JP2004221807A - Distribution routing table management system and router - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the problems with a conventional routing table management system that not only incurs difficulty in the mount because of a high required memory capacity and increase in the cost but also is bottlenecked on the throughput. <P>SOLUTION: Proposed is an NUMA type or a CC-NUMA type and furthermore a COMA type for an architecture of a routing table memory. They are not provided with a shared routing table and manage path information by a distribution routing table distributed and placed to each interface to utilize position locality. The COMA type particularly can much more utilize the position locality because the home position of a path information entry is freely transited through referencing. The CC-NUMA type and the COMA type are provided with a cache to cope with a temporal locality and also to cope with address locality by adopting a mixing work for a hash function of the cache. Moreover, they cope with space locality by changing the hush function for each zone and registering the result to a cache memory to enhance a hit rate of the cache. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a low-latency switching system for large-capacity packets in communication, and more particularly to a routing table required for packet routing work in a low-latency switching system for large-capacity packets and a management and operation method of the routing table. As the packet, in particular, an IP packet and an Ethernet (registered trademark) frame are accepted.
[0002]
[Prior art]
[Non-patent document 1]
"Multi-zone Caches for Accelerating IP Routing Table Lookups, IL Chvetsand MH MacGregor, pp. 121-126, in Proceedings of HPSR 2002, May. 200.
In the field of communications, various router and switch configuration technologies have been proposed that meet the demands for wider bandwidth, higher quality, and lower delay. In particular, the routing table is a part related to the future enhancement of functions, bandwidth, quality, and delay of the network, and the operation performance of the routing table affects the performance of the router and the network.
[0003]
FIG. 12 is a configuration diagram of a conventional basic router. A basic packet processing procedure will be described with reference to FIG. The destination information of the packet input to the plurality of interfaces 101 of the router is examined in the packet processing unit 102 of the router. By referring to the routing table 1201 based on the destination information, information necessary for the output destination interface and the QoS is obtained. By transmitting this information to the switching fabric 105, the packet is transferred to the output destination interface which is the destination. As the switching fabric 105, a crossbar is often used. An arbiter attached to the crossbar accepts the request, and performs arbitration in consideration of QoS information, priority, and the like. A crossbar is set based on the arbitration result, and the input interface and the output interface are correlated.
[0004]
The routing table includes only one shared routing table 1202 (which may be referred to as a primary routing table, a default routing table, or the like) in a router, and a distributed routing table 1201 (secondary) distributed to each interface accompanying the router. Routing table, simple routing table, etc.).
[0005]
The contents of the distributed routing table 1201 are the same as the contents of the shared routing table 1202, and distribute the access to the shared routing table. Especially in the case of a cache, it is called a routing table cache. As a conventional technique relating to caching of a routing table, see "Multi-zone Caches for Accelerating IP Routing Table Lookups, IL Chvets and MH MacGregor, pp. 121-126, in Proceeding Rs. I want to be. For the first time in this dissertation, IP-specific caches were considered. This paper proposes a multi-zone cache that makes use of temporal locality and spatial locality existing in IP traffic. The time locality indicates that once access is made from an IP address in a router, access to the same IP address is likely to occur many times in a short time thereafter. The spatial locality is that when a router outputs a packet with a certain IP address to a certain destination (for example, an interface number), there is a possibility that a packet having an IP address in the vicinity (an IP address in which only the lower few bits are different) is also output to the same destination. Is higher.
[0006]
In the technique of FIG. 12, the shared routing table and other routing tables have the same contents, and the contents of the shared routing table are copied to other routing tables as appropriate. Further, in the conventional example of FIG. 12, communication is not performed between the routing tables of other interfaces.
[0007]
If route information is centrally managed by a shared routing table (1202 in FIG. 12) provided only in the entire router, it is not possible to cope with the concentration of table reference requests due to an increase in the processing speed of the router. (Routing table cache) and a corresponding configuration has been proposed.
JP-A-2002-164899, JP-A-2000-83055, JP-A-11-289435, and JP-A-11-122289 have been proposed as having a distributed routing table equivalent to a routing table cache. These have both a single shared routing table as a whole and a routing table cache provided for each interface. The shared memory is connected via a common bus or switch, and such a memory management architecture is called a UMA (Uniform Memory Access model).
[0008]
[Problems to be solved by the invention]
An object of the present invention is to improve a routing table in a router and its reference mechanism, and to increase the capacity, reduce the delay, and reduce the cost of the router. This requires simplification and miniaturization of the memory used for the routing table. As a result of miniaturization of the memory, cost reduction can be achieved by reducing the number of chips and the board area, and the operation speed is improved, so that the throughput is improved.
[0009]
In the conventional UMA type routing table management structure, the routing table (1201 in FIG. 12) and the shared routing table (1202) are mounted in different places but have the same contents. This structure has been devised to distribute the access concentration of the shared routing table by giving the same contents to each interface so that the load is not concentrated on the shared routing table.
[0010]
The present invention focused on the following problems in the conventional method.
1. The shared routing table is a union of the routing tables in each interface. Therefore, the size of the shared routing table becomes large, and a search cost and an implementation cost when referring to the contents are required, which cannot contribute to an improvement in the effective throughput and a reduction in the price. When each interface has the same contents as the shared routing table, not only costs are increased, but also it becomes difficult to cope with the enlargement of the routing table accompanying the development of the network.
[0011]
When both the shared routing table and the routing table cache are provided, if the difference between the routing information in each interface is large, the amount of information to be remembered in the shared routing table becomes enormous. Generally, as the amount of information in the routing table increases, the access cost also increases. Therefore, in a large-scale and high-speed router, the access cost of the shared routing table becomes a problem, and it becomes difficult to improve the execution throughput.
2. When there is a strong interface-dependent locality (position locality) in referencing the routing table, there are few references to the shared routing table, and it is difficult to obtain a performance improvement corresponding to the hardware cost required for providing a huge shared routing table. .
3. In the caching of the routing table, not only the temporal locality and the spatial locality but also the above-mentioned positional locality and further, the address locality need to be considered. The address locality indicates that valid or frequently used IP addresses and MAC addresses are not uniformly distributed over the entire address space but are unevenly distributed. Therefore, the hash function in the cache of the routing table is not a simple one that collects partial bit information of a given IP address, but a cache hash function corresponding to address locality is required. The Multi-Zone cache effectively utilizes the spatial locality, but does not consider the position locality and the address locality.
4. When implementation costs and processing costs are required to realize a hash, it is difficult to provide different hashes for each Zone. Japanese Patent Application Laid-Open No. 2001-320420 uses a search result based on a hash for adjusting traffic, and obtains a CRC of the entire IP address.
5. Japanese Patent Application Laid-Open No. 2001-326679 is an invention for a routing table search technique, in which an inode method of a multiple-time indirect reference method used as a standard in a file system such as UNIX (registered trademark) is applied to routing table reference. It is effective when the target is a large capacity (a hard disk targets a capacity such as giga or tera) as used in a file system that is lower in the storage hierarchy. Is slow.
[0012]
[Means for Solving the Problems]
In order to solve the above-mentioned problems, the present invention uses the following means.
1. The structure is such that a common routing table is not provided and a routing table is distributed and arranged for each port. That is, unlike the example of FIG. 12, each routing table has a different combination of information. The fact that a routing table can have its own content means that each table can be optimized as desired using locality. This structure is called a NUMA (Non-Uniform Memory Access model). Furthermore, a structure of a CC-NUMA (Cache Coherent-NUMA) having a cache in the NUMA is adopted. Alternatively, a structure is adopted in which a routing table cache is distributed and arranged for each port. This structure is called COMA (Cache Only Memory Architecture). As a specific router configuration, a router having a plurality of interfaces, each of which includes a packet processing unit for processing an input packet, and a set of information dependent on the packet and a destination of the packet as information. A routing table for storing and referencing the packet-dependent information as a key, and searching for a destination of the packet. If a search is not found in the routing table of the first interface, the routing of the second interface is performed. Search the destination by referring to the table. With this configuration, it is possible to avoid a problem that the routing table is enlarged due to the development of the network. The information depending on the packet is, for example, an IP address or a MAC address.
[0013]
Each routing table has a routing table cache that stores a copy of a part of the information held by the routing table. When a search is not found in a certain routing table cache, the routing table corresponding to the routing table cache is referred to. In this case, CC-NUMA for searching for a destination may be used.
[0014]
Each routing table is configured as a routing table cache. If a search is not found in the routing table cache on the interface where the packet arrives, the destination is referred to by referring to the routing table cache on the other interface. If a search is made and a hit is found, the hit data may be copied to the first routing table cache as COMA.
[0015]
When the second routing table returns a reference result, the effect of the cache can be obtained by storing the reference result in the first routing table.
[0016]
In addition, if the first path for transferring the input packet between the interfaces and the second path for transferring the storage information of the routing table between the interfaces are provided independently, the transfer speed of both can be improved. Can be.
[0017]
Further, a stirring means can be added to the hash function of the routing table cache to improve the memory use efficiency.
[0018]
An entry in another routing table having the same entry as the routing table changed when the contents of the routing table are changed may be invalidated, or the change may be notified to a routing table in another interface by an update message. .
[0019]
According to another aspect of the present invention, there are provided a plurality of packet processing units and a plurality of routing tables corresponding to each of the packet processing units, and the first routing table is processed by the corresponding first packet processing unit. Is used first to find the destination of the packet that has been sent, and if the search is not hit in the first routing table, a search is made in the second routing table, and the second routing table is Is used first to search for the destination of the packet processed by the second packet processing unit. If the search is not hit in the second routing table, the search is performed in the first routing table. Provide a router.
[0020]
According to another aspect of the present invention, there is provided a routing method for a router including a plurality of interfaces each receiving a packet, wherein each of the interfaces includes a packet processing unit that processes an input packet, and an address of the input packet. A routing table for searching a destination based on the routing table, and a routing table cache for storing at least a part of a table stored in the routing table, and based on an address of a packet input to an interface, A first step of searching for a destination in the routing table cache; a second step of searching for a destination in the routing table of the interface when the destination cannot be searched in the first step; and a second step of searching for a destination in the routing table of the interface. When bought, a third step, to find the destination in the routing tables of other interfaces.
2. If the structure of the routing table is NUMA and a distributed routing table that each interface independently manages is provided, location locality can be utilized. As a result, a memory amount smaller than the memory amount required for the shared routing table is sufficient. Also, if CC-NUMA is used, the content managed by each interface is cached even when referring to the content managed by each interface regardless of the interface itself, so that the location locality can be effectively utilized, and the reference speed can be similarly increased. Improvement can be expected. If the management structure of the routing table is set to COMA, each piece of routing information does not stay in a fixed place but moves to a place where access frequently occurs, so that ideal routing information is arranged. Therefore, the management becomes complicated, but the position locality can be effectively utilized.
3. In order to make effective use of address locality, a hash function that performs agitation rather than bit extraction is required instead of a hash function that simply extracts a certain bit that is usually used. For example, an IP address of 10.113.53.152 is expressed in binary notation as 00001010.011110001.00101011.10011000. If a memory is drawn using all of them, a 32-bit address space is required, which makes mounting difficult. Therefore, when a hash function for extracting a bit at a location that is a multiple of 4 as a certain bit is used, the hash value becomes 01010011, and a hash in which a memory is subtracted based on this value can be considered. However, considering the locality of the address, the hash value obtained in this way has a high probability of obtaining a similar value, and the memory cannot be used efficiently. Therefore, a stirring means is required. Since complicated stirring requires hardware and processing costs, a hash function using a simple CRC (Cyclic Redundancy Check) remainder is used. Further, in order to effectively utilize the spatial locality in a single cache without providing a plurality of the same cache structures, a different hash structure is provided for each zone and a dual-port memory is used.
4. In the CRC, a better stirring result is obtained by the generator polynomial, the number of bits at the source and the destination can be freely taken (the number of bits after conversion needs to be smaller than the number of bits before conversion), and the implementation cost And the processing cost are both small. CRC can be easily parallelized and performed at high speed. The CRC is used as a hash function of the IP address and used for a routing table cache.
5. Speeding up is achieved by searching for the routing table cache not by indirect reference but by obtaining a single reference to the hash as in a normal cache. The reference method of the cache is set associative.
According to one aspect of the present invention, a router for communication includes the arbiter for ensuring bandwidth and real-time performance and an associated or integrated switch.
[0021]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following embodiments, a router is used as an example. However, the present invention is not limited to a router, and is expected to have an effect of increasing capacity, reducing delay, and reducing cost by applying the present invention when equivalent localities exist. it can.
[0022]
Embodiment 1
An example of the CC-NUMA type routing table structure and operation according to the present invention will be described.
[0023]
FIG. 1 shows a configuration example of a CC-NUMA type routing table. In the CC-NUMA type, each interface 101 includes a packet processing unit 102 that performs various processes on a packet, a routing table cache 103, and a distributed routing table 104. Further, a switching fabric 105 for exchanging packets is provided outside each interface. When the routing table cache 103 is not provided, it is called a NUMA type routing table structure.
[0024]
When an interface refers to a distributed routing table 104 (remote distributed routing table) in another interface, there are a method of sharing a signal line used for transmitting and receiving a packet and a method of separately providing a unique signal line. . In the case of common use, another fabric 106 for exchanging routing table information is not required, and new hardware costs and wiring costs on a board can be reduced. However, there is a possibility that a decrease in bandwidth occurs due to collision between packet communication and routing information communication. When not shared, a fabric 106 for exchanging routing table information separately is provided as shown in FIG. Examples of fabrics include crossbars and rings (bidirectional rings, unidirectional rings).
[0025]
Each interface has its own distributed routing table and routing table cache. By providing a cache, it supports time locality. This is because a plurality of accesses to the same address that occur in a short time due to the cache almost all hit the cache, and a high-speed address reference can be expected. At the same time, it corresponds to the position locality. This is because, even when there is an access pattern peculiar to each interface, it can be expected that a distributed routing table or a routing table cache independently provided for each interface can optimize entries according to the access pattern of the corresponding interface.
[0026]
The operation of the routing table cache differs slightly depending on whether the address to be handled is a layer 3 IP address or a layer 2 MAC address.
1. When performing layer 3 switching (layer 3 forwarding)
In Layer 3 switching, routing is performed using an IP address. For the IP address, all of the time locality, the position locality, the space locality, and the address locality are valid. The time locality and the location locality are handled by the distributed routing table and the routing table cache. To take advantage of spatial locality, a different CRC agitation hash is used for each Zone. Each zone means that when a zone is divided by a certain prefix length, the zone is distinguished by whether the zone is longer or shorter than the zone, each zone is set to a different zone, and stirring is performed using a different CRC. CRC operates at high speed with a small mounting cost. When the stirring is performed by the CRC, the addresses are randomly distributed, and the density of the IP address space becomes sparse as a result of the stirring. Therefore, it is possible to prevent the cache performance from deteriorating due to the concentration of registration at a specific address of the cache, and to uniformly access the cache address, so that the cache memory can be used effectively.
[0027]
FIG. 2 shows a flowchart in accessing the routing table cache 103 and the distributed routing table 104. Here, IPv4 is shown as an example, but the same applies to IPv6. Further, the mask length of the IPv4 address is discriminated from the zone at the boundary of 22. If the mask length is smaller than 22, the hash is registered in the Zone-A cache using the hash A, and if the mask length is larger, the zone is registered in the Zone-B cache using the hash B. Will be described. These processes are controlled by the control unit 107 by software or hardware.
[0028]
Assuming that there is access to the IP address 192.168.34.56, first, in step S201, a hash value for each zone is obtained. If the cache has a dual-port configuration, two zone-specific hash values can be searched simultaneously. That is, since it can be divided into two Zones by the break of Prefix, each has a different hash generation and search structure. In S202, the routing table cache is referred to. If there is a hit, the packet is forwarded in S206. If there is a mishit, the distributed routing table (local distributed routing table) existing on the same interface is referred to in S203. If there is a mishit, the distributed routing table (remote distributed routing table) of another interface is referred to in S204. If there is still a miss hit, a process for when the destination is unknown is performed in S205. Specifically, processing such as forwarding or discarding the packet to the default route is performed.
[0029]
If the local distributed routing table or the remote distributed routing table is hit, the packet is forwarded in S207. At the same time, registration in the cache is also performed. At the time of registration, reference is made not only to the destination but also to the mask value.
[0030]
On the Internet, those having the same IP address prefix are treated as a group. It is the mask value that specifies how far the same prefix is. For example, in the case of WINDOWS XP (registered trademark) or the NT system, when an ipconfig instruction is executed from a command line, a mask value called a subnet mask appears together with an IP address. In the following, ZONE is a concept classified according to how long the mask value is, that is, how far the prefix is to be viewed.
[0031]
When the hit entry is 192.168.0.0/16, since the mask length is 16 and less than 22, the address of 192.168.34.56 is stored in the Zone-A cache using the hash A. The destination information is registered. When the hit entry is 192.1684.34.0/24, the mask length is 24 and is larger than 22, so that the entry is registered in the Zone-B cache using the hash B.
2. When performing layer 2 switching
In layer 2 switching, routing is performed using a MAC address. As for the MAC address, temporal locality, spatial locality, and address locality are valid. The operation in this case can also be described using the flowchart in FIG.
[0032]
Since there is no mask value in the MAC address, Zone is not distinguished. Therefore, stirring is performed simply by CRC. If there is an access to the MAC address 00: 11: AA: 22: BB: 33, a hash value is obtained in step S201. In S202, the routing table cache is referred to. If there is a hit, the packet is forwarded in S206. If a mishit occurs, a distributed routing table (local distributed routing table) existing on the same interface is referred to in S203. If there is a mishit, the distributed routing table (remote distributed routing table) of another interface is referred to in S204. If there is still a miss hit, a process for when the destination is unknown is performed in S205. Specifically, processing such as forwarding or discarding the packet to the default route is performed.
[0033]
If the local distributed routing table or the remote distributed routing table is hit, in step S207, the packet is forwarded and registered in the routing table cache according to the hash value.
[0034]
Embodiment 2
A procedure for performing route information processing on the CC-NUMA type routing table according to the present invention will be described. When performing route information processing such as BGP (Border Gateway Protocol) or OSPF (Open Shortest Path First), it is necessary to read the routing table because it is necessary to convey the own route information to the other party. In the shared routing table method, it is only necessary to read the contents of the corresponding shared routing table. However, in the distributed routing table method, since information is dispersed, a separate procedure for summarizing the information is required. Therefore, the following policy is adopted. The route control processing itself is performed by dedicated hardware or a processor capable of accessing each distributed routing table.
1. Information received by a certain interface in the routing process is managed only in the local distributed routing table. Therefore, no notification is made to the remote distributed routing table.
2. When an interface sends information associated with the route control processing and needs the contents of the remote distributed routing table, the contents of the corresponding remote distributed routing table are acquired and notified.
[0035]
According to this procedure, information comprehensively including the contents of the entire distributed routing table can be obtained. The same applies to the COMA type routing table. Basically, the routing information is handled only by the locally existing home, and the content is acquired and notified only when it is necessary to refer to the home of the remote distributed routing table.
[0036]
Cache injection (speculative registration of cash) is also possible in accordance with these transfer procedures. Since the information registered in the cache is a perfect match, it is not possible to directly register an address to which a mask exchanged in the normal route control processing is added. Therefore, a value obtained by embedding an appropriate value in a masked portion, or an address specified without a mask in IPv4 or IPv6 is cached.
[0037]
Embodiment 3
When the existing data is updated by adding new data to the distributed routing table by the route control processing according to the second embodiment, the contents of the routing table cache existing in another interface sharing the information. Need to be updated. This procedure will be described for the case where bitmap information is added to entries of each shared routing table and managed. When the management is performed by adding the bitmap information, the management information increases, but the communication amount at the time of updating can be reduced due to the fine management.
[0038]
FIG. 3 is a structural diagram showing an entry structure of a shared routing table and a distributed routing table in a case where a sharing situation is managed by a bitmap, with respect to the CC-NUMA type routing table management structure in the present invention.
[0039]
FIG. 3 shows an example of the structure of an entry of the distributed routing table corresponding to FIG. The entry includes IP address, destination interface number, QoS information, Shared bitmap information, and other additional information such as management information and mask value information for management using a Patricia tree structure or the like. The IP address is managed by integrating the destination and the destination. The QoS information describes, for example, occupied bandwidth and information of QoS class (how to guarantee the bandwidth). The shared bitmap information indicates whether the information of the entry is shared with another interface. For example, if there are eight interfaces and the second and seventh interfaces have the same entry, the shared bitmap information Becomes 01000010.
[0040]
FIG. 4 shows a protocol in this method. If there is a mishit in the routing table cache (401 in FIG. 4), a reference request is issued to the distributed routing table on the same interface (402 in FIG. 4). If a hit is found, the corresponding portion (the portion corresponding to the same interface number) is marked in the shared bitmap of the entry. Here, “Shad” indicates that the same entry is shared by a plurality of interfaces.
[0041]
If there is a mishit, a reference request is issued to the distributed routing table existing on another interface (403 in FIG. 4). When an entry exists in a certain shared routing table, a corresponding part (a part corresponding to the issuer interface number of the reference request) is marked in the Shared bitmap of the entry (404 in FIG. 4). Because it is a bitmap, it is possible to know which interface is caching. For example, in FIG. 4, since it is a reference request for the interface 1, the first bit of the bitmap is marked. When the entry is updated (405 in FIG. 4), an Invalidate message is issued to the routing table cache of the corresponding interface according to the Shared bitmap to invalidate the cache (406 in FIG. 4). At the same time, the mark of the shared bitmap is deleted.
[0042]
FIG. 5 is a structural diagram showing an entry structure of a routing table cache having information indicating a home position in the CC-NUMA type routing table management structure according to the present invention.
[0043]
As shown in FIG. 5, a home ID indicating which interface existing data is cached is managed. When erasing a cache entry (such as eviction of a cache entry), the shared bitmap of the distributed routing table of the corresponding interface is changed in accordance with the home ID so that a useless Invalidate message is not generated even if the entry is updated.
[0044]
Here, the invalidation of the cache by the Invalidate message has been described, but the cache may be updated by the Update message.
[0045]
Embodiment 4
Different from the third embodiment, a case in which an identifier is added to an entry of each shared routing table and managed is described. This method increases the amount of communication at the time of updating, but reduces the amount of management information.
[0046]
FIG. 6 is a structural diagram showing an entry structure of a distributed routing table in a case where a shared state is managed by a single bit in a CC-NUMA type routing table management structure according to the present invention.
[0047]
FIG. 6 shows an example of the structure of the entry of the shared routing table in this case. It has a Shared flag instead of a Shared bitmap.
[0048]
FIG. 7 shows a protocol in this method. When a miss occurs in the routing table cache (701 in FIG. 7), a reference request is issued to the distributed routing table on the same interface (702 in FIG. 7). If there is a hit, the shared flag of the entry is marked. If there is a mishit, a reference request is issued to the distributed routing table existing on another interface (703 in FIG. 7). If an entry exists in a certain shared routing table, the shared flag of the entry is marked (704 in FIG. 7). Because it is a flag, it is not possible to know which interface is caching. When the entry is updated (705 in FIG. 7), if the Shared flag is marked, an Invalidate message is broadcast to invalidate the cache (706 in FIG. 7). If the interface receiving the Invalidate message has a corresponding entry, the cache entry is invalidated. If you don't have it, ignore it. At the same time, the shared flag mark is deleted. By providing two Shared flags for the local interface and the remote interface, it is possible to avoid broadcasting for updating entries to be resolved locally.
[0049]
On the routing table cache side, a home ID indicating which interface existing data is cached is managed as in the third embodiment. When erasing a cache entry (such as flushing a cache entry), the shared bitmap of the distributed routing table of the corresponding interface is changed according to the home ID so that even if the entry is updated, useless broadcast of Invalidate message is not generated. I do.
[0050]
Here, the invalidation of the cache by the Invalidate message has been described, but the cache may be updated by the Update message.
[0051]
Embodiment 5
FIG. 8 is a structural diagram showing a CC-NUMA type routing table management structure according to the present invention, particularly showing an entry structure of a routing table cache having no management information.
[0052]
In the third and fourth embodiments, a method in which the management information is not added to the routing table cache side as shown in FIG. 8 will be described.
[0053]
When the cache entry is deleted and used together with the third embodiment (e.g., eviction of a cache entry), a request to delete Shared information in the shared routing table is broadcast. When used together with the fourth embodiment, nothing is particularly performed when a cache entry is evicted.
[0054]
In addition, a method of not adding management information to each shared routing table and routing table cache is also conceivable. This method does not require management information, but always updates the entry to another interface when updating the entry.
[0055]
FIG. 9 shows a configuration example of a NUMA type routing table as a CC-NUMA cacheless structure. While the NUMA type routing table memory contributes to the reduction in the size of each memory, it is difficult to reduce the delay because it does not have a cache.
[0056]
Embodiment 6
An example of a COMA type routing table structure according to the present invention will be described.
[0057]
FIG. 10 shows a configuration example of the COMA type routing table. In the COMA type, each interface 101 includes a packet processing unit 102 that performs various processes related to a packet, and a routing table cache 103. Further, a switching fabric 105 for exchanging packets is provided outside each interface. When this switching fabric is shared for exchanging routing information, another fabric 106 for exchanging routing table information is not required. If not shared, a fabrics 106 for exchanging routing table information is provided separately.
[0058]
Each interface has its own routing table cache so that it can accommodate temporal and location localities. In particular, in the case of the COMA type, unlike the CC-MUMA type, the movement of the home (location where a certain address is stored) of each data is permitted. Here, the data is an entry, that is, a set of information in a horizontal row shown in FIG. In the CC-NUMA type, an entry in the distributed routing table is set by the routing protocol in each interface, and this entry (the original entry, not the cache, particularly the home) is not moved to the distributed routing table in another interface. In the COMA type, the remote routing table cache that frequently refers to the home caches the home (this is called a copy) and continues to access the same entry (a copy of the home). If the access is delayed, the entry that was previously home is evicted, and the copy becomes home. In this way, each entry is located closer to the interface that is more frequently referenced, so that the placement of the entries is optimized with use, resulting in lower access costs.
[0059]
In order to take advantage of this feature, when the number of interfaces is large in the COMA type, the fabric 106 has a hierarchical structure.
[0060]
FIG. 11 shows an example of a fabric having a hierarchical structure, which has a binary tree structure. Each branch has a directory 1101. This binary tree structure may be an n-ary tree structure using a larger integer n. The following describes a routing table management protocol using COMA. Each directory holds the next state of the entry.
[0061]
a) Invalid: Indicates that there is no data in the corresponding entry. Abbreviated S state
b) Exclusive: No other copy exists. Abbreviation E state
c) Shared: There may be other copies. Abbreviated S state
d) Reading: A reading request is issued to the upper hierarchy, and a waiting state is set. Abbreviation R state
e) Answering: A waiting state is issued after issuing a read or write request to the lower hierarchy. Abbreviation A state
f) Writing: A write request is issued to the upper hierarchy, and the state is in a waiting state. Abbreviation W state.
[0062]
First, the operation at the time of reading will be described. The entries in the E and S states can be read freely. If the reading is missed, the corresponding entry is added to make the state R, and the upper directory is inquired. If there is a hit in reading, the state is set to A and an inquiry is made to the lower hierarchy. If not, the state is set to the R state, and an inquiry is made to an upper layer. By repeating this operation recursively, the user eventually arrives at home or copy. The information of this entry is returned to the hierarchy based on the R and A entries, and each entry is sent to the accessed interface while changing each entry to the Shared state.
[0063]
Next, the operation at the time of writing, that is, at the time of updating will be described. The entry in the E state can be freely written. In the case of an S state or a write error, it is necessary to search for a home or copy and update the data. In the case of the S state, the entry is updated, the state is changed to the W state, and the upper directory is inquired. In the case of a writing error, an inquiry is directly made to the upper directory. If an entry exists in that hierarchy, the state is set to A and an inquiry is made to the lower hierarchy. If not, the state is set to the W state, and an inquiry is made to an upper layer. By repeating this operation recursively, the home or copy is reached. The information of this entry is returned to the hierarchy based on the entries of W and A, and only the entry in the A state is sent to the accessed interface while changing to the Shared state. When the user arrives at the highest level without arriving at home or copy, the information is to be held as Exclusive. Therefore, the user goes down the level while changing the W state to the S state, and adds a new entry as the E state.
[0064]
Embodiment 7
If the routing table cache entry for an interface becomes full and cannot be registered any more, there is a method for avoiding the next discard.
1. Find other open entries by broadcast
The broadcast message is used to search for a free entry in another routing table cache and transferred. The free routing table cache reserves the entry and returns a response message. An interface that has no entry that can be registered additionally passes the entry according to the response message that has returned the earliest. This message is also broadcast, and the routing table cache, which has been transferred, registers the entry in the reserved area, and the routing table cache, which has not been transferred, releases the reservation.
2. Send transfer request to each interface in order
In the method using a broadcast message, the transfer operation can be performed quickly, but the communication may be congested because the broadcast is used. Therefore, there is a method in which a transfer request is issued to each interface in order and a sequential confirmation is taken. The transfer operation can be time-consuming, but does not require the use of broadcast messages.
3. Prepare a separate memory and save it there.
[0065]
Although the difference in structure from UMA, NUMA, and CC-NUMA is small, this memory is only for compensating for the lack of entries in COMA, and is used as swap space.
[0066]
【The invention's effect】
By using the NUMA, CC-NUMA, and COMA type routing memory management system according to the present invention, the amount of routing table memory can be reduced. Further, a method using a cache such as the CC-NUMA or COMA type enables a high-speed routing table reference corresponding to various localities.
[0067]
The routing memory management method according to the present invention is scalable, and is especially important when constructing a large-scale and high-throughput router.
[0068]
In the advanced information society, a router having a larger capacity is required in terms of both communication speed and routing table capacity. The performance can be improved by alleviating the increase in the routing table reference time, which is one of the factors that limit the effective throughput of the router.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a CC-NUMA type routing table management structure as an example of a management structure of a routing table memory in the present invention.
FIG. 2 is a state transition diagram showing a procedure for performing an address search on a CC-NUMA type routing table management structure according to the present invention.
FIG. 3 is a structural diagram showing an entry structure of a distributed routing table in a case where a sharing status is managed by a bitmap, with respect to a CC-NUMA type routing table management structure according to the present invention.
FIG. 4 is a flowchart showing a processing flow when an update operation is performed on a CC-NUMA type routing table management structure using a bitmap according to the present invention.
FIG. 5 is a structural diagram showing an entry structure of a routing table cache having information indicating a home position in the CC-NUMA type routing table management structure according to the present invention.
FIG. 6 is a structural diagram showing an entry structure of a distributed routing table in a case where a shared state is managed by a single bit in a CC-NUMA type routing table management structure in the present invention.
FIG. 7 is a flowchart showing a processing flow when an update operation is performed on a CC-NUMA type routing table management structure using a single bit in the present invention.
FIG. 8 is a structural diagram showing an entry structure of a routing table cache having no management information, particularly regarding a CC-NUMA type routing table management structure according to the present invention.
FIG. 9 is a block diagram showing a NUMA type routing table management structure as an example of a management structure of a routing table memory in the present invention.
FIG. 10 is a block diagram showing a COMA type routing table management structure as an example of a management structure of a routing table memory in the present invention.
FIG. 11 is a structural diagram showing an example of a binary tree structure as an example of a fabric having a hierarchical structure in the COMA type routing table management mechanism of the present invention.
FIG. 12 is a block diagram showing a conventional UMA type routing table management mechanism.
[Explanation of symbols]
101 Interface
102 Packet processing unit
103 Routing table cache
104 Distributed routing table
105 Switching fabric for packet switching
106 Switching fabric for management information exchange
S201 Hash operation state
S202 Reference state of routing table cache
S203 Reference state of local distributed routing table
S204 Reference state of remote distributed routing table
S205 Processing status when destination is unknown
S206 Packet forwarding state
S207 Packet forwarding and cache registration status
401 Action missed in routing table cache
402 Action to issue reference request to local distributed routing table
403 Action to issue reference request to remote distributed routing table
404 Action to Mark Shared Bitmap
405 Update entry action
706 Invalidate Message Issuing Action
701 Action missed in routing table cache
702 Action to issue reference request to local distributed routing table
703 Action to send reference request to remote distributed routing table
704 Action to mark the Shared flag
705 Update entry action
706 Invalidate Message Issuing Action
1101 directory
1201 routing table
1202 Shared routing table.

Claims (12)

A router having a plurality of interfaces,
Each of the plurality of interfaces is
A packet processing unit that processes the input packet;
A routing table that stores a set of the information dependent on the packet and the destination of the packet as information, is referred to using the information dependent on the packet as a key, and searches for a destination of the packet,
A router that searches for a destination by referring to the routing table of the second interface when the search is not hit in the routing table of the first interface. Each of the routing tables has a routing table cache for storing a copy of a part of the information held by the routing table,
2. The router according to claim 1, wherein when a search is not found in a certain routing table cache, a destination is searched by referring to a routing table corresponding to the routing table cache. Each of the above routing tables is configured as a routing table cache,
If no search is found in the routing table cache at the interface where the packet arrived, the destination is searched by referring to the routing table caches at the other interfaces. 2. The router according to claim 1, wherein the data is copied to a routing table cache of the router. 2. The router according to claim 1, wherein when the second routing table returns a reference result, the reference result is stored in the first routing table. 2. The router according to claim 1, further comprising a first path for transferring a packet input between the interfaces and a second path for transferring information stored in the routing table between the interfaces. 4. The router according to claim 2, wherein a stirrer is used for the hash function of the routing table cache. 2. The method according to claim 1, wherein when the content of the routing table is changed, an entry of another routing table having the same entry as the changed routing table is invalidated, or the change is notified to a routing table in another interface by an update message. Router. 2. The router according to claim 1, wherein an entry in the routing table includes a bitmap or a flag for indicating that another routing table holds the same content as the content of the entry. 2. The router according to claim 1, wherein the information dependent on the packet is an IP address or a MAC address. A plurality of packet processing units, and a plurality of routing tables that exist corresponding to each packet processing unit,
The first routing table is used first to search for the destination of the packet processed by the corresponding first packet processing unit. If the search is not hit in the first routing table, the first routing table is used. A search is performed in the second routing table,
The second routing table is used first to search for the destination of the packet processed by the corresponding second packet processing unit. If the search is not hit in the second routing table, the second routing table is used. The router whose search is performed in the first routing table. A routing method for a router having a plurality of interfaces each receiving a packet,
A packet processing unit for processing an input packet; a routing table for searching for a destination based on an address of the input packet; and a routing table for storing at least a part of a table stored in the routing table. Have a cache,
A first step of searching a destination in the routing table cache of an interface based on an address of a packet input to the interface;
A second step of searching for a destination in the routing table of the interface when the destination cannot be searched in the first step;
A third step of searching for a destination in a routing table of another interface when the destination cannot be searched in the second step;
A routing method comprising: 12. The routing method according to claim 11, wherein when a destination can be searched in the second or third step, data of the destination is registered in the routing table cache searched in the first step.

Images