through that node, and the state of each of the packets at the time it passed through, including a record of the last known good packet. and its sequence number. When a producer node detects an error condition in an acknowledgment for a packet, resends all packets beginning with the last known good packet Each consumer node is able to process or reject the resent packets, including packets that may already have been processed, which it is aware of due to the packet and state records for all packets. Strong ordering is thus supported, since the sending order of packets can be maintained for processing due to the resends, and nonidempotent commands are supported due to the consumer nodes' ability to determine whether they have already processed a given packet, and to send an acknowledge-done reply if that is the case. The system is equipped to operate successfully in the presence of a failed or very busy node by maintaining a queue of busy acks from each node, or from nodes that are particularly busy, and when the queue is full taking steps to bypass the busy or failed node. Alternatively, the system can detect that some predetermined period of time has passed during which either no response or only busy responses have been received from a given node, and at that point determine to effectively bypass the node for future transactions, at least for some timeout period, so that packet transactions for the rest of the network can proceed unhindered by the ill-behaved node.

A system for maintaining reliable packet distribution in a ring network with support for strongly ordered, nonidempotent commands. Each consumer node on the network maintains a record of the sequence of packets that have passed through that node, and the state of each of the packets at the time it passed through, including a record of the last known good packet and its sequence number. When a producer node detects an error condition in an acknowledgment for a packet, resends all packets beginning with the last known good packet. Each consumer node is able to process or reject the resent packets, including packets that may already have been processed, which it is aware of due to the packet and state records for all packets. Strong ordering is thus supported, since the sending order of packets can be maintained for processing due to the resends, and nonidempotent commands are supported due to the consumer nodes' ability to determine whether they have already processed a given packet, and to send an acknowledge-done reply if that is the case. The system is equipped to operate successfully in the presence of a failed or very busy node by maintaining a queue of busy acks from each node, or from nodes that are particularly busy, and when the queue is full taking steps to bypass the busy or failed node. Alternatively, the system can detect that some predetermined period of time has passed during which either no response or only busy responses have been received from a given node, and at that point determine to effectively bypass the node for future transactions, at least for some timeout period, so that packet transactions for the rest of the network can proceed unhindered by the ill-behaved node.

System for maintaining strongly sequentially ordered packet flow in a

In a computer system including a packet-switched bus, a method for requesting transactions such that memory accesses are initiated quickly. A master transmits a first portion of a transaction request packet having multiple portion. A memory controller receives the first portion of the transaction request, which includes a row address portion of a memory address. The memory controller initiates a memory access by applying a row address strobe signal to the row of the memory location in response to receiving the first portion of the request packet, and the master transmits any remaining portion of the transaction request. After the full memory address has been received, it is determined whether data stored at the memory location is to be read from a source other than the memory location. The memory controller aborts the memory access by inhibiting assertion of a column access strobe signal to the memory location if the data is to be read from a source other than the memory location.

Method and apparatus for quickly initiating memory accesses in a

This invention describes a link-by-link flow control method for packet-switched uniprocessor and multiprocessor computer systems that maximizes system resource utilization and throughput, and minimizes system latency. The computer system comprises one or more master interfaces, one or more slave interfaces, and an interconnect system controller which provides dedicated transaction request queues for each master interface and controls the forwarding of transactions to each slave interface. The master interface keeps track of the number of requests in the dedicated queue in the system controller, and the system controller keeps track of the number of requests in each slave interface queue. Both the master interface, and system controller know the maximum capacity of the queue immediately downstream from it, and does not issue more transaction requests than what the downstream queue can accommodate. An acknowledgment from the downstream queue indicates to the sender that there is space in it for another transaction. Thus no system resources are wasted trying to send a request to a queue that is already full.

Method and apparatus for flow control in packet-switched computer system

A multiprocessor computer system has a multiplicity of sub-systems and a main memory coupled to a system controller. Some of the sub-systems are data processors, each having a respective cache memory that stores multiple blocks of data and a respective set of master cache tags (Etags), including one Etag for each data block stored by the cache memory. Each data processor includes an interface for sending memory transaction requests to the system controller and for receiving cache transaction requests from the system controller corresponding to memory transaction requests by other ones of the data processors. The system controller includes transaction activation logic for activating each said memory transaction request when it meets predefined activation criteria, and for blocking each said memory transaction request until the predefined activation criteria are met. An active transaction status table stores status data representing memory transaction requests that have been activated, including an address value for each activated transaction. The transaction activation logic includes comparator logic for comparing each memory transaction request with the active transaction status data for all activated memory transaction requests so as to detect whether activation of a particular memory transaction request would violate the predefined activation criteria. With certain exceptions concerning writeback transactions, an incoming transaction for accessing a data block that maps to the same cache line a pending, previously activated transaction, will be blocked until the pending transaction that maps to the same cache line is completed.

Transaction activation processor for controlling memory transaction

An apparatus and method for handling interrupt requests from any of a plurality of interrupters to any of a plurality of interrupt handlers. Each interrupt handler includes an interrupt input request queue for holding a plurality of incoming interrupt requests. A system controller is connected to the interrupters and the interrupt handlers, and includes an input queue coupled to each interrupter for receiving a plurality of interrupt requests. The system controller includes a processor and a memory storing instructions for controlling its operation. The system controller also includes an output queue coupled to each interrupt handler (which in many cases will also be interrupters), and a counter for monitoring the current number, at any given time, of interrupt requests pending in each of the interrupt input queues in the interrupt handlers. When an interrupt request is sent from an interrupter, the system controller determines a target for the request, which may be by a target ID in the request or may be based upon a predetermined method to ensure even distribution of interrupt requests among all or a subset of the interrupt handlers. The system controller determines whether the target interrupt handler's input queue is full, and if not then it passes on the interrupt request and sends a positive acknowledgment to the interrupter. If the queue is full, then a negative acknowledgment is sent, and the interrupter then waits a random period of time and sends the interrupt request again. The target interrupt handler may thus accept multiple interrupt requests and process them in order without negative acknowledgments.

Method and apparatus for interrupt communication in packet-switched

A system for resolving packet errors and busy acknowledgments for packets in a ring network, to maintain continued packet distribution on the network with support for strongly ordered, nonidempotent commands. Each consumer node on the network maintains a record of the sequence of packets that have passed through that node, and the state of each of the packets at the time it passed through, including a record of the last known good packet and its sequence number. When a producer node detects an error condition in an acknowledgment for a packet, resends all packets beginning with the last known good packet. Each consumer node is able to process or reject the resent packets, including packets that may already have been processed, which it is aware of due to the packet and state records for all packets. Strong ordering is thus supported, since the sending order of packets can be maintained for processing due to the resends, and nonidempotent commands are supported due to the consumer nodes' ability to determine whether they have already processed a given packet, and to send an acknowledge-done reply if that is the case. A retry queueing system is used which allows the system to effectively interleave busy retry loops with an error retry loop, thus accommodating both error retries and busy retries while preserving strong sequential ordering, if desired, of the sent packets.

System for preserving sequential ordering and supporting nonidempotent

The present invention provides a scalable, modular and pipelined distributed bus arbitration system for efficiently resolving bus contention between sub-systems, e.g., processors, coupled to a common system bus. The arbitration system includes a plurality of distributed bus arbiters which receives the bus requests from the sub-systems and independently determine the next bus master. The arbitration protocol enables the arbitration process to be eliminated from the critical timing path thereby allowing the system to operate at the maximum system clock frequency possible for a given integrated circuit (IC) technology to reduce overall system clock latencies. Any change among the sub-systems during an arbitration clock cycle is based on any system bus request(s) which are active during a clock cycle immediately preceding the arbitration clock cycle, and is independent of any system bus request(s) asserted during the arbitration clock cycle. In addition, the arbitration protocol treats a current bus master, i.e., the bus master driving the system bus, preferentially. Each arbitration task is completed within a system clock cycle regardless of processor speed. As a result, the arbitration latency for retaining the current bus master is one system clock cycle while the latency for selecting and switching bus masters is two system clock cycles. In this implementation, a last port driver is the only sub-system permitted to assert a bus request in a clock cycle and immediately drive the system bus in the next immediate clock cycle. Conversely, when a second sub-system which is not the last port driver needs to drive an inactive system bus, the second sub-system asserts its bus request line in a first clock cycle, and arbitration occurs within all the respective bus arbiters occurs in a second clock cycle.

Growing community of inventors

Palo Alto, CA, United States of America

William C Van Loo