KR100364433B1 - IP address look-up method using a bit-vector table - Google Patents

Abstract

The present invention changes the trie structure so that the forwarding table used for IP address retrieval of the router can be searched at high speed so that it can be applied to a Gbps (Giga bit per second) high speed router. Is compressed, and the compressed information is expressed as a separate field value, and the IP address retrieval method shortens the tri search path.

An IP address retrieval method using a bit-vector table according to the present invention comprises the steps of: (a) constructing a routing table used for retrieving an IP address into a bit-vector table represented by a prefix tree; (b) dividing the IP address to be searched in the bit-vector table by m bits; obtaining the number of child nodes and the number of prefix bits in a bit-vector table by indexing the value divided by the m bits; And (d) searching for an IP address using a cumulative value of accumulating the number of prefix bits as a routing index.

Description

IP address look-up method using a bit-vector table}

The present invention relates to an IP address lookup method of an internet router, and more particularly, a forwarding table used for address lookup of a router may be applied to a gigabit per second (Gbps) -class high speed router. It's about how you can search at high speed.

That is, the data structure and search algorithm of the forwarding table that can be quickly searched by configuring the forwarding table stored in the DRAM memory small enough to be stored in the cache memory.

1 illustrates an embodiment of a conceptual design of a general router.

In FIG. 1, routers are connected to each other via a network, a forwarding engine and a network processor via a switching fabric. Only the headers of packets entering the network interface are forwarded to the forwarding engine to determine the network interface to be used for forwarding. The network interface performs the routing protocol, handles other parts of the packet header, and so on. A representative example employing the design of FIG. 1 is a BBN Multigigabit router.

2 illustrates another embodiment of a conceptual design of a router.

In the design of Figure 2, the processing elements included in the network interface perform IP address lookups individually, a representative example being a router from Ascend Communications. Both the forwarding engine of FIG. 1 and the processing elements of the interface of FIG. 2 have a local routing table, that is, a forwarding table. The routing table, which is constantly changed by the routing protocol, is dynamically managed by the network processor. The forwarding table also allows the network processor to build on the routing table and download it from the forwarding engine. The reason why the routing table and the forwarding table are used separately is that the routing table should be easy to change and manage dynamically, and the forwarding table should be able to support fast search instead of the dynamic structure. Among the structures and functions of the routers described above, the IP address lookup function processed by the forwarding engine is evaluated as the most important factor determining the performance of the router, and other functions are known to be handled with a relatively small burden.

The router performs a routing lookup that searches the routing table to determine the output link to forward the IP datagram.

Figure 3 shows an entry in the routing table, which shows a combination of an IP address and a next hop.

An IP address is represented by a prefix of unspecified length (a network address on the Internet, up to 32 bits long), and the result of a routing search is an output link to the final destination of the IP datagram, that is, This is the next hop value. Routing lookup is performed based on the Longest Prefix Match (LPM) technique, which selects entries whose longest prefix matches. For example, when there is a routing table as shown in FIG. 3, the prefix length of each entry is configured to be in an inconsistent state. If 10000111 is given as the address to be searched, an exact match is not searched. However, there are 100 * and 1000 * entries that match at least partly, and the longer entry value, 1000 *, is selected from the two prefixes and L6 is selected as the next hop.

The traditional routing table implementation uses the Patricia tree, which was invented decades ago and modified for LPM. A typical implementation is NetBSD 1.2. In the NetBSD implementation, the inner and terminal nodes of the tree are each 24 bytes in size, and a routing table with about 40,000 entries is implemented in 2MB. This size routing table is so large that it cannot fit in a single-chip (L1) or L2 cache (SRAM: 10-20 nsec), so it uses memory (DRAM: 60-100 nsec). Because of its very slow access speed, several memory reads required for IP address lookups can be fatal. As a result, in order to implement a high-speed router, a new search method is required to compensate for the existing shortcomings and enable fast search.

There are three main methods for implementing a high-speed router capable of fast IP address lookup.

First, the technique using hardware is using ①CAM (Content-Addressable Memory) and ② applying parallelism and cache.

Second, fast forwarding is implemented by changing the communication protocol such as recording a label or additional information in the packet header.

Third, the software-based method is to apply a more efficient data structure and retrieval method, ④ a study to apply an exact matching scheme, and ⑤ a study to change the trie structure. Can be classified as

The first way of using hardware is to have good performance, but at the cost of high cost.

The second method of changing the communication protocol is difficult to use because of the difficulty of changing the existing Internet protocol. Of course, it can be applied gradually by providing compatibility as an extension concept of the existing communication protocol, but there is a disadvantage in that it cannot be realized unless it is a large-scale application of a network unit.

The last method, based on software, is to apply binary search technique by extending various length prefixes to the same length as in ④, and extend binary search to multiple searches based on prefix extension. This method has a disadvantage that it is difficult to implement and use because the data structure is complex because it must have various fields that prevent back-tracking from binary search. The method of using tri as in ⑤ compresses the steps of the tri again, expresses the compressed information as a separate field value, shortens the tri search path, and devises various data structures representing the tri to enable quick retrieval. Can be divided into

The technical problem to be achieved by the present invention is to change the trie structure so that the forwarding table used for the address search of the router can be applied at high speed to be applied to a Gbps (Giga bit per second) -class high-speed router By compressing the tri step, and expressing the compressed information as a separate field value to provide an IP address search method to shorten the tri search path.

1 illustrates an embodiment of a conceptual design of a general router.

2 illustrates another embodiment of a conceptual design of a router.

Figure 3 shows an entry in the routing table, which shows a combination of an IP address and a next hop.

4 is a flowchart illustrating an IP address searching method using a bit-vector table according to the present invention.

FIG. 5 illustrates an embodiment in which all prefix entries of a routing table used when searching an IP address are represented by a binary tree of depth 32. Referring to FIG.

FIG. 6 illustrates a forwarding table configured by dividing a prefix tri into four layers of 8 bit depth in order to shorten a search depth of the prefix tri.

FIG. 7 illustrates an embodiment in which a tree is extended to represent an IP prefix as a tree.

FIG. 8 shows a boot-ray configured in units of 8 bits by indicating prefix bit information and child node information of each leaf node 256 bits in two bits.

9 shows the data of a bit-vector table for organizing the entire prefix tri into a wide-priority search order with a bootray of depth 8, and configuring a bit-vector table having each boot-lay as one node. The structure is shown.

FIG. 10 illustrates an embodiment in which a bit-vector table is configured by applying FIGS. 8 and 9.

11 is a program of an IP address searching method using a bit-vector table.

12 shows a table of forwarding experiment results using a bit-vector table.

An IP address retrieval method using a bit-vector table according to the present invention for solving the above technical problem comprises the steps of: (a) constructing a routing table used in retrieving an IP address as a bit-vector table represented by a prefix tri; (b) dividing the IP address to be searched in the bit-vector table by m bits; obtaining the number of child nodes and the number of prefix bits in a bit-vector table by indexing the value divided by the m bits; And (d) searching for an IP address using a cumulative value accumulating the number of prefix bits as a routing index.

In addition, the step (a) may include (a1) expressing a binary tree having a length n of a depth by expanding an IP prefix entry of a routing table used when searching for an IP address; (a2) increasing the number of prefix bits by changing every node to a tri having no child node or having two child nodes in order to express the IP prefix of the binary tree as a tri; (a3) dividing the prefix tri into a booti of depth m to calculate how many bits of the prefix bits exist at the depth m; And (a4) configuring a final bit-vector table by configuring a booti of depth m in the prefix tri as a sequential node of a bit-vector table in a width-first search order.

In addition, in the step (a1), when the routing table is represented as a binary tree, there are k prefix bits, which are the number of routing entries, in the binary tree.

Further, in the step (a2), the prefix bit of the IP address at the depth m of the prefix tri is a count value indicating the number of bits between 0 and k + α bits.

Further, in step (b), the m-bit value means a boot ray and an m-th specific leaf node including the m-th node at the depth m of the prefix tri.

In addition, the specific leaf node is characterized by indicating the position of the prefix bit field and the child bit field of the bit-vector node.

In addition, in step (c), (c1) the value divided by the m bits is used as an index of the bit-vector table, and when there are child nodes at the m th bit position, the number of child nodes accumulated to the current position and the prefix bits Obtaining a number; And (c2) if there are no child nodes in the bit-vector table node indexed with the m bits, and if there are prefix bits, calculating the number of prefix bits again.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

4 is a flowchart illustrating an IP address searching method using a bit-vector table according to the present invention.

First, in step S410, a routing table used in retrieving an IP address is configured as a bit-vector table represented by a prefix tri.

FIG. 5 illustrates an embodiment in which all prefix entries of a routing table used when searching an IP address are represented by a binary tree of depth 32. Referring to FIG.

At this time, the depth of the tree at ³² is the maximum number of 2 ³² IP addresses.

That is, an IP address having a length of 32 bits has a value between 0 and 2 ³² , and a routing table entry also has a value between 0 and 2 ³² . Since the actual routing table entries are composed of less than 2 ³² numbers, such as the prefix bits indicated by dots in FIG. 5, when the routing table is represented as a binary tree, there are prefix bits as many as the number of routing entries at depth 32.

Therefore, the index value of the routing table entry to be found in the routing search of the given IP address is equal to the count value indicating the number of bits between 0 and 2 ³² of the prefix bit of the IP address at the depth 32 of the binary tree. In other words, instead of comparing the IP lookup with the prefix value in the routing table, it can be replaced by searching the binary tree to calculate the number of prefix bits reached at depth 32.

FIG. 6 illustrates a forwarding table configured by dividing a prefix tri into four layers of 8 bit depth in order to shorten a search depth of the prefix tri.

Bits represented by dots in FIG. 6 become places where an IP address prefix exists. The prefix points are the result of extending from the root node according to the bit value (1 or 0), and the specific prefix represented by the point is chosen as the next hop entry by the representative, ie LPM, for the IP address in the range r of FIG. . Also, as shown in FIG. 6, a longer length prefix included in a range of a specific prefix replaces an upper prefix by its range.

FIG. 7 illustrates an embodiment in which a tree is extended to represent an IP prefix as a tree.

That is, every node must be a trie that must have two child nodes or no child nodes. The prefix try expressed in this way is equivalent to arranging all representable prefix bits in order from the perspective of the Leafs node at depth 32, with r = 2 ^(32-n) when the length of the prefix is n It can be seen that having.

In FIG. 7, a node marked black is a node having a prefix, a white node is a node having no corresponding prefix, and an oblique node is a node having a child node or a lower prefix. However, sequential searching of all the original 32-tris consumes a lot of time, so it is necessary to reduce the searching time by compressing the depth of the tri. Applying this concept, a method of constructing a tri layer into three layers (16 bits, 8 bits, 8 bits) and using a bit-vector representing a presence or absence of a prefix has been published. In the method, the forwarding table is composed of several arrays representing the order of appearance of prefix bits using bit-vectors.

In the present invention, in order to shorten the search depth of the prefix tri, the prefix tri is divided into four layers of 8 bit depth as shown in FIG. 6 to create a forwarding table. Can be found. Dividing the entire prefix tri into an 8-bit-depth bootstrap results in a maximum of 2 ⁸ leaf nodes at depth 8, which constitutes one forwarding table node. That is, since all prefix bits of depth 8 become one forwarding table node, and at depth 16, a maximum of 256 forwarding table nodes can be created because a depth of 8 bootray can be formed at which the bits of depth 8 are rooted.

FIG. 8 shows a bootstrap configured in units of 8 bits, and prefix bit information and child node information of each leaf node 256 bits are represented by two bits.

One forwarding table node is a field value that accumulates the prefix bits (tp) and the number of child node bits (tc) of all preceding nodes to reduce the number of shift operations required for IP address lookups, and every 8 bytes of the current node. It has three fields (p1, p2, p3 / prefix, c1, c2, c3 / child) in which the number of bits is recorded.

FIG. 9 shows the data of a bit-vector table for organizing the entire prefix tri into a wide-priority search order with a bootray of depth 8, and configuring a bit-vector table having each boot-lay as one node. The structure is shown.

FIG. 10 illustrates an embodiment in which a bit-vector table is configured by applying FIGS. 8 and 9.

The routing table used in the example of FIG. 10 is composed of six entries, and is first changed to a bit-vector table represented by a graph described in FIGS. 6 and 7. In the example of FIG. 10, the depth of the step is changed to 4 bits for easy explanation, and the p1 to p3 fields are changed to use only the p1 field that is responsible for 8 bits. The configured forwarding table is composed of 6 nodes, and since the depth is 4 bit units, one node can express a total of 16 bits, that is, 16 prefix values.

The number of nodes in the bit-vector table is about 3,700 for about 54,000 real routing tables and consists of 284 Kbytes. This size is small enough to be stored in the L2 cache of the Pentium II 400MHz processor, and the average search speed is 320-530 nsec depending on the size of the routing table.

Second, in step S420, the prefix tri is layered and divided in units of m bits from the IP address of the bit-vector table.

The use of an IP address divided by 8 bits is because the prefix tree is layered into 8 bits. The obtained 8-bit value refers to a leaf node corresponding to the BootRie, and means the positions of the prefixBit and childBit fields of the bit-vector node.

Third, in operation S430, the number of child nodes and the number of prefix bits are obtained from the prefix tri layered into m bits.

When you know that at line 3 and there are child nodes, find the number of child nodes and the prefix bits up to the current position, and use this as an index of the bit-vector table as the search node for the next 8 bits, as shown in line 5. do. Finally, if there are no more child nodes and there are prefix bits, the number of prefix bits is calculated again.

Fourth, an IP address is searched using a cumulative value of accumulating the number of prefix bits as a routing index (S440).

The number of prefix bits obtained as a result of the search becomes the index of the forwarding table entry. However, since the number of prefix bits in the bit-vector table is larger than the number of entries in the forwarding table, the forwarding table cannot be directly accessed by the obtained index. The reason for the large number of prefix bits in the bit-vector table is that binary trees are extended to tri as shown in FIG. Therefore, the prefix bit of the bit-vector table is the index of the extended tri and must use a separate data structure (pointer) that maps the index of the binary tree to the index of the tri. The number of data structures required is equal to the number of prefix bits of the extended trie, and an additional memory read occurs after the search is completed.

If the search for the address 12.5 (0000 1100 0000 0101) is performed using the example of FIG. 10, the process is as follows. The order of the first node to be searched is the order of node 0-> node 1-> node 3-> node 5. Since childBit is in the set (1) state at nodes 0, 1, and 3, as shown in line 3 of FIG. 11, the cumulative number of set bits of the childBit up to the corresponding bit is obtained and used as an index of the next node to move to the next node. Since the prefixBit is set (1) in nodes 1 and 5, the cumulative number of set bits of the prefixBit is obtained as shown in lines 3 and 4 of FIG. Used as an index. However, in the example of FIG. 10, since the value of pointers and the value of the routing table are the same, if the prefix accumulation value is used as the routing index, the IP address 12.5 is completed.

12 is a table showing the results of the forwarding experiment using the bit-vector table.

The forwarding experiment using the bit-vector table was performed using a real routing table in the United States. As shown in FIG. 12, the size of the routing table varies from more than 48,000 backbone routers to enterprise-class routers with thousands of entries. The build time of the forwarding table is about 90 msec to 280 ms, which is the time to construct the bit-vector table from the prefix try in memory. Previous studies have shown that the forwarding table of a backbone router is small enough to be modified once per second.

We used a Pentium II 400MHz processor (L1 cache 32Kbyte, L2 cache 512Kbyte) in Linux kernel 2.2.12 to measure the search time using bit-vector forwarding table. The average search time for Paix routers was 320 nsec and the largest Mae-East router was 530 nsec. The IP address used for the search was generated by a random function, and performance was measured by repeating the search a million times. Since the current general processor has no special way of resident certain data blocks in the L2 cache within our understanding, the performance measurement result of the present invention is the result of the forwarding table mixed in the L2 cache and memory.

The drawings and specification are merely exemplary of the invention, which are used for the purpose of illustrating the invention only and are not intended to limit the scope of the invention as defined in the appended claims or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

According to the present invention, a quick IP address search is possible only by software. A 5 Gigabit high speed router should support a speed that can process packets of an average size of 512 bytes within 800x10 ^-9 seconds. In the present invention, the average packet throughput tested using an actual routing table is 320x10 ^-9 to 530x10- ^{. It} takes ⁹ seconds to implement a 5 Gbps or faster router.

Claims (8)

(a) expressing a binary tree having a length n of a prefix by extending an IP prefix entry to configure a routing table used in retrieving an IP address into a bit-vector table represented by a prefix tri; (b) changing all nodes to tri, which should have two child nodes or no child nodes, increasing α prefix bits; (c) dividing the prefix tri into a booti of depth m to calculate the number of bits of the prefix bits present in the depth m; (d) constructing a final bit-vector table by configuring a booti of depth m in the prefix tri as a sequential node of a bit-vector table in a width-first search order; (e) dividing the IP address to be searched in the bit-vector table by m bits; obtaining the number of child nodes and the number of prefix bits in the bit-vector table by indexing the value divided by the m bits; And and (g) searching for an IP address using a cumulative value of the number of prefix bits as a routing index. delete The method of claim 1, wherein step (a) When the routing table is represented as a binary tree, an IP address retrieval method using a bit-vector table includes k prefix bits, which are the number of routing entries, in the binary tree. The method of claim 1, wherein the length of the prefix in step (a) is IP address retrieval method using a bit-vector table characterized by having a maximum depth of 32. The method of claim 2, wherein the index value of the routing table in step (b) And a count value indicating the number of bits between 0 and k + α bits at the depth m of the prefix tri. 2. The IP using the bit-vector table according to claim 1, wherein the m-bit value in step (e) means a boot ray and an m-th specific leaf node including the m-th node at the depth m of the prefix tri. How to search address. The method of claim 6, wherein the specific leaf node is IP address retrieval method using a bit-vector table, characterized in that the position of the prefix bit field and the child bit field of the bit-vector node. The method of claim 1, wherein step (f) (f1) calculating the number of child nodes and prefix bits accumulated to the current position when there are child nodes at the m th bit position using the value divided by the m bits as an index of the bit-vector table; And (f2) IP address retrieval using a bit-vector table, comprising calculating the number of prefix bits if there are no child nodes in the bit-vector table node indexed with the m-bits and if there are prefix bits Way.