Hitachi, Ltd.

Hitachi Vlsi Engineering Corp.

Other

Hitachi Ulsi Engineering Corp.

Hitachi Microcomputer Eng. Ltd.

Hitachi Engineering Co., Ltd.

Hitachi, Vlsi Eng. Corp.

Shigeo Takeuchi

Tatsuo Higuchi

Junji Nakagoshi

Naoki Hamanaka

Yasuhiro Ogata

Taturu Toba

Tatsuru Toba

Shinichi Shutoh

Teruo Tanaka

Yasuhiro Inagami

Tohru Shonai

Hiroyuki Chiba

Toshiaki Tarui

Katsuyoshi Kitai

Yooichi Shintani

Eiki Kamada

Machiko Asaie

Yoshiko Yasuda

Tadaaki Isobe

Hiroaki Fujii

Shin'ichi Shutoh

Takayuki Nakagawa

Kiyohiro Obara

Masahiro Yamada

Hideo Wada

Yuji Saeki

Toshimitsu Ando

Masato Ishii

Mitsuyoshi Igai

Osamu Arakawa

Yasuhiro Inagaki

A computer system including a plurality of processing nodes, at least one resource provided for use by any of the processing nodes and a plurality of register sets. Each register set is provided in each processing node for storing in parallel use status information indicating whether the resource is in exclusive use status. The computer system includes a plurality of request issue circuits, each being provided in each processing node, for issuing requests for exclusive use of the resource, a message exchanging circuit for serializing requests issued by the request issue circuits into a serialized order and broadcasting the request to the processing nodes and a plurality of status control circuits. Each status control circuit is provided in each processing node to update a corresponding register set depending on use status information and each request received at a corresponding node.

EXCLUSIVE CONTROL METHOD WITH EACH NODE CONTROLLING ISSUE OF AN EXCLUSIVE USE REQUEST TO A SHARED RESOURCE, A COMPUTER SYSTEM THEREFOR AND A COMPUTER SYSTEM WITH A CIRCUIT FOR DETECTING WRITING OF AN EVENT FLAG INTO A SHARED MAIN STORAGE

A computer system including a plurality of processing nodes, at least one resource provided for use by any of the processing nodes and a plurality of register sets. Each register set is provided in each of the processing nodes for storing in parallel use status information indicating whether the resource is in exclusive use status or not. The computer system can also include a plurality of request issue circuits, each being provided in each of the processing nodes, for issuing individually requests for exclusive use of the resource, a message exchanging circuit for serializing requests issued by the request issue circuits into a serialized order and broadcasting the request to all of the processing nodes in the serialized order and a plurality of status control circuits. Each status control circuit is provided in each of the processing nodes corresponding to each of the register sets to update individually a corresponding register set depending on use status information stored in the corresponding register set and each of the requests for exclusive use of the resource received at a corresponding node.

Exclusive control method with each node controlling issue of an exclusive use request to a shared resource, a computer system therefor and a computer system with a circuit for detecting writing of an event flag into a shared main storage

A data processor includes a register group having registers of the number larger than the number of registers which can be designated by a register specifier field of an instruction. The register group consists of a plurality of register queues with respect to logical register numbers designated in the instruction, each register queue including a plurality of physical registers. In the data processor, a physical register number forming section is provided for converting the logical register number to a physical register number in the register queue corresponding to the logical register number, by using queue control information designated in the register specifier field and read/write information decided by the kind of the instruction and the position of the register specifier field in the instruction.

Data processor with multiple register queues

In order to determine a transfer path of a message to a receiving-end processor group, a processor includes a routing bit generation circuit, and an exchange switch includes partial broadcast path control circuits and a path control information alteration circuit. In order to define the range of a receiving-end processor group, a network includes transfer control circuits. A crossbar switch includes transfer control circuits associated with output ports and a boundary register group. When a partial broadcast message is transferred from an input port in the downstream direction of an output port, it is decided whether a belonging to the partial broadcast range associated with a connected to the particular input port is connected to the particular output port, whereby the particular partial broadcast message is transferred from the same output port.

Partial broadcast method in parallel computer and a parallel computer

In a method of transferring packets in a network for a parallel processor system handling a one-to-one transfer packet to be transferred from a processor to another processor and a broadcast packet to be transferred from a processor to a plurality of other processors, a transfer request of a broadcast packet is preferentially selected and a check is made to detect whether or not a plurality of processors specified as receivers are in a state in which the packet can be received. The broadcast packet is transferred to the processors found to be in the state in which the packet can be received. The packet transfer is delayed for the other processors in a state in which the packet cannot be received. Namely, only when the state of the processors is changed to the state in which the packet can be received, the broadcast packet is transferred thereto.

Switching system for transferring broadcast packet held in broadcast

In a parallel processor system comprising a plurality of processor elements constituting a network, a source processor element wishing to broadcast data to a plurality of destination processor elements sends a broadcast request message containing the target data to a broadcast exchanger. The broadcast exchanger converts the received message into a broadcast message and sends it over the network to the destinations. A plurality of broadcast request messages, if transmitted parallelly to the broadcast exchanger, are serialized thereby so that only one broadcast message will be transmitted at a time over the network. This prevents deadlock from occurring between different broadcast messages. The routes for transmitting broadcast request messages and those for transmitting broadcast messages are arranged so as not to overlap with one another. This suppresses deadlock between any broadcast request message and broadcast message. The broadcast exchanger is replaced alternatively with one of the partial networks. These schemes all apply where long messages are transmitted through worm-hole routing.

Parallel processor system with a broadcast message serializing circuit

A parallel computer system includes a plurality of processor units, a data transfer network for interconnecting the processor units, a synchronizing network for allowing program execution to be performed synchronously by the individual processor units, a connecting unit for connecting the individual processor units and the synchronizing network, and an input unit connected to the synchronizing network. The connecting unit connects selectively the individual processor units to the synchronizing network in accordance with information inputted via the input unit. With the parallel computer system, a program can be executed synchronously in parallel by using a desired number of processor units of those incorporated in the system, whereby availability of processor resources of the system can be enhanced.

Parallel computer system

In order to reduce load at a resource managing node for exclusive control of a shared resource, each node has a group of lock state registers each corresponding to one of the nodes. Before one node issues a lock request to a resource managing node, the node checks the register group to see if the resource managing node is unlocked. With the target node found to be accessible, the access requesting node sends to a broadcast message exchange circuit a broadcast request message including a lock request regarding the resource managing node. The broadcast message exchange circuit receives such broadcast request messages from access requesting nodes, and changes them serially into broadcast messages for broadcast to all nodes. Of these broadcast messages, the first message received by each node is processed by its lock control circuit so that the lock request in that message is allowed to lock the resource managing node. The lock control circuit writes the number of the access requesting node into the register corresponding to the resource managing node. The access requesting node checks the register contents to see if the lock request it issued has been successfully accepted.

Exclusive control method with each node controlling issue of an

Parallel processors communicate with each other over a network by transmitting messages that include destination processor information. A message controller for each processor in the network receives the messages and checks for faults in the message, particularly in the destination processor number contained in a first word of the message. If a fault occurs in the destination processor number, then the faulty message is transmitted to an appropriate processor for handling the fault. In this way the network operation is not suspended because of the fault and the message is not left in the network as a result of the error occurring in the destination processor number. The processor to which the faulty message is directed is determined by a substitute destination processor number contained in the message or is predetermined and set in another way, such as by a service processor. To recover from the fault, the processor receiving the faulty message can request that the message be retransmitted or the error can be corrected using an ECC, for example. If the faulty message cannot be retransmitted, then the processor or the host processor can request that the job to which the faulty message pertains be canceled by all of the processors executing that job without affecting the simultaneous execution of other jobs by the same processors.

Fault handling and recovery for system having plural processors

A parallel processor system including a plurality of processors. When packets of same destination PE number are inputted from different ports, the destination PE number is added with ID numbers of leading ports of split crossbar switches to which the different input ports belong, respectively, by using respective addition circuits, to thereby determine a transfer destination output port for the packets. A plurality of the split crossbar switches having different numbers of input/output ports are realized by partitioning a crossbar switch. By means of an input port select circuit provided in association with each of the output ports, an output request for the packet from the input port belonging to the split crossbar switch to which the associated output port belongs is accepted, while output requests for the packets from the input ports belonging to the other split crossbar switches are inhibited from being accepted, whereby transfer of broadcast packets are inhibited between the split crossbar switches belonging to a physically same crossbar switch. Such situation can be evaded in which same broadcast packets arrive at one and the same processor a number of times.

Switch circuit comprised of logically split switches for parallel

An operation complete signal and a convergence result signal from each processor are transferred to the X-direction interconnection switches, AND of these signals is obtained in switch units in each interconnection switch, the signal is sent out to all the Y-direction interconnection switches through a synchronizing signal relay switch and the like in each relay switch, AND of these signals is obtained in each switch unit in the interconnection switches, and the result thereof is transferred to each processor through each synchronizing signal relay switch. With this, a logical product of an operation complete signal and a logical product of a convergence result signal from all the processors are sent in parallel to all the processors.

Parallel computing system for synchronizing processors by using partial

A parallel computer network wherein an arbitration circuit for performing arbitrating operation over a plurality of processing requests at the same time at high speed is provided in a crossbar network control circuit to thereby prevent the processing requests not selected from being kept awaited for a long time. The arbitration circuit includes a priority bit change circuit which has a plurality of adders for adding a preset value to the priority information of the each awaited processing request and also has a plurality of comparators for detecting the requests being awaited.

Arbitration circuit capable of changing the priority and arrival time of

In a parallel computer, in order to reduce the overhead of data transmissions between the processes, a data transmission from the virtual space of a process in a certain cluster to the virtual space of a process in other cluster is executed without copying the data to the buffer provided within the operating system. The real communication area resident in the real memory is provided in a part of the virtual space of the process, and an identifier unique within the cluster is given to the communication area. When the transmission process has issued a transmission instruction at the time of data transmission, the cluster address of the cluster in which the transmission destination process exists and the identifier of the communication area are determined based on the name of the transmission destination process. Then, the data is directly transmitted between the mutual real communication areas of the transmission originating process and the transmission destination process. Overhead for the data transmission between the processes can be reduced by avoiding making a copy of the data between the user space and the buffer provided within the operating system at the time of data transmission between the processes.

Inter-processor communication method for transmitting data and processor

Crossbar switches having 2.sup.n +1 ports and computing clusters are arranged so that each crossbar switch is connected to 2.sup.n processors. Auxiliary processors that perform parallel processing administrative functions and input/output functions are arranged at the remainder ports of the crossbar switches. Exchangers are provided to connect each processor and its crossbar switches. Parallel processing may be executed by the 2.sup.n processors independently of processing by the auxiliary processors for speed. One mounting unit is formed of a crossbar switch of one dimension, the processor group connected to that crossbar switch, and all of the crossbar switches of a different dimension that are connected to one of the processors of the one processor group. The parallel processor system is mounted by just combining mounting units with no need for special LSIs or frames or the like on which to mount the crossbar switches and without the interfaces that connect the processor and the network becoming concentrated in one place.

Parallel processor system having computing clusters and auxiliary

An information processing apparatus for executing instructions in parallel includes circuitry which, when a first instruction requesting reading of an operand from a certain address of the main storage or buffer storage has been decoded, detects among instructions in execution the presence of a second instruction requesting writing of an operand held by a register such as a general-purpose register into that address of the main storage without implementing an operation on the operand. If the second instruction has been detected, the invention reads out an operand from the register specified for operand reading by said second instruction before operand writing into the main storage by the second instruction is completed.

Growing community of inventors

Hanno, Japan

Shigeo Takeuchi