JP2009194510A - Priority arbitration system and priority arbitration method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a priority arbitration system and a priority arbitration method which suppress deterioration of latency by long path packets or packets waiting due to contention on a path. <P>SOLUTION: The priority arbitration system is provided with: a plurality of CPUs; a plurality of hardware resources to be accessed from the plurality of CPUs; a plurality of crossbars which perform routing between them; and a plurality of routing tables which store latency information showing latency between the plurality of CPUs and the plurality of hardware resources. When transmitting transmission packets to the destination hardware resources, each of the plurality of CPUs adds the latency information corresponding to the destination hardware resources to the transmission packets in the routing tables of the own CPU. When receiving the plurality of packets, each of the plurality of crossbars executes priority arbitration so as to preferentially transmit a packet whose latency is high to the destination hardware resources. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to packet arbitration.

  There is known a large-scale SMP (Symmetric Multi Processing) computer system in which a plurality of crossbars having arbitration circuits exist on a path from a plurality of CPUs to an access destination shared resource (memory, IO, etc.).

The prior art documents that the applicant has known are listed below.
JP 2005-340959 A Japanese Patent Laid-Open No. 9-321795 Japanese Patent Laid-Open No. 10-135952 JP 11-312093 A

  Access to a shared resource physically located far from the path is performed via a plurality of LSIs. Therefore, there is a drawback that the latency becomes long. Further, when a conflict occurs in the arbitration circuit in the middle of the route, the request may be awaited. In this case, the latency required to complete the request further increases.

  If there is a preceding request in the CPU that requests access to a shared resource located far in the path and a subsequent request issued after completion of the preceding request, the increase in latency of the preceding request This can delay the issuance of requests. As a result, there has been a problem that the performance of the entire computer system is degraded.

  The priority arbitration system according to the present invention is set in a plurality of CPUs, a plurality of hardware resources accessed from the plurality of CPUs, and a storage area corresponding to each of the plurality of CPUs. A plurality of routing tables for storing latency information indicating the latency with each of them, and a plurality of crossbars for routing packets exchanged between a plurality of CPUs and a plurality of hardware resources. Each of the plurality of CPUs corresponds to the transmission destination hardware resource in the routing table corresponding to the own CPU among the plurality of routing tables when transmitting the transmission packet to the transmission destination hardware resource among the plurality of hardware resources. Latency information is added to the transmission packet. Each of the plurality of crossbars, when receiving a plurality of packets, refers to the latency information of each of the received plurality of packets and preferentially transmits a packet having a large latency to the destination hardware resource. Execute.

  In the priority arbitration method according to the present invention, a step of setting a plurality of routing tables for storing latency information indicating latency between the own CPU and each of a plurality of hardware resources in a storage area corresponding to each of the plurality of CPUs; When each of the plurality of CPUs transmits a transmission packet to a destination hardware resource among the plurality of hardware resources, the CPU corresponds to the destination hardware resource in the routing table corresponding to the own CPU among the plurality of routing tables. A step of adding latency information to the transmission packet, and when each of the plurality of crossbars receives the plurality of packets, the destination of the packet having a large latency is preferentially referred to by referring to the latency information of each of the plurality of received packets. Executes the priority arbitration process sent to the hardware resource. Tsu and a flop.

  According to the present invention, a priority arbitration device and a priority arbitration method that suppress a decrease in latency due to a long path packet or a packet waited due to a competition on the path are realized.

  Hereinafter, the best mode for carrying out the present invention will be described. First, the present embodiment will be schematically described.

  In the priority arbitration system and the priority arbitration method of the present embodiment, there are a plurality of crossbars having arbitration circuits on the path from the requester (CPU) to the shared resource to be accessed (hardware resources such as memory and IO). Applies to large scale SMP computer systems.

  A field (latency field) describing a latency value is provided in the header portion of the request packet issued by the requester. A routing table is provided that can obtain the routing information from the destination information of the access request to the shared resource of the destination and the minimum required time (minimum latency). With reference to this routing table, the CPU sets a latency value required for route passage obtained from the latency table in the latency field of the packet header when generating a packet.

  In some cases, a plurality of packets simultaneously use the same path on the packet transfer path, thereby causing a conflict and requiring arbitration. In such a case, a mechanism for preferentially passing a packet having a large latency value required by the packet by comparing the latency information of the packet header is provided. This mechanism improves long path packet latency.

  Further, when a certain packet is temporarily waited due to arbitration contention, a mechanism is provided for adding the waited time to the latency value of the packet header. This mechanism allows the packet to be a higher priority packet than before in the next arbitration. As a result, further latency deterioration in the next arbitration competition of the packet is suppressed.

  Next, a specific configuration for realizing such a priority arbitration system and a priority arbitration method will be described. FIG. 1 shows the configuration of the present embodiment. Referring to FIG. 1, the system to which the priority arbitration system and the priority arbitration method according to the present embodiment are applied includes a plurality of LSIs. This system includes four requester CPUs 1010, CPU 1011, CPU 1110, and CPU 1111 that have a function of sending a request packet, four shared resources IOH 1220, IOH 1221, IOH 1320, and IOH 1321 that are access destinations of requests, and requester CPUs 1010, CPU 1011, The crossbars NC1000, NC1100, NC1200, and NC1300, which are located between the CPU 1110, the CPU 1111 and the shared resources IOH1220, IOH1221, IOH1320, and IOH1321, and form a path through which packets pass are provided.

  Between any two LSIs of these LSIs, there is a bi-directional bundle that bundles two, a unidirectional connection interface that transfers data from one to the other and a unidirectional connection interface that transfers data from the other to the other. Connected via connection interface.

  The requester CPU 1010 and CPU 1011 are connected to the crossbar NC 1000 via a bidirectional connection interface, and a function for outputting an access request for the shared resources IOH 1220 to IOH 1321 to the cross bar NC 1000 as a request packet, and a reply sent as a processing completion notification from the shared resource. It has a function to receive packets.

  Similarly, the requester CPU 1110 and the CPU 1111 are connected to the crossbar NC 1100 via the bidirectional connection interface, and receive a reply packet from the shared resource and a function of outputting an access request packet to the shared resource IOH 1220 to IOH 1321 to the cross bar NC 1100. It has a function.

  The storage areas corresponding to the requester CPU 1010, CPU 1011, CPU 1110, and CPU 1111 include, for each shared resource accessible from the own CPU, routing information used in the crossbar at the time of packet transfer, and packets from the own CPU to the destination shared resource. Routing tables R1010, R1011, R1110, and R1111 that hold minimum latency information (indicating one-way latency in the present embodiment) indicating the minimum required time for transferring. Each requester CPU 1010, CPU 1011, CPU 1110, and CPU 1111 refers to the routing table of its own CPU at the time of request packet output, routing information corresponding to the shared resource of the destination and the shortest latency value in the destination designation field and latency field of the packet header. It has a function to set.

  FIG. 4 shows the configuration of any one of the routing tables R1010, R1011, R1110, and R1111 and the setting operation for the packet header. In the routing table R, routing information and latency information (shortest latency value) are held for each shared resource. Each requester CPU 1010, CPU 1011, CPU 1110, CPU 1111 extracts the corresponding routing information and the shortest latency value by designating the shared resource to be the destination, and these values are respectively stored in the destination designation field F 1 and the latency field F 2 of the packet header. Can be set.

  The shared resources IOH1220 and IOH1221 are connected to the crossbar NC1200 via a bidirectional connection interface, and have a function of receiving request packets from all requester CPUs via the crossbar NC1200. Similarly, the shared resources IOH 1320 and IOH 1321 are connected to the crossbar NC 1300 via a bidirectional connection interface, and have a function of receiving request packets from all requester CPUs via the crossbar NC 1300.

  The shared resource IOH 1220 to 1321 that has received the request packet analyzes the request packet from the requester CPU, and performs appropriate processing (such as reading and writing on the shared resource). Thereafter, the shared resource IOH 1220 to 1321 generates a reply packet for the requester CPU as the transmission destination CPU and sends it to the crossbar to be connected.

  The shared resources IOH 1220 to 1321 internally have hardware resource side routing tables R1220, 1221, 1320, and 1321 having the same functions as the routing table of the requester CPU, and have the same functions as the request packet generation by the requester CPUs 1010 to 1111. . That is, the shared resource IOH 1220 to 1321 sets the requester requester CPU as the transmission destination CPU when the reply packet is output. The shared resources IOH 1220 to 1321 set return route routing information corresponding to the destination CPU and return route latency information indicating the shortest latency value in the destination designation field and latency field of the packet header of the return packet.

  The crossbar NC1000 is connected to the adjacent crossbars NC1100 and NC1200 and the requester CPU1010 and CPU1011 via a bidirectional connection interface, and has a function of inputting and outputting packets between LSIs. A port number is assigned to each connection interface in the crossbar NC1000. The connection interface to the crossbar NC 1100 is assigned to port 2, the connection interface to the crossbar NC 1200 is assigned to port 3, the connection interface to the requester CPU 1010 is assigned to port 0, and the connection interface to the requester CPU 1011 is assigned to port 1.

  Similarly, the crossbar NC 1100 is connected to the adjacent crossbars NC1000 and NC1300 and the requester CPU 1110 and CPU 1111 via a bidirectional connection interface, and has a function of inputting and outputting packets between LSIs. Each connection interface in the crossbar NC 1100 is assigned a port number as in the crossbar NC1000. The connection interface to the crossbar NC1000 is assigned to port 3, the connection interface to the crossbar NC1300 is assigned to port 2, the connection interface to the requester CPU 1110 is assigned to port 0, and the connection interface to the requester CPU 1111 is assigned to port 1.

  Similar to the crossbar described above, the crossbar NC1200 is connected to the adjacent crossbars NC1000 and NC1300 and the shared resources IOH1220 and IOH1221 via a bidirectional connection interface, and has a function of inputting and outputting packets between LSIs. Yes. Each connection interface in the crossbar NC 1200 is assigned a port number in the same manner as the crossbar described above. The connection interface with the crossbar NC1000 is assigned with port 2, the connection interface with the crossbar NC1300 is assigned with port 3, the connection interface with the shared resource IOH 1220 is assigned with port 1, and the connection interface with the shared resource IOH 1221 is assigned with port 0.

  Similar to the crossbar described above, the crossbar NC1300 is connected to the adjacent crossbars NC1100 and NC1200 and the shared resources IOH1320 and IOH1321 via a bidirectional connection interface, and has a function of inputting and outputting packets between LSIs. Yes. Each connection interface in the crossbar NC 1300 is assigned a port number similarly to the crossbar described above. The connection interface with the crossbar NC 1200 is assigned with port 2, the connection interface with the crossbar NC 1100 is assigned with port 3, the connection interface with the shared resource IOH 1320 is assigned with port 1, and the connection interface with the shared resource IOH 1321 is assigned with port 0.

  Each of the crossbars NC1000 to NC1300 does not distinguish between a request packet and a reply packet, and operates for the purpose of transferring the packet to the destination IOH 1220 to 1321 or the CPUs 1010 to 1111 according to the routing information existing in the packet header. Each of the crossbars NC1000 to NC1300 has a packet reception unit and an output arbitration unit for each connection interface. Packets received by the crossbars NC1000 to NC1300 from the connection interface are temporarily held by the receiving unit. The packet output from the arbitration control unit is received by another LSI via the connection interface.

  FIG. 2 shows the operation of the packet receiver and output arbitrator in the crossbar. In the crossbar configuration of FIG. 1 that is the configuration of the present embodiment, there are four ports per crossbar. Therefore, in practice, a diagram of four packet receivers and four output arbitration units is appropriate, but in FIG. 2, this is simplified, and two packet receivers and one output arbitration unit are represented. Yes.

  Each packet receiver includes a reception buffer and a destination analysis circuit C1. The reception buffer once holds the received packet. The destination analysis circuit C1 analyzes the destination designation field of the packet header of the packet held in the reception buffer, determines the output destination port of the packet, and outputs an arbitration request to the output arbitration circuit C2 of the output destination port. To do. Each packet receiving unit also has a function of outputting an arbitration request to the output arbitration circuit C2 of the destination port determined by the destination analysis circuit C1, and simultaneously outputting the value of the latency field of the packet header.

  Each packet receiving unit has a function of taking out a packet from the reception buffer and sending the packet to the output port when receiving a use permission signal (GNT signal) from the output arbitration unit.

  Each packet receiving unit has a function of rewriting the latency value according to the result of arbitration for each packet, as will be described below. If each packet receiver outputs an arbitration request to the output arbitration circuit C2 of the output destination port and does not immediately output a GNT signal, it waits until it receives the GNT signal without taking out the packet from the reception buffer ( Waiting for arbitration). During this time, each packet receiving unit adds a waiting time to the latency field value of the packet header and updates the latency field. That is, when the packet receiving unit is in an arbitration waiting state, the latency field of the header is continuously added during that period, and the value is continuously output to the output arbitration unit as a latency field value.

  The output arbitration unit in the crossbar has an arbitration circuit C2. When the arbitration circuit C2 receives an arbitration request from the packet reception unit of each port in the crossbar, the arbitration circuit C2 determines only one use-permitted port from the arbitration request, and sends a GNT as an output port use permission notification to the port. Output a signal. The arbitration circuit C2 has a priority arbitration function that permits the use of the output port based on the latency field value in the packet header output from each packet receiver. With this function, when the use permission port is determined, the use of the output port is preferentially permitted for an arbitration request having a large latency value.

  Further, the output arbitration unit in the crossbar has a packet transmission unit having a function of receiving a packet from the use permission port determined by the arbitration circuit C2 and sending it to the connection interface.

  Next, the operation of the system having the above configuration will be described. FIG. 3 shows an operation example of this embodiment. FIG. 3 is substantially the same as the configuration of FIG. 1, and the numbers of the elements are only in the thousands range from the thousands. The part that doesn't appear in the explanation of operation was deleted. Hereinafter, with reference to FIG. 3, an example of the priority arbitration operation in the long-path packet transfer and the priority arbitration operation for the packet delayed due to the arbitration conflict will be described.

[Assumptions in explanation of operation example]
In this operation example, the path from the requester CPU 3010 to the shared resource IOH 3320 as the transmission destination hardware resource is via the crossbar 3000, the crossbar 3200, and the crossbar 3300 (path 1).
In this operation example, the path from the requester CPU 3011 to the shared resource IOH 3220 is via the crossbar 3000 and the crossbar 3200 (path 2).
In this operation example, the path from the requester CPU 3110 to the shared resource IOH 3220 is via the crossbar 3100, the crossbar 3300, and the crossbar 3200 (path 3).

The shortest latency value of the path is set such that the passing latency per crossbar is 10 in the operation example of the present embodiment. That is, the shortest latency value of the route 1 and the route 3 is 30, and the shortest latency of the route 2 is 20.
In the operation example of the present embodiment, the arbitration waiting time is uniformly set to 15.
Note that the routing information set in the packet header is equivalent to general routing information, although it will not be described in detail in the operation description of the present embodiment. The routing information is treated as including information that can be determined by the output destination port by analyzing in the crossbar.

[Priority arbitration in long-path packet transfer]
The requester CPU 3010 generates a request packet for the shared resource IOH 3220 and sends it to the crossbar NC3000. At this time, the requester CPU 3010 obtains the routing information to the shared resource IOH03220 and the shortest latency value (= 30) from the routing table of the own CPU, and sets them in the destination designation field and the latency field of the packet header. The request packet from the requester CPU 2010 is received by the packet receiver at port 0 of the crossbar NC3000. The packet receiver of port 0 of the crossbar NC3000 refers to the destination designation field in the packet header of the received packet, determines that the output destination is port 3, and sends an arbitration request to the output arbitrator of port 3 of the crossbar NC3000. Output. At the same time, the latency field value (= 30) of the request packet from the requester CPU 3010 is also output to the output arbitration unit of the port 3 of the crossbar NC3000.

  At the same time that the packet receiver at port 0 of the crossbar NC3000 receives a request packet from the requester CPU 3010, the packet receiver at port 1 of the crossbar NC3000 receives a request packet for the shared resource IOH3220 from the requester CPU 3011. The packet receiver of port 1 of the crossbar NC3000 refers to the destination designation field in the packet header of the received packet, determines that the output destination is port 3, and sends an arbitration request to the output arbitrator of port 3 of the crossbar NC3000. Output. At the same time, the latency field value (= 20) of the request packet from the requester CPU 3011 is also output to the output arbitration unit of the port 3 of the crossbar NC3000.

  The output arbitration unit of the port 3 of the crossbar NC3000 executes the priority arbitration process as follows. The output arbitration unit simultaneously receives an arbitration request from the port 0 packet reception unit of the crossbar NC3000 and the port 1 packet reception unit of the crossbar NC3000. At this time, the output arbitration unit of port 3 of the crossbar NC3000 compares the latency field values received from port 0 and port 1 (port 0: port 1 = 30: 20) and gives priority to the request on the port 0 side having a larger value. Therefore, the GNT signal is output to port 0 of the crossbar NC3000.

  The packet receiving unit of port 0 of the crossbar NC3000 that has received the GNT signal from the output arbitration unit of port 3 of the crossbar NC3000 sends the packet to port 3 of the crossbar NC3000. At this time, since there is no waiting time due to arbitration, the latency field of the packet header is not added.

  The request packet from the requester CPU 3010 output from the port 3 of the crossbar NC 3000 is received by the packet receiving unit of the port 2 of the crossbar NC 3200 and sent to the output arbitration unit of the port 3 of the cross bar NC 3200. Here, since there is no contention with other packets, the packet passes without arbitration. Since there is no waiting in arbitration, the latency field value of the packet header is not added.

  The request packet from the requester CPU 3010 sent from the port 3 of the crossbar NC 3200 is received by the packet receiving unit of the port 2 of the crossbar NC 3300 and outputted to the output arbitration unit of the port 1 of the crossbar NC 3300. Again, since there is no contention with other packets, the packet passes without arbitration. Since there is no waiting in arbitration, the latency field value in the packet header is not added.

  The total latency in the transfer path of the request packet from the requester CPU 3010 that has reached the shared resource IOH 3320 from the port 1 of the crossbar NC 3300 matches the value of the latency field of the packet header. Therefore, in the case of this operation example, the packet passes through the path to the shared resource IOH 3320 with the shortest latency (latency field value = 30).

  In this operation example, there was only one conflict in the output arbitration unit in the transfer path of the request packet from the requester CPU 3010. However, when contention occurs in all crossbars on this path, mediation contention occurs three times in total, and priority mediation is performed by comparing the latency values in each mediation unit. In this case, if the competitor packet is a packet that has not waited for arbitration, there is no packet having priority over the request packet from the requester CPU 3010 that is the longest path in this configuration example. Therefore, even in the case of contention, there is a high possibility that the packet can pass with the shortest latency.

[Priority arbitration of packets delayed due to arbitration conflict]
The requester CPU 3010 generates a request packet for the shared resource IOH 3220 and sends it to the crossbar NC3000. At this time, the requester CPU 3010 obtains the routing information and the shortest latency value up to the shared resource IOH03220 from the routing table R, and sets them in the destination designation field F1 and the latency field F2 of the packet header.

  The request packet from the requester CPU 3011 is received by the packet receiver of port 1 of the crossbar NC3000. The packet reception unit of port 1 of the crossbar NC3000 refers to the destination designation field F1 in the packet header of the received packet, determines that the output destination is port 3, and requests an arbitration request to the output arbitration unit of port 3 of the crossbar NC3000. Is output. At the same time, the value = 20 in the latency field F2 of the request packet from the requester CPU 3011 is also output to the output arbitration unit of the port 3 of the crossbar NC3000.

  At the same time that the packet receiver at port 1 of the crossbar NC3000 receives a request packet from the requester CPU 3011, the packet receiver at port 0 of the crossbar NC3000 receives a request packet (latency field value = 30) from the requester CPU 3010 to the shared resource IOH3320. Receive. The packet receiving unit at port 0 of the crossbar NC 3000 outputs an arbitration request to the output arbitrating unit at port 3 of the crossbar NC 3000, similarly to the packet receiving unit of port 1.

  The output arbitration unit of the port 3 of the crossbar NC3000 simultaneously receives the arbitration request from the packet reception unit of the port 0 of the crossbar NC3000 and the packet reception unit of the port 1 of the crossbar NC3000. At this time, the output arbitration unit of the port 3 of the crossbar NC3000 compares the latency field values received from the port 0 and the port 1 (port 0: port 1 = 30: 20), and gives priority to the request on the port 0 side having a larger value. The GNT signal is output to port 0 of the crossbar NC3000. The output arbitration unit of the port 3 of the crossbar NC 3000 suppresses the output of the GNT signal to the port 1 until the packet output from the packet reception unit of the port 0 to the port 3 is completed. Meanwhile, the packet receiver of port 1 of the crossbar NC3000 enters the arbitration wait state.

  The packet receiving unit of the port 1 of the crossbar NC 3000 continues to add the arbitration waiting time to the latency field of the packet header of the request packet from the CPU 3011 during the arbitration waiting state. Further, the added latency field value is continuously output to the output arbitration unit of the port 3 of the crossbar NC3000 so as to be used as the latency value for the next arbitration.

  The output arbitration unit of port 3 of the crossbar NC3000 outputs a GNT signal to the packet reception unit of port 1 of the crossbar NC3000 when the packet reception unit of port 0 of the crossbar NC3000 completes the output of the packet. Having received the GNT signal from the output arbitration unit of the port 3 of the crossbar NC3000, the packet reception unit of the port 1 of the crossbar NC3000 adds the waiting time (= 15) due to the arbitration to the latency field of the packet header. As a result, the packet is output to the port 3 of the crossbar NC 3000 with the latency value = 35.

  The request packet from the requester CPU 3011 output from the port 3 of the crossbar NC3000 is received by the packet receiver of the port 2 of the crossbar NC3200. The packet receiver of port 2 of the crossbar NC 3200 analyzes the packet header of the received request packet from the requester CPU 3011 and sends an arbitration request and latency value = 35 to the output arbitration unit of port 1 of the crossbar NC 3200 connected to the shared resource IOH 3220. Output.

  At the same time that the packet receiver of port 2 of the crossbar NC 3200 receives the request packet from the requester CPU 3011, the packet receiver of port 3 of the crossbar NC 3200 receives a request packet (latency field value = 30) from the requester CPU 3110 to the shared resource IOH 3220. Receive. The packet reception unit at the port 3 of the crossbar NC 3200 outputs an arbitration request to the output arbitration unit at the port 1 of the crossbar NC 3200, similarly to the packet reception unit of the port 2.

  The output arbitration unit of port 1 of the crossbar NC 3200 simultaneously receives requests from the packet reception unit of port 2 of the crossbar NC 3200 and the packet reception unit of port 1 of the crossbar NC 3200.

  Due to the configuration of the present embodiment, the requester CPU 3110 present in the packet receiver of port 3 of the crossbar NC 3200 is more than the distance from the request packet from the requester CPU 3011 present in the packet receiver of port 2 of the crossbar NC 3200 to the shared resource IOH 3220. The distance from the request packet to the shared resource IOH 3220 is longer. For this reason, the request packet from the requester CPU 3110 is a higher priority request at the minimum latency value. However, in the case of this embodiment, the latency value of the request packet from the requester CPU 3011 is added to the time waited by the crossbar NC3000. Therefore, in the output arbitration unit of the port 1 of the crossbar NC 3200, the port 2 of the crossbar NC 3200 is determined as a high priority port, and the GNT signal for the port 2 is output first.

  The packet reception unit of port 2 of the crossbar NC 3200 that has received the GNT signal from the output arbitration unit of port 1 of the crossbar NC 3200 sends the packet to port 1. At this time, since there is no waiting time due to arbitration, the value of the latency field of the packet header is not added.

  The total latency of the request packet sent from the port 1 of the crossbar NC 3200 and received from the requester CPU 3011 that has reached the shared resource IOH 3220 matches the value of the latency field of the packet header. Therefore, the total latency is a value obtained by adding the time waited by the crossbars NC3000 and NC3200 to the shortest latency. In the case of the present embodiment, the arbitration waiting time is only the waiting time 15 in the crossbar NC3000, and since the packet has passed preferentially in the crossbar NC3000, there is no waiting time, and the packet can pass with a result latency value = 35. As a result, a decrease in latency after the occurrence of waiting for arbitration at the crossbar NC3000 is suppressed.

  With the mechanism described above, a request packet having a long transfer path can preferentially pass through the arbitration unit in the crossbar. In addition, a packet that has been waiting due to a route conflict may be able to pass through the route preferentially in the subsequent conflict.

  From these effects, it is possible to suppress a decrease in latency of a long-path packet that can be a bottleneck and a packet that is waited due to competition on the path. As a result, it becomes possible to improve the performance of a large-scale SMP computer system.

1 shows a configuration of a system to which a priority arbitration system and a priority arbitration method are applied. It is a figure for demonstrating the structure and operation | movement of a crossbar. It is a figure for demonstrating transfer of a packet. A routing table and a packet header are shown.

Explanation of symbols

CPU 1010, 1011, 1110, 1111 Requester R1010, 1011, 1110, 1111, 1220, 1221, 1320, 1321 Routing table NC1000, 1100, 1200, 1300 Crossbar IOH1220, 1221, 1320, 1321 Shared resource C1 Destination analysis circuit C2 Arbitration circuit CPU3010 , 3011, 3110 Requester R3010, 3011, 3110, 3220, 3320 Routing table IOH3220, 3320 Shared resource R Routing F1 Destination designation field F2 Latency field

Claims (10)

Multiple CPUs;
A plurality of hardware resources accessed from the plurality of CPUs;
A plurality of routing tables that are set in a storage area corresponding to each of the plurality of CPUs, and store latency information indicating latency between the own CPU and each of the plurality of hardware resources;
A plurality of crossbars for routing packets exchanged between the plurality of CPUs and the plurality of hardware resources;
When each of the plurality of CPUs transmits a transmission packet to a transmission destination hardware resource among the plurality of hardware resources, the transmission destination hardware in the routing table corresponding to the own CPU among the plurality of routing tables. Adding the latency information corresponding to the resource to the transmission packet;
Each of the plurality of crossbars, when receiving a plurality of packets, refers to the latency information of each of the received plurality of packets and preferentially transmits a packet having a large latency to the destination hardware resource. Priority arbitration system that executes priority arbitration processing. The priority arbitration system according to claim 1,
The latency information indicates one-way latency from each of the plurality of CPUs to each of the plurality of hardware resources,
Each of the plurality of hardware resources stores return path latency information indicating one-way latency to each of the plurality of CPUs when a reply packet is returned to a destination CPU of the plurality of CPUs. With reference to the routing table, the return path latency information corresponding to the destination CPU is added to the reply packet,
When each of the plurality of crossbars receives a plurality of packets from the plurality of hardware resource sides, the plurality of crossbars refer to the return latency information of each of the received plurality of packets and give priority to a packet having a large latency. A priority arbitration system that transmits to the destination CPU. The priority arbitration system according to claim 1 or 2,
Each of the plurality of routing tables stores routing information indicating a route of packet transfer between a CPU corresponding to its own routing table among the plurality of CPUs and each of the plurality of hardware resources. system. The priority arbitration system according to claim 3,
The routing information indicates a route from each of the plurality of CPUs to each of the plurality of hardware resources;
Each of the plurality of hardware resource side routing tables includes a packet transfer route from the hardware resource corresponding to the own hardware resource side routing table of the plurality of hardware resources to each of the plurality of CPU resources. Priority arbitration system that stores return routing information to indicate. The priority arbitration system according to any one of claims 1 to 4, wherein
Each of the plurality of crossbars updates the latency information of the received packet to a value obtained by adding a waiting time by the priority arbitration process. Setting a plurality of routing tables for storing latency information indicating latency between the CPU and each of the plurality of hardware resources in a storage area corresponding to each of the plurality of CPUs;
When each of the plurality of CPUs transmits a transmission packet to a transmission destination hardware resource among the plurality of hardware resources, the transmission destination hardware in the routing table corresponding to the own CPU among the plurality of routing tables. Adding the latency information corresponding to a resource to the transmission packet;
When each of the plurality of crossbars receives a plurality of packets, it refers to the latency information of each of the received plurality of packets and preferentially transmits a packet having a large latency to the destination hardware resource. A priority arbitration method comprising: executing a priority arbitration process. The priority arbitration method according to claim 6, wherein
The latency information indicates one-way latency from each of the plurality of CPUs to each of the plurality of hardware resources,
Further, when each of the plurality of hardware resources returns a reply packet to a destination CPU of the plurality of CPUs, hardware for storing return path latency information indicating one-way latency to each of the plurality of CPUs. Referring to a resource side routing table, adding the return path latency information corresponding to the destination CPU to the reply packet;
When each of the plurality of crossbars receives a plurality of packets from the plurality of hardware resources, the packet having a large latency is preferentially referred to by referring to the return path latency information of each of the received plurality of packets. A priority arbitration method comprising: transmitting to the destination CPU. The priority arbitration method according to claim 6 or 7, wherein
Each of the plurality of routing tables stores routing information indicating a route of packet transfer between a CPU corresponding to a self routing table of the plurality of CPUs and each of the plurality of hardware resources,
Further, each of the plurality of CPUs adds the routing information corresponding to the transmission destination hardware resource to the transmission packet in a routing table corresponding to the CPU among the plurality of routing tables;
Each of the plurality of crossbars includes a step of transferring a received packet to a routing destination set in the routing information. The priority arbitration method according to claim 8, wherein
The routing information indicates a route from each of the plurality of CPUs to each of the plurality of hardware resources;
Each of the plurality of hardware resource side routing tables includes a packet transfer route from the hardware resource corresponding to the own hardware resource side routing table of the plurality of hardware resources to each of the plurality of CPU resources. Store the return routing information shown,
Further, each of the plurality of hardware resources adds the return route information corresponding to the transmission destination CPU to the return packet in a return route routing table corresponding to the own hardware resource among the plurality of return route routing tables. ,
Each of the plurality of crossbars includes a step of transferring the received reply packet to a routing destination set in the return path routing information. A priority arbitration method according to any one of claims 6 to 9, wherein
Further, each of the plurality of crossbars updates the latency information of the received packet to a value obtained by adding a waiting time by the priority arbitration process.

Images