WO2010137572A1

WO2010137572A1 - Network-on-chip, network routing method, and system

Info

Publication number: WO2010137572A1
Application number: PCT/JP2010/058780
Authority: WO
Inventors: 淳鳥居
Original assignee: 日本電気株式会社
Priority date: 2009-05-25
Filing date: 2010-05-25
Publication date: 2010-12-02
Also published as: JPWO2010137572A1; JP5083464B2

Abstract

Disclosed is a method which lowers latency in a network-on-chip, reduces costs, and increases reliability. Uniform routers of a plurality of routers existing in the system are grouped into two classes: main routers (4) and sub routers (3). Through the plurality of sub routers (3), a main router (4) handles all of the routing calculation up to the next main router or to the destination network interface. As the sub routers (3) do not need to perform routing calculation, logic can be omitted, latency when packets are passing the sub routers can be reduced, and cost can be lowered. Further, the reliability of the main routers (4) can be increased, and the sub routers (3) become able to respond to faults, when generated, by sending packets to the closest main router at the time of the fault, thereby increasing reliability.

Description

Network-on-chip and network routing methods and systems

[Description of related applications]
The present invention is based on the priority claim of Japanese patent application: Japanese Patent Application No. 2009-125580 (filed on May 25, 2009), the entire contents of which are incorporated herein by reference. Shall.
The present invention relates to a semiconductor integrated circuit device and a network routing method, and in particular, a data communication network between IP cores in an LSI integrated circuit (System On a Chip (SoC)) in which a plurality of IP (Intellectual Property) cores are integrated. The present invention relates to an apparatus and a method suitable for application to a chip (Network on a Chip: NoC).

Conventionally, in an LSI integrated circuit (SoC: System On a Chip) in which a plurality of IP (Intellectual Property) cores are integrated, connection between IP cores is performed using a bus. For example, AMBA (registered trademark of ARM) proposed by ARM has become an industry standard for such buses. This AMBA has evolved as AHB (Advanced High performance Bus), Multi-layer AHB, AXI (Advanced eXtensible Interface) in response to the increase in the number of cores and the complexity of chips.

However, in the bus configuration, the wiring length becomes long and it is difficult to cope with the improvement of the frequency. In addition, it is necessary to centralize arbitration against access competition from a plurality of IPs. As a result, problems such as the need to recreate the bus have arisen with respect to fluctuations in the number of IPs.

Currently, a method of dealing with these problems by providing a locality by hierarchically connecting buses has been actively performed in a high-performance SoC.

On the other hand, NoC (Network on Chip), which can ensure scalability by distributing arbitration to each router while effectively utilizing the locality of communication by packetizing and sending data, is the University of Stanford University, W. J. et al. The following non-patent document 1 has been proposed and has been actively researched in recent years. Examples of actual chips adopted include TILEERA's TILE64, Intel's 80 core chip, CEA-Let's FAUST chip, STMicro's STNoC, and the like.

As shown in FIG. 15, these basically constitute a network by a router 302 and a communication channel 305. By sending a packet composed of a plurality of flits as shown in FIG. 16 to these networks, Realize communication.

Here, each router 302 sends the received packet to an appropriate output port for sending to the destination network interface (NIF).

At this time, after looking at the address and ID of the destination in the router 302, based on a certain rule, a routing protocol method for determining a destination for the next hop,
In the network interface on the sender side, all the routing to the destination is determined, and the source routing method that includes the routing information in the packet,
Exists.

<Routing protocol method>
The routing protocol method analyzes a packet that has entered from an input channel, extracts an address and ID, searches a table provided in the router based on this information, and determines which output channel the packet is output to. After that, arbitration is performed for the output channel, and a packet is output.

Therefore, several steps of logical operations are required to determine the output destination.

Also, when changing the routing, it is necessary to update the routing rules in all routers involved in the change.

As a technique for reducing the number of steps required for routing, a “Next Routing Computation” method in which routing calculation is performed by the previous router has been proposed. Taking FIG. 17 as an example, when there is a packet from the router R0 to the router R5 via the router R1, the routing calculation of the router R1 is performed by the router R0 and the routing calculation of the router R5 is performed by R1. is there.

Furthermore, a method of virtually setting up an Express channel, a method of performing predictive routing, and the like have been proposed. For example, the following

non-patent documents

2 and 3 are examples.

In the method of Non-Patent Document 2, as shown in FIG. 17, the number of pipeline stages required in the relay router is reduced by establishing a virtual channel that bypasses routing calculation and arbitration by several relay routers. When bypassed, it passes through the router in two cycles.

In FIG. 17, routers 351 are connected by a bidirectional channel 352. Here, when there are many communications from the router R1 to the router R15, the virtual express channel 354 is defined.

Routers R2, R7, and R11 can omit the routing calculation and switch arbitration of packets arriving on this channel. In this way, the latency from the router R1 to the router R3 and the latency from the router R3 to R15 are shortened.

On the other hand, the method of Non-Patent Document 3 predicts the output direction of a packet speculatively based on a certain prediction rule in parallel with performing the routing calculation for the arrived packet, and switches the direction of the packet. If the speculation is successful by performing arbitration and passing through the switch, the latency is shortened for the cycle of the routing calculation.

If this prediction is not correct as a result of the routing calculation, the flit of the packet existing in the output latch, the input latch in the router in the direction of speculative output, or the input data FIFO (First In First Out) is canceled and correct. Perform switch arbitration and switch passage for the output direction again.

<Source routing method>
In the source routing method, the output channel information in each router is embedded in a packet in advance at the network interface on the transmission side, and each router can complete the routing simply by extracting this information.

For example, Intel's 80-Tile 1.28TFLOPS Network-on-Chip adopts this source routing method (Non-patent Document 4). FIG. 18 shows a flit 0 of the packet in this chip. As shown in FIG. 18, in this chip, 10 hops of DID (Destination IDs) 361 indicating the output destination for each router are generated by the output side network interface and defined in the packet. According to this, the router determines the output port.

Therefore, each router can perform routing only by extracting the DID value. As described above, the source routing method can simplify the structure of each router.

The source routing method is supported as an option even in a LAN (Local Area Network) between computers such as TCP / IP (Transport Control Protocol / Internet Protocol).

This specifies the physical location of the network through which the packet passes in the packet as an IP address and routes the packet accordingly. There are two types, a strict source routing method and a loose source routing method.

<Strict source routing method>
In the strict source routing method, all network points to be hopped are specified in a packet at the network interface on the transmission side, and a packet that reaches a network point not specified is discarded.

<Loose source routing method>
On the other hand, in the loose source routing method, only a part of the network points are specified in the packet, and the route supplemented by the router is allowed in the meantime, and the routing protocol method is used except for the specified network points. Then, routing is performed at each network point and packet transfer is performed.

In any case, the maximum number of hops of the router that can be specified is 9. These are often used for router testing.

FIG. 19 shows a packet format in the case of this source routing. In FIG. 19, when the option type 381 is a designated option, the source routing method is used, and source routing for the IP address designated by the option length can be performed.

The option length 382 is followed by an option pointer 383 and an IP address 384 corresponding to the designated option length. Since n is 9 at the maximum, it is not possible to specify an IP address of 10 or more.

Patent Document 1 discloses a route repair method capable of repairing a route with a small load in a large-scale ad hoc network having a large number of hops between a transmission source and a transmission destination. A route for relaying the nine wireless nodes B, C... J is established between the source node A and the destination node K, and the management nodes A, F, and K are provided for each predetermined number of hops on the route. Has been placed. When the route between the wireless nodes I and J disappears, the wireless node I that has detected this transmits an RERR (route invalid notification) with the nearest management node F as the destination in the source direction. The management node F that has received this RERR transmits a Repair_REQ (path repair request) destined for the nearest management node (here, the destination node K) in the destination direction to perform local path repair, A new route that bypasses the lost route is established between the two management nodes F and K. Patent Document 2 discloses an electronic device (system-on-system) having interconnection means (N) for connecting a plurality of modules (IP; AD, M) and enabling communication between the modules (IP; AD, M). Chip) and communication resources are related to time division multiple access based on time slots that divide and share the available communication band, and the electronic device has at least one of a plurality of modules (IP; AD, M) And at least one network interface coupled to the interconnection means (N), the network interface (NI) being adapted to establish at least one connection to at least one further network interface (NI); At least one connection has at least one channel (ad). At least one network interface (NI) has at least one slot table (ST1-ST11) that reserves time slots of at least one channel (a-d). A configuration is disclosed in which time slots are shared between channels (a-d) associated with the same network interface (NI).

JP 2008-283260 A Special table 2009-502080

The entire disclosures of

Patent Documents

1 and 2 and Non-Patent Documents 1 to 4 are incorporated herein by reference.
The analysis according to the invention is given below.

Network on chip (NoC) aims to improve efficiency and performance by replacing the system bus with network configuration and packet-based communication.

However, in the routing based on the conventional technique in which each router performs a routing calculation, the delay and area of each router increase due to the routing calculation. Specifically, a table and arithmetic logic for determining the network output destination from the ID and address are required.

When the ID increases, the amount of the table, the power of the table pull, and the delay increase.

When the delay increases, it is necessary to decrease the network frequency and decrease the throughput or increase the number of pipeline stages.

The increase in the number of pipeline stages causes a problem that the required amount of the buffer for storing packets increases during the routing calculation.

Also, various proposals for suppressing an increase in routing latency also increase the complexity of the router and increase the amount of material.

Furthermore, in the case of these methods, when routing is changed due to a failure or network congestion, it is necessary to rewrite the table contents of each router or the operation logic.

On the other hand, in the case of the source routing method, the router can be simplified and speeded up, but the number of router hops that pass from the sending side to the receiving side is not constant, so the area for storing routing information in the packet Need to be secured with a variable length.

Or, if the fixed length is used, the number of useless fields increases in communication with a small number of hops, and if this field is restricted, the number of hops is limited and scalability is hindered.

In addition, when the number of hops increases, the amount of computation for calculating the routing at the source network interface first increases, the amount of delay increases, or if there is a situation during the route, specify it first. It is difficult to change the routed routing, and the applicability to dynamic routing is inferior.

Therefore, an object of the present invention is to provide a method, apparatus, and system for achieving low latency, low cost, and high reliability of a network-on-chip.

According to the present invention, a plurality of information processing units, storage units, and interface units are integrated on one or a plurality of chips, and a network including a network interface, a communication path, and a network router is integrated between the units. Information processing semiconductor element that performs information processing by exchanging information between packets via the network, and the network router is divided into a plurality of groups of a main router group and other router groups, and the number of times it passes through the router If the number of hops that can not reach the destination within the specified number of times, pass through the main router, each time through the main router, reset the number of times through the router, and perform routing to satisfy the specified number of times, A network router is provided.

According to the present invention, it is possible to achieve low latency, low cost, and high reliability of the network on chip.

1 is an overall block diagram of a multi-core SoC using an on-chip network according to an embodiment of the present invention. It is a block diagram of the main router in one Embodiment of this invention. It is a block diagram of a subrouter in one embodiment of the present invention. It is a figure which shows the format of the packet in one Embodiment of this invention. It is a figure which shows the format of

flits

0 and 1 at the time of the network interface output of B3 in embodiment of this invention. It is a figure which shows the format of the flit 0 at the time of the main router output of E3 in embodiment of this invention. It is a figure which shows the format of the flit 0 at the time of the main router output of F6 in embodiment of this invention. It is a figure which shows the format of the flit which stored the route information of the packet in embodiment of this invention. It is a figure which shows the format of the flit 0 at the time of the network interface output of B3 in embodiment of this invention. It is a figure which shows the whole structure of the multi-core SoC using the on-chip network in embodiment of this invention. It is a figure which shows the format of the packet at the time of the network interface output of B3 in embodiment of this invention. It is a figure which shows the format of the packet at the time of the main router output of E3 in embodiment of this invention. It is a figure which shows the format of the packet at the time of the main router output of F6 in embodiment of this invention. It is a figure which shows the format of the response packet produced | generated by the network interface of G6 in embodiment of this invention. It is a block diagram of SoC which employ | adopted the network on chip. It is a figure which shows the format of the packet in a network on chip. It is a figure which shows the concept explaining the Express Path of related technology. It is a figure which shows the format of the packet which showed the route information of related technology. It is a packet format figure at the time of source routing of TCP / IP of related technology.

First, the basic principle of the present invention will be described, followed by a specific example.

In the present invention, the routers existing in NoC are divided into two types: main routers and sub-routers.

For network packets, first, the source network interface takes the source routing method to the appropriate main router, or when the adjacent network interface is the destination, the routing up to that point is collectively performed.

Packets that have arrived at the main router are routed to the next main router or the destination network interface in a lump by the source routing method. Through these series of operations, the sub-router performs routing according to the partial routing information performed by the network interface or the main router. As a result, the sub-router is simplified, and the quantity, latency, and power consumption can be reduced.

On the other hand, the routing calculation for the main router is slightly complicated, but in any case, it is necessary to calculate the route for the output from its own router. Does not occur.

Also, it is possible to reduce the overall cost by reducing the ratio of the main router in the sub-router network.

In the present invention, the basics of NoC routing are proposed.

In the network, prepare a network topology in which the main router is arranged within the specified hop count and reachable to another main router, and the others are arranged as sub-routers, and packets are within the specified hop count. Thus, there is provided a routing method that is controlled so as to pass through a path from the main router to the next main router or to the receiving side network interface.

That is, routing that does not pass the main router beyond the specified number of hops is not performed.

In the second aspect of the present invention, the main router collectively performs routing within the specified number of hops to the next main router, and the sub-router processes the packet according to the routing information performed by the main router. Therefore, the sub-router performs routing on the arrived packet based on the routing information performed by the main router. The sub-router does not need to perform routing calculation on its own.

In the third aspect of the present invention, the routing information for the sub-router performed by the main router is written in a specific flit in the packet. The sub-router performs routing by referring to this specific flit. Sub router too. It is possible to modify the routing information of this specific flit as required. The subrouter does not calculate routing.

In the fourth aspect of the present invention, when the routing of the subsequent main router can be calculated by the upstream main router, the downstream main router is based on the routing information of the upstream main router, similarly to the sub router. , Do the routing. This makes it possible to reduce the routing cost, particularly when the IP connected to the main router is the receiving destination.

In the present invention, the limitation on the number of hops results from the restriction of the routing information that can be written in the limited data area of a specific flit in the packet.

When the routing information of each sub-router is expressed as it is in a fixed bit number area, the limit on the number of hops is uniquely determined by how many areas of the fixed bit number can be taken. However, in particular, in an on-chip network that employs a mesh configuration, it is expected that routing with the most straight packets will be performed.

Therefore, in the fifth and sixth aspects of the present invention, by adding the concept of the number of iterations to the routing information, the routing information is compressed, the number of reachable hops is increased, and the number of hops to the main router is decreased. Let

Even if such a device is added, packets may be concentrated on the main router compared to the conventional network-on-chip due to the hop count specification. Therefore, there is a possibility that the performance may deteriorate due to network congestion in the main router near the IP where communication is concentrated.

Therefore, in the seventh aspect of the present invention, in order to cope with these problems, the number of main routers is increased in the vicinity of the IP where communication is concentrated, or a router arranged as a spare main router is replaced with a sub-router or A mechanism for distributing the load using the main router is provided.

In the eighth to tenth aspects of the present invention, network reliability is improved. The on-chip network in the present invention optimizes the entire network by allowing the main router to have a certain autonomous function and the sub-router to operate in a dependent manner.

In the eighth aspect of the present invention, the main router grasps the failure status and congestion status of the nearby network and performs routing according to it.

In the ninth aspect of the present invention, when a subrouter or link failure is detected, the subrouter converts the arrived packet into information in the packet by using an available channel for which no failure is detected. Regardless, it performs routing back to the main router.

In the tenth aspect of the present invention, the main router grasps the situation of the failure and performs avoidance routing based on this.

Also, in the eleventh aspect of the present invention, the reliability of the entire network is maintained by improving the hardware reliability of the main router over that of the sub-router without increasing the redundancy of the entire network. In the eleventh aspect of the present invention, flits are added to the packet every time the routing information for the sub-router passes through the main router. As a result, when the packet is transmitted from the transmission side network interface, only the routing information up to the first main router is retained, but when the packet reaches the reception destination network interface, the transmission side receives the reception destination. All the routing information up to this point is held, and the number of packet flits is increased accordingly.

As a result, the packet is returned to the transmission side without performing a new routing calculation based on this routing information at the time of retransmission processing when a failure is detected, data reception confirmation processing, and response processing to the data read request. It becomes possible.

Such routing can be performed for all packets, or only for some of the packets that are highly necessary.

Further, when returning a packet to the transmission side, in the present invention, the main router also performs packet transfer without performing autonomous routing like the sub router.

<Embodiment 1>
An embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 shows the overall configuration of a multi-core SoC to which one embodiment of the present invention is applied. In this embodiment, an on-chip network configuration having a two-dimensional mesh structure is adopted. In FIG. 1, the multi-core SoC 10 has a plurality of IP (Intellectual Property) cores 1. Each IP core 1 is connected to a communication channel 5 via a network interface 2. The communication channel 5 is used for connection and communication between the network interface 2 and the main router 4 or the subrouter 3. The main router 4 and the sub router 3 are arranged on the multi-core SoC 10 at a certain ratio.

FIG. 2 is a diagram showing a configuration of the main router 4 of FIG. Referring to FIG. 2, the main router includes input latches 21n, 21e, 21s, 21w, 21i that receive input signals from IP (Intelligent Property) directly connected to the four directions of the NESW and the main router itself.

Route calculators

22n, 22e, 22s, 22w, 22i for calculating routes from destinations;
Routing

flit generation units

23n, 23e, 23s, 23w, 23i for generating a routing flit based on the result of route calculation;

Selectors

24n, 24e, 24s, 24w, 24i for selecting one of the results of the routing

flit generation units

23n, 23e, 23s, 23w, 23i and direct information from the input latches 21n, 21e, 21s, 21w, 21i, respectively ,
Input FIFOs (First In First Out Storage) 25n, 25e, 25s, 25w, 25i for storing selection results of the

selectors

24n, 24e, 24s, 24w, 24i,
The output destination of each input signal is arbitrated by the arbiter 27, and the

selector

28n, 28e, 28s, 28w, 28i selects the output from the

input FIFO

25n, 25e, 25s, 25w, 25i in any direction. When,

Output buffers

29n, 29e, 29s, 29w, 29i for latching the selection results by the

selectors

28n, 28e, 28s, 28w, 28i of the crossbar switch 26, respectively;
The latch results in the

output buffers

29n, 29e, 29s, 29w, and 29i are output.

FIG. 3 is a diagram showing the configuration of the sub-router 3 in FIG. Referring to FIG. 3, since the sub-router does not have a routing mechanism, it is greatly simplified as compared with the main router 4 of FIG. That is, the input data is stored in the

direct input FIFOs

45n, 45e, 45s, 45w, and 45i and then reaches the crossbar switch 46. The crossbar switch 46 arbitrates the output destination of each input signal by the arbiter 47, and selects the output from the input FIFO 45 in either direction by each

selector

48n, 48e, 48s, 48w, 48i. The selection result is latched by the

output buffers

49n, 49e, 49s, 49w, and 49i, and then output.

FIG. 4 shows an example of a packet format in this embodiment. Referring to FIG. 4, the packet includes flits 0 to 6. Each frit includes a sideband 82 and a packet body 86. The sideband 82 indicates an attribute of the packet body 86, and includes a valid bit 83, a head bit 84, and a terminal bit 85, and indicates whether the flit data is valid, the head flit of the packet, or the terminal flit. The packet body 86 has different data meanings for each frit.

In this embodiment, flit 0 includes a plurality of route information 87 (route information 1 to 4) for source routing and control information 88 related thereto.

The flit 1 includes a flit length 89 of the entire packet, position information 90 of the target IP, and control information 91 related thereto.

Flit 2 includes an access attribute 92 and an access address 93 of this packet.

In the frit 3 to the frit 6, data 94 is stored.

The route information 87 indicates the direction in which the packet is output when entering the subrouter 5.

In this embodiment, since four pieces of information from route information 1 to route information 4 can be stored, the number of hops of the subrouter passing to the main router or the destination is four at the maximum.

In addition, pointer information indicating which of a plurality of route information 87 should be referred to is stored in the control information 88.

The access attribute 92 includes attributes such as a read request, a read response, and a write request.

Hereinafter, the operation of this embodiment will be described by showing an example of packets in time order. Here, a case will be described in which a packet is sent from IP core 1 located at B3 in FIG. 1 to IP core 1 located at G6.

First, the network interface 2 connected to the IP on the B3 side performs routing to the main router suitable for routing to the destination within 4 hops on the way.

In this embodiment, since the target IP core cannot be reached within 4 hops, it is sent to the main router located at E3 or the main router located at C5. Here, it is assumed that it is sent to E3. In that case, flits 0 and 1 of the packet at the time of output of the network interface 2 of B3 are as shown in FIG.

That is, in FIG. 5, route information 1 to 4 (87) in FIG. 4 is stored in 101 to 104, route information 101 to 103 is in the E direction, and route information 104 is not a valid value. It becomes.

Also, the control information 88 of FIG. 4 is 3 at 105, indicating that there are three valid data and that the route information 103 should be referenced next.

Also, “G6” is stored in the position 107 of the target IP, and 8 is stored in the frit length 106.

Since this packet is the sub-router 3 of B3 and the control information 105 is 3, it refers to the route information 103 and is output to the E side, that is, the C3 side. At this output time point, the control information 105 is subtracted by 1 by the sub-router 3 of B3 to be 2.

In the C3 sub-router, similarly, the control information 105 is 2, so the route information 102 is referred to and output to the E side, that is, D3. The control information 105 is further subtracted by 1 and becomes 1.

The same applies to D3, and the control information becomes 0, and the packet arrives at the main router 4 located at E3.

In the main router 4 located at E3, the route calculation unit 22w decodes that the target position information 90 of the flit 1 is G6 as indicated by 108, and further G6 cannot be reached within 4 hops. Judge.

Therefore, routing to the nearest F6 main router is performed.

In this case, the main router output of E3 is E direction, that is, F3. To advance 3 hops in the S direction, packet 0 is replaced as shown in FIG. That is, since the route information 111 to 113 is in the S direction and the route information 114 is not a valid value, it is Invalid.

Further, the control information 88 (see FIG. 4) is 3 at 115, and the route information 113 is indicated as a pointer. This flit is generated by the routing flit generator 23w (see FIG. 2).

This leads to the main router of F6 via the sub-routers 3 of F3, F4 and F5 as in the previous operation.

In the main router 4 located at F6, the route calculation unit 22n decodes that the target position information 90 (see FIG. 4) of the flit 1 is G6 as indicated by 108, and further, the number of hops is 4 for G6. Judge that it can be reached within.

Therefore, routing to the target position G6 is performed. In this case, the main router output of F6 is the E direction, that is, G6, and packet 0 is replaced as shown in FIG.

That is, the route information 121 indicates that it is taken into the IP core, and since the route information 122 to 124 is not a valid value, it becomes Invalid.

Further, the control information 88 (see FIG. 4) becomes 1 at 125, and the route information 121 is indicated as a pointer. This flit is generated by the routing flit generation unit 23n (see FIG. 2).

Thereby, the sub router 3 of G6 can send the packet to the IP core 1 of G6 based on the route information 121 of the fetched packet. It can be seen that data communication from B3 to G6 becomes possible by the above routing.

At this time, only the network interface of B3 and the main router 4 of E3 and F6 performed the actual routing calculation, and the sub-router 3 of B3, C3, D3, F3, F4, F5, and G6 has flit 0. Routing is performed with reference to only the route information 87. For this reason, it is possible to simplify the router and shorten the delay through the router.

The routing calculation in the main router 4 adds an absolute value obtained by subtracting the coordinates of the main router 4 from the target IP position information 90 for each of the X axis and the Y axis. Since this addition result is the shortest number of hops, it is determined whether or not this addition result exceeds a specified value. If the addition result does not exceed the specified value, the routing information of each sub-router up to this receiving destination is written in the route information 87 of flit 0.

When the addition result exceeds the specified value, the main router closest to the receiving destination is selected from the reachable main router group, and the routing information of the sub-router to this main router is the route information of flit 0. This is realized by writing to 87.

In the examples of FIGS. 5 to 7, only the routing information of the sub-router is written in the field of the route information 87 of flit 0, but it is also possible to have a specification in which the output direction of the main router is written in the same manner. In this case, routing information from the main router to the next main router or reception destination remains in flit 0 in a complete form.

Further, when the main router 4 reads the control information 88 (see FIG. 4) when the packet arrives and the value is not 0, the routing performed by the main router 4 in the previous stage is not completed. Therefore, the main router 4 determines that autonomous routing is not performed, performs routing based on the route information 87 in the flit 0, and transfers the packet to the adjacent main router 4 or sub-router 3.

By adding this determination logic, it is not necessary to perform autonomous routing calculation in all the main routers 4 in the on-chip network, so that the load on the main router 4 can be reduced.

However, since the main router 4 has the capability of performing autonomous routing, the routing including the main router 4 is performed, and the route information 87 ( It is also possible to discard the packet (see FIG. 4), perform routing again, and send a packet after writing new route information 87.

On the other hand, in an on-chip network adopting a two-dimensional mesh or a network topology based on the two-dimensional mesh, the probability that the packet goes straight increases. For this reason, it is possible to write routing information with a larger number of hops in a limited frit area by adding the number of transfers in the same direction as the number of repetitions to the route information by taking advantage of this characteristic.

Claims

5 and 6 are based on this concept.

FIG. 8 shows an example of frit 0 in which such information is described. In FIG. 8, route information 137 and control information 138 are the same as the route information 87 and control information 88 shown in FIG. 4, and information on which direction is output from the router and which part of routing is currently completed. It shows the pointer of whether or not. Here, iteration number information 139 is added.

The iteration count information 139 indicates how many times each route information 137 is repeated. In each of the information on the number of iterations currently performed, the iteration count information 139 may be updated by the sub-router, or the pointer in the control information 138 may have information including the iteration count.

9 is connected to the IP on the B3 side when the present invention is applied when a write packet is sent from the IP core 1 located at B3 in FIG. 1 to the IP core 1 located at G6. This shows a flit 0 of a packet output from the network interface 2.

In the present embodiment, the network interface 2 of B3 performs routing to the main router 4 located at F6 in FIG. 1, and unlike the example shown above, the main router 4 located at E3 According to the invention, autonomous routing is not performed in the same manner as the sub-router.

In FIG. 9, flit 0 is assumed to be 4 hops in the E direction and 3 hops in the S direction after leaving B3 to reach the main router 4 of F6, and a total number of hops of 7 is written in the control information 146. . Since there are two types of traveling directions, 144 and 145 entries of route information 3 and repetition information 3 are respectively Invalid. On the other hand, the route information 2 is first advanced by 4 hops, so that 142 of the route information 2 is E and 143 of the repetition information 2 is 4.

Thus, the routers B3, C3, D3, E3, F3, F4, and F5 transfer the packet to the main router 4 of F6 while subtracting 1 from the control information 146.

Since the control information 146 of the arrived packet is 0, the main router 4 of F6 performs routing, generates flit 0 similar to FIG. 7, and sends the packet in the E direction. As a result, the packet can be transferred to a desired IP.
If, for example, a failure or a mixture occurs in the F3 router or channel when passing through the E3 main router 4, the E3 main router 4 performs a new routing calculation and writes the result in flit 0 to create another route. Of course, it is also possible to perform the transfer.

The number of hops is limited by a unique integer value due to the limitation of the bit information of the route information written to the flit. However, if concepts such as the number of iterations are added to the route information in this way, the number of hops is limited at a high number of iterations. Can be relaxed. For this reason, the upper limit of the number of hops can be set to a range within a certain route information field and repetition number field. Even in this case, it is possible to leave the property of avoiding failure and congestion by performing routing through the main router 4 at a certain frequency.

Here, the number of iterations and the route information field are clearly separated, but the route information may be written in a specific frit by a method with a higher compression rate.

When the routing method according to the present invention is used, the hardware cost, power, and latency of the sub-router 3 can be reduced. However, in order to perform routing via the main router 4, the main router 4 Tend to concentrate on access.

In the present invention, the main router 4 and the sub-router 3 can be allocated within the upper limit of the number of hops, and the main router 4 and the sub-router 3 can be freely arranged otherwise. Accordingly, when there is an IP core 1 where traffic is likely to be concentrated, it is possible to reduce the access to a specific main router 4 by centrally arranging the main router 4 around the IP core 1.

FIG. 10 is a diagram showing the overall multi-core SoC configuration. The network shown in FIG. 10 is the same as that shown in FIG. Therefore, loads on the main router 154 such as B2 and E3 in the vicinity may be concentrated. In order to avoid this, spare main routers 156 are arranged at B3, D4, E1, and F2.

These spare main routers 156 normally operate in the same manner as the sub-routers 153. In other words, since the other main router 154 and the network interface 153 do not generate a packet to which the routing information up to the spare main router 156 is added, autonomous routing is not performed.

However, when access to the IP core 151 located in D2 is concentrated and the channel 155 to the main router 154 located in B2 and E3 is congested, the information is arranged in B3, D4, E1, and F2. To the spare main router 156. In this case, these spare main routers 156 reroute the passing packets within the spare main router 156, and route the packets to D2 without passing through the main router 4 located at B2 and E3. Reroute. As a result, it is possible to reduce the load on the main router 4 of B2 and E3.

For such a load distribution and failure avoidance mechanism, the present invention (Claim 8) includes means for grasping the congestion status of nearby links and sub-routers. For example, the main router 4 uses the dedicated signal line to the main router 4 closest to each subrouter and the subrouter 3 to display the packet passage rate on the link, the packet contention rate of each output channel of the subrouter 3, and the information when the operation is abnormal. To the main router 4 using control packets.

Based on these pieces of information, the main router 4 grasps the packet congestion status and failure status in the vicinity, and performs routing for arriving packets to avoid congestion.

Further, according to the present invention (Claim 4), even when a packet that does not need to be routed autonomously arrives at the main router 4 or the spare main router 6, in such a situation, the route information 87 is discarded. It is also possible to perform routing avoiding a failure or congestion again and send a packet after writing new route information 87.

Also, when the sub-router 3 fails, since it does not have an autonomous routing capability, it cannot be determined where to send a packet from among available channels.

Therefore, when detecting a failure, the sub-router 3 in the present invention (Claim 9) selects a route that can reach the nearest main router 4 with the closest number of hops from the available channels, and flits 0. The control information 88 is sent as 0.

When the router that receives such a packet is the main router 4, the routing is performed again with the target IP location information 90 as it is.

On the other hand, in the case of the sub-router 3, as in the previous sub-router, the route that can reach the nearest main router 4 with the closest hop number is selected from the available channels, and the control information 88 of flit 0 is set to 0. Send out as By repeating this process, the main router 4 is reached, and packet transmission can be resumed with correct routing information.

Thereby, the sub-router 3 can improve the reliability of the network with almost no increase in hardware complexity.

On the other hand, since the main router 4 is required to operate autonomously, as described in the tenth invention, the main router 4 is provided with redundant logic for error or an error correction detection function to In combination with the invention, it is also possible to improve the reliability of the entire on-chip network.

Finally, the present invention (claim 11) will be described. In the invention described so far, the route information is generated by the main router 4 or the network interface 2, and the main router 4 deletes the route information to the main router held by the arrived packet from the flit and creates a new one. The route information is written in the same flit, in this embodiment, flit 0, and the packet is transferred.

In this method, the packet length, that is, the number of flits that one packet has does not change from the transmission source to the reception destination, which is one of the problems of the source routing method. When the number of routing hops increases, the packet length increases. Further, it is possible to solve the problem that the cost of routing on the transmission side increases.

However, for example, for a packet such as a read request, the IP core that has received the request needs to respond with data toward the IP core of the transmission source.

In addition, there are cases where it is necessary to send a response packet back to the IP core of the transmission source in the event of an acknowledgment or an error. If all the route information is stored from the transmission source, the main router 4 or the network interface 2 can omit the routing calculation in the response packet.

The present invention (Claim 11) proposes a method of maintaining all the route information from the transmission source in the packet by adding a flit containing the route information every time the main router 4 performs routing.

FIG. 11 shows the configuration of the packet 201 when the packet is sent from the IP core 1 located at B3 in FIG. 1 to the IP core 1 located at G6, when the network interface 2 located at B3 is sent. . Here, the data stored in the route information 207, the control information 208, and the like are the same as those in FIG. 6, and the processing performed by the network interface 2 is equivalent.

When the packet 201 arrives at the main router 4 located at E3, routing is performed in the same manner, but as shown in FIG. 12, flit 0 at the time of FIG. 11 is moved to flit 2, and the control information 212 is As shown in FIG. 12, a flag indicating that the log is a past routing log is set.

Also, since the flit length of the packet 201 is extended, the flit length bit 209 is incremented by 1 and becomes 8. In addition, in the newly added flit 0, since the output channel from the main router 4 is also added to the route information 207, the value SEEE is stored.

Similarly, in the main router 4 located at F6, as shown in FIG. 13, the frit 2 is moved to the frit 3, the frit 0 is moved to the frit 2, and a new frit 0 is added. 9

Thus, when the packet reaches G6, all routing information up to B3 is described in the flit.

Thereafter, when a response packet is sent from G6 to B3, the direction and order of the received routing information are reversed as shown in FIG. As a result, the response packet can be sent back to the IP core 1 located at B3 without performing routing calculation.

Of course, such routing that leaves route information in a packet can be performed only for packet transfer with a specific attribute.

Hereinafter, the function and effect of this embodiment will be described.

Since the main router performs the routing calculation of a plurality of routers in a lump and the other sub-routers perform packet transfer accordingly, the sub-router can be simplified. Overall, network latency can be shortened, and area costs and power costs can be reduced.

In addition, it is possible to easily cope with scalability problems, which are the disadvantages of the source routing system, and local problems and congestions, so that high reliability of the network can be achieved.
It should be noted that the disclosures of the above patent documents are incorporated herein by reference. Within the scope of the entire disclosure (including claims) of the present invention, the embodiments and examples can be changed and adjusted based on the basic technical concept. Various combinations and selections of various disclosed elements are possible within the scope of the claims of the present invention. That is, the present invention of course includes various variations and modifications that could be made by those skilled in the art according to the entire disclosure including the claims and the technical idea.

1 IP core 2 network interface 3 subrouter 4 main router 5 channel 10 multicore SoC (System on a Chip)
DESCRIPTION OF SYMBOLS 21 Input latch 22 Route calculation part 23 Routing flit production | generation part 24 Selector 25 Input FIFO (First In First Out) buffer 26 Crossbar switch 27 Arbiter 28 Selector 29 Output buffer 45 Input FIFO (First In First Out) buffer 46 Crossbar switch 47 Arbiter 48 Selector 49 Output buffer 81 Packet 82 Sideband attribute signal 83 Valid bit 84 First bit 85 End bit 86 Packet body 87 Route information field 88 Control information field 89 Flit length field 90 Target IP position information field
91 Control information field 92 Access attribute field 93 Access address field 94 Data field 101 Route information field 1
102 Route information field 2
103 Route information field 3
104 Route information field 4
105 Control information field 106 Flit length field 107 Target IP location information field 108 Control information field 111 Route information field 1
112 Route information field 2
113 Route information field 3
114 Route information field 4
115 Control information field 121 Route information field 1
122 Route information field 2
123 Route information field 3
124 Route information field 4
125 control information field 137 route information field 138 control information field 139 iteration number field 140 route information field 141 iteration number field 142 route information field 143 iteration number field 144 route information field 145 iteration number field 146 control information field 151 IP core 152 network interface 153 Sub-router 154 Main router 155 Channel 156 Spare main router 160 Multi-core SoC (System on a Chip)
201 packet 202 sideband attribute signal 203 valid bit 204 first bit 205 end bit 206 packet body 207 route information field 208 control information field 209 flit length field 210 target IP location information field
211 Control information field 212 Control information field 301 Multi-core SoC (System on a Chip)
302 router 303 IP core 304 network interface 305 channel 321 packet body signal 322 sideband attribute signal 323 target IP location information 324 flit length 325 control information 326 access address 327 access attribute etc. 328 data 351 router 352 communication channel 353 used for communication example Communication channel 354 Virtual Express channel 361 DID (Destination IDs) used for communication example
362 Chained Header
381 Option type field 382 Option length field 383 Pointer field 384 IP address field

Claims

A network in which a plurality of information processing units or storage units and interface units are integrated on one or a plurality of chips, and the units are connected to each other, and includes one or more networks including a network interface, a communication path, and a network router. Integrated on the chip,
In a semiconductor device that performs information processing by communicating information between the units in the packet via the network,
Divide network routers into multiple groups of main routers and other routers,
If the number of hops that pass through the router cannot reach the receiver within the specified number of times, the number of times that the router passes through the main router is reset every time the main router passes through the main router. A semiconductor device that performs routing so as to satisfy.
The main router comprises means for calculating routing;
The main router and a sub-router that performs routing according to information routed by the main router or the network interface;
The semiconductor device according to claim 1, comprising:
The main router is
From the packet, from the ID or address indicating the data receiving destination, the routing of the output destination from the main router of the packet, and means for routing the sub-router within the specified number of times,
The routing information for the sub-router is updated by writing a flit that describes the route information for reaching the main router in the packet, and the packet is sent to the corresponding router.
The semiconductor device according to claim 1, wherein the sub-router that has received the packet performs routing according to the routing information and updates the routing information.
The main router is
If routing is determined by another main router that has passed through the previous stage within the specified number of times or the network interface on the sending side, routing is not performed autonomously, but the packet is routed according to the routing specified by the main router at the previous stage. The semiconductor device according to claim 1.
The main router is
Means for routing the sub-router;
Sub-routing information is compressed and encoded in a frit and stored.
The semiconductor device according to claim 2, wherein the sub-router performs routing according to information obtained by decompressing the compression-encoded information.
Sub-routers are allowed to hop consecutively up to the number of hops determined by routing information that can be compressed and stored in a specified flit,
The semiconductor device according to claim 5, wherein the specified number of hops varies depending on a compression ratio for routing to a routing destination.
2. The router according to claim 1, wherein the main router is disposed in a vicinity of a router connected to an IP where network congestion is expected in advance and in the vicinity thereof, more than the ratio of the main router to other network routers. Semiconductor device.
3. The semiconductor device according to claim 2, wherein the main router includes means for grasping congestion and failure of surrounding links and the sub-router, and performs routing of the sub-router so as to avoid congestion or failure.
The semiconductor device according to claim 2, wherein the sub-router in which the failure has occurred performs control to return an arriving packet to a nearby main router using an available resource in the router.
The main router has redundant logic, error detection correction function,
3. The semiconductor device according to claim 2, wherein a failure occurring in the link or the sub-router is corrected when a packet arrives at the main router, and is recovered by the redundancy logic when the main router fails.
The routing information performed by the main router is updated by writing the flit written on the route for reaching the main router in the packet, and the flit information before rewriting is moved to another position of the packet. The semiconductor device according to claim 1, wherein the route information from the transmission source of the packet is held in the packet by appending.
A network-on-chip device comprising the semiconductor device according to any one of claims 1 to 11 on-chip.
A plurality of information processing units or storage units, interface units are integrated on one or a plurality of chips, a network connecting the units, and a network composed of a network interface, a communication path, and a network router are integrated together,
Inter-unit information is communicated in packets via the network for information processing,
Network routers are divided into groups of main routers and other routers,
If the number of hops that pass through the router cannot reach the receiver within the specified number of times, go through the main router,
A network routing method for performing routing so as to satisfy the specified number of times by resetting the number of times of passing through the router every time the main router is passed.
The main router calculates the routing,
A sub-router that performs routing according to information routed by the main router or the network interface;
The network routing method according to claim 13, comprising:
The main router is
From the packet, from the ID or address indicating the data reception destination, the routing of the output destination from the main router of the packet, and the routing of the sub-router within the specified number of times,
The routing information for the sub-router is updated by writing a flit describing the route information for reaching the main router in the packet, and the packet is sent to the corresponding router.
The network routing method according to claim 13, wherein the sub-router that receives the packet performs routing according to the routing information and updates the routing information.
The main router is
If routing is determined by another main router that has passed through the previous stage within the specified number of times or the network interface on the sending side, routing is not performed autonomously, but the packet is routed according to the routing specified by the main router in the previous stage. The network routing method according to claim 13.
The main router is
Route the sub-router,
Sub-routing information is compressed and encoded in a frit and stored.
The network routing method according to claim 14, wherein the sub-router performs routing according to information obtained by decompressing the compression-encoded information.
It is allowed to hop continuously in the sub-router up to the number of hops determined by routing information that can be compressed and stored in a specified flit,
The network routing method according to claim 17, wherein the specified number of hops varies depending on a compression ratio for routing to a routing destination.
14. The router according to claim 13, wherein a larger number of the main routers are arranged in a vicinity of a router connected to an IP where network congestion is expected in advance and in the vicinity thereof than other main routers. Network routing method.
15. The network routing method according to claim 14, wherein the main router grasps the congestion and failure of surrounding links and the sub-router, and performs routing of the sub-router so as to avoid congestion or failure.
The network routing method according to claim 14, wherein the sub-router in which a failure has occurred performs control to return an arrived packet to the neighboring main router using resources available in the router.
The main router has redundant logic, error detection correction function,
The network routing according to claim 14, wherein a failure occurring in a link or the sub-router is corrected when a packet arrives at the main router, and is recovered by the redundancy logic when the main router fails. Method.
The routing information performed by the main router is updated by writing a flit describing the route information for reaching the main router in the packet, and the flit information before rewriting is moved to another position in the packet. The network routing method according to claim 13, wherein the route information from the transmission source of the packet is held in the packet by additionally writing.
A network connecting a plurality of information processing units or storage units and interface units, comprising a network composed of a network interface, a communication path, and a network router, and the information between the units in a packet A network system that communicates and processes information via
Divide network routers into multiple groups of main routers and other routers,
If the number of hops that pass through the router cannot reach the receiver within the specified number of times, the number of times that the router passes through the main router is reset every time the main router passes through the main router. A network system that performs routing to meet the requirements.
The main router comprises means for calculating routing;
The main router and a sub-router that performs routing according to information routed by the main router or the network interface;
The network system according to claim 24, comprising:
The main router is
From the packet, from the ID or address indicating the data receiving destination, the routing of the output destination from the main router of the packet, and means for routing the sub-router within the specified number of times,
The routing information for the sub-router is updated by writing a flit that describes the route information for reaching the main router in the packet, and the packet is sent to the corresponding router.
The network system according to claim 24, wherein the sub-router that has received the packet performs routing according to the routing information and updates the routing information.
The main router is
If routing is determined by another main router that has passed through the previous stage within the specified number of times or the network interface on the sending side, routing is not performed autonomously, but the packet is routed according to the routing specified by the main router in the previous stage. The network system according to claim 24.
The main router is
Means for routing the sub-router;
Sub-routing information is compressed and encoded in a frit and stored.
26. The network system according to claim 25, wherein the sub-router performs routing according to information obtained by decompressing the compressed and encoded information.
It is allowed to hop continuously in the sub-router up to the number of hops determined by the routing information that can be compressed and stored in the specified flit,
26. The network system according to claim 25, wherein the specified number of hops varies depending on a compression ratio for routing to a routing destination.
25. The router according to claim 24, wherein the main router is disposed more in the vicinity of the router connected to the IP that is expected to be congested in the network than the main router in the other network routers. Network system.
26. The network system according to claim 25, wherein the main router includes means for grasping a congestion and a failure of a surrounding link and the sub router, and performs routing of the sub router so as to avoid the congestion or the failure.
26. The network system according to claim 25, wherein the sub-router in which a failure has occurred performs control to return an arrived packet to a nearby main router using an available resource in the router.
The main router has redundant logic, error detection correction function,
26. The network system according to claim 25, wherein a failure occurring in a link or the sub-router is corrected when a packet arrives at the main router, and is recovered by the redundancy logic when the main router fails.
The routing information performed by the main router is updated by writing the flit written on the route for reaching the main router in the packet, and the flit information before rewriting is moved to another position of the packet. 25. The network system according to claim 24, wherein route information from the transmission source of the packet is held in the packet by additionally writing.
The main router is
An input latch for receiving an input signal;
A route calculator that calculates the route from the destination,
A routing frit generation unit that generates a routing frit based on a route calculation result in the route calculation unit;
A selector for selecting one of the routing flit generated by the routing flit generation unit and the information from the input latch;
An input FIFO (First In First Out) for storing the selection results of the selectors;
An output buffer for latching the selection result by the crossbar switch, and
A pair containing
Corresponding to IP (Intellectual Property) directly connected to each route and the main router itself,
A crossbar switch having an arbiter for adjusting an output destination of an input signal and a selector for selecting an output from the input FIFO in any direction, a plurality of the input FIFOs corresponding to each path and IP, and a plurality of the output buffers The network system according to claim 24 or 25, wherein the network system is provided in between.
The sub-router
An input FIFO (First In First Out) for inputting an input signal;
An output buffer for latching the selection result by the crossbar switch, and
A pair containing
Corresponding to IP (Intellectual Property) directly connected to each route and the sub-router itself,
A crossbar switch including an arbiter that adjusts an output destination of an input signal and a selector that selects an output from the input FIFO in any direction, includes a plurality of input FIFOs corresponding to each path and IP, and a plurality of input FIFOs. 26. The network system according to claim 25, provided between said output buffer.