TW508925B - A hardware architecture for IP forwarding table lookup - Google Patents

Abstract

An IP forwarding table lookup function is required in a router to decide the next hop of an IP packet. Upon receiving an IP packet, a router extracts the destination address of the packet for routing table lookup. Because of subnetting and supernetting, IP table lookup is based on longest prefix matching. In this invention, a hardware architecture of forwarding table lookup is designed for IPv4 backbone routers. Our invention uses conventional memory and burst mode SRAM. The advantages of our invention include high memory usage efficiency, large table capacity, low update complexity, and high lookup speed.

Description

V. Description of the invention (1) -------, Fields of application in the industry-This paper proposes a hardware look-up table architecture for backbone routers using 1 Pv4 'using traditional memory with clustered static memory ( Burst mode 6) Make the table lookup function possess the characteristics of memory usage efficiency, large transfer table, low data update complexity, and fast search. Background IP forwarding table (IP forwarding table) is a key technology of routers. Routers must determine the routing of IP packets by searching forwarding tables. There are two types of lookup tables: software lookup tables; for example, Small Forwarding Tables for Fast Routing Lookups, and Fast Address Lookup Tables for Internet Routers for Internet Routers), Scalable High Speed IP Routing Lookups and IP Address Lookup in Hardware for High-Speed Routing The memory access speed is performed by the Hardware Lookups in ψ Hardware at Memory Access Speeds. At present, the software lookup table relies on the central processing unit (CPU) in the router to process the transfer table with a special data structure algorithm. The purpose is to enable the entire transfer table to

508925 V. Description of the invention (2) It is stored in the internal cache memory of the central processing unit to improve the query speed, but the processor will spend a lot of power to perform tedious data structure calculations, and it is easy to encounter bottlenecks in search speed, so it is suitable For low-speed area routers, low cost is its advantage. On backbone networks where speed is critical, fast hardware lookup tables must be used. The hardware look-up table can use the access to the memory to perform the table look-up action. As long as the memory access speed is fast enough, it can meet the requirements of practical applications. Although it costs more hardware, it has a shorter search time. Known technology 丨 _ Routers can be roughly divided into backbone routers and area routers based on the network scale connected to them. These routers have different requirements for table lookup technology. Generally, the routing table recorded in the area router used in the local area network or the corporate network is small, about one thousand pieces of backup data, and the speed of the input and output ports is not high, and the time for checking the table is relatively loose. The backbone router has a much larger forwarding table because it is connected to the WAN. According to (IPM A) statistics for the four major routers in the United States (as shown in Table 1, its forwarding table contains about 30,000 to 40,000 pieces of data on request, so it cannot be executed directly using (CAM, content addressable memory) Hardware checklist. In addition, the line rate of its input and output ports is much higher than that of the local area network, and the search time must be shorter. As the network speed increases and the number of users increases, the performance of the router

Page 6 508925 V. Description of the invention (3) It must be improved, some actions handled by software will be replaced by hardware, and the lookup table is no exception. Although "Nicli Mckeow, Steven Lin and Pankaj

Gupta published his paper "Roting L〇〇kup in Hardware at

Mem ° T ^ Access Speedn also proposes a low-speed high-speed data rate for the (ΪΡν4) router, and encounters the transfer table update: ΐ: go 'but the cost of memory is too high, you must use 3 3 MB of memory space for storage, Difficulty using (update). Another type of hardware lookup table is f X ,, and CAM (Content Addressable [e m r r y) is used.

Alas, the next day's signature for this Λ can quickly find the exact equivalent of the search to achieve the longest prefix match :: a memory space. The disadvantages are :! Table, and then to compare the correct results, you must use multiple sets of CAMs to check the height at the same time, and the power consumption is too large to make large = 'C touch the hardware to form the complexity of the CAM memory set by the table lookup tool of the backbone network router , So uninventive

The primary object of the present invention is to disclose a hardware look-up table architecture, which uses properties of burst mode SRAM, large transfer table capacity, and lookup. Φ 2 Backbone router using (IPv4) proposed ¥ ^ first memory with clustered static memory to make the table lookup function with high memory usage and the complexity of data update and fast

Page 7 508925 V. Description of the invention (4) Anyone who is familiar with the technology will undoubtedly know the objects and advantages disclosed by the present invention after reading the following detailed description of the preferred embodiments shown by different diagrams. The detailed description of the invention is the way; the transfer tree table is continuously displayed as a route table that can find a router, and the node outputs the second continuous storage range. The router needs to obtain the current request. The feature layer makes the bottom layer exist. It is necessary to have the function of the IP forwarding lookup table to receive the destination bit routing information in the IP header. The IP forwarding table first uses traditional memory to directly narrow the search range, and then uses the comparison and comparison technique in the bottom (3 2-η) layer. Therefore, this memory has a higher internal backup node. Not much, expected. The following is the details of the present invention. After determining the IP packet's routing address, it is necessary to perform a linear search by searching the data structure for binary static memory points recorded in layers 1 through n. The efficiency of the nodes that must be compared. Because each time can be explained in detail after several comparisons. The present invention provides a hardware look-up table architecture for IP packet forwarding, which includes traditional memory, cluster static memory (B-SRAM), comparator, and data selector. The traditional memory is used to store the first n bits of IP value. The data structure corresponding to the element. Its data output is connected to the cluster static memory (B-SRAM) address line and data selector. The function is to output the search result or instruct the second stage search.

Page 8 508925 V. Starting address of invention description (5). The cluster static memory is used to store the lean structure corresponding to the last (3 2 _ n) position of the I p value. Its data is connected to the comparator in turn, and 丨? The last (3 2-n) values are compared and the control signal of the selector is accepted. A set of comparators 'for parallel search of the last IP value (32 ~ n) bits (LSB) and cluster static memory (B-SRAM) round-out data to improve table lookup speed' and compare the results Out to the selector; the number of comparators can be determined by the actual application. The invention also uses a selector, which is used to select the result that meets the longest prefix principle from the traditional memory or comparator, and judge whether the search ends or not; it is stored in the traditional memory and the cluster static memory (B-SRAM) The special data structure inside makes the memory highly efficient. Functional block diagram: The functional block diagram of the present invention is shown in Figure 1. It includes the I p address (most significant n bits of 1F, 1), traditional memory (RAM '2), and cluster static memory ( Burst mode SRAM, hereinafter referred to as B-SRAM '3), a comparator (comparat〇r, 4), and a data selector (selector' 5), a control signal (9), and next hop informati On, 10). The traditional § self-memory body is used to store the query results of the first n bits (MSB s) of the I p value for the deer; the cluster static memory (B-SRA Μ) is used to store the end of the I p value (3 2- η) query result corresponding to bit (LSB s); the comparator is used for

Page 9 刈 8925

(2) == SBs 丄 's check; the selector selects the first, XT, and XT results, and judges whether the query is over or not. The following describes the data format: · Traditional memory stores data

1 Port number 0 Index to B-SRAM, and the ^ set of static memory (MSB) is used to indicate whether it has been ^ at the _th level, soil =. If it is 1, it means to find the data of ‘symbol = first son of bin’ in the first-level tree data structure, and the remaining bits are the output port number of the packet. If ϋ 〇, you must go to the second layer to search, and the remaining bits record the B-SRAM index. , Jing ~ ,. Clustered static memory (B-SRAM) stores data; 2 part IP address Port number Prefix mask This data format provides the second layer of remaining IP bits for comparison, including the remaining bits of the requested data, and The output number corresponding to this request data. Since the longest prefix requirement must be followed, a prefix mask must be provided

Page 10 508925 V. Description of the invention (7) Tell the comparator (4) the number of remaining bits to be compared, and design this mask—the largest value that can be represented is greater than the length of the second layer of IP to be checked to represent the last stroke data. Table lookup procedure: · The table lookup can be divided into two stages. First, the first n bits of the IP address are directly mapped to the address memory of the traditional memory, and the first layer is searched. If the MSB of the output data is 1, then The data selector outputs the remaining bits and instructs the search to complete. & If the search is not completed in the first layer, the output is an indicator pointing to the cluster static memory (B-SRAM). At this time, the cluster static memory (B-SRAM) will be sent out for each clock cycle. The data is compared with the unmatched IP bits (32-η bits) in the comparator. As long as the bus width of the cluster static memory (B-SRAM) is sufficient, multiple sets of comparators can be used for parallel comparison. In order to make the search efficient, the backup data in the cluster static memory (B-SRAM) can be arranged in accordance with the prefix length when building the table, so that the data successfully matched for the first time must match the longest word. First, the data selector can select a port number corresponding to the largest prefix mask among these comparators, and stop the table lookup operation. You do not need to search the entire section to ensure that the data that meets the principle is found. In addition, when reading When the entered prefix mask is greater than the number of unmatched IP bits, it indicates that the segment data has been retrieved and the search must be stopped.

Page 11 508925 V. Description of the invention (8) Data update: When the transfer table was first created, a fixed proportion of space was reserved in the cluster static memory (B-SRAM) according to the number of data on request with the prefix length exceeding η. New nodes are added, so there is no need to change the data in the entire clustered static memory (B-SRAM). To update the information on request with the prefix length less than η, you only need to change the data in the traditional memory. When a node whose prefix length exceeds η is updated, it must first use traditional memory to correspond to cluster static memory (B-SRAM). If it is required to arrange the prefix length from large to small, you need to segment this section. All the data on request are reordered and updated, otherwise you only need to add one data to the reserved space. Because there is not much backup data in each clustered static memory (B-SRAM) section, up to 64, plus a few paths with a prefix length greater than 25, and only the data that needs to be updated, so This design prevents the search bottleneck when updating the transfer table. Hardware architecture and operation: The hardware architecture of the present invention is shown in Figure 2. It includes a data selector (5), memory (2 5 6K * 2 RAM, 6), and cluster static memory (32K * 9). B-SRAM, 7), special application integrated circuit (8), next hop information (10), comparator (comparator 11), comparator 12 (comparator 12), comparator (comparator) 13) 、 Cluster mode control line

Page 12 508925 V. Description of the invention (9) line, 15), special application integrated circuit (table lookup ASIC, 16) that performs table lookup function. The address line input of the traditional memory is the first n bit of the IP value to be checked, and the data output line is connected in parallel with the address line of the cluster static memory (B-SRAM) and the input of the selector. The selector checks the MSB of this data. If it is 1, it outputs the port number and ends the search. Otherwise, the cluster static memory (B-S R AM) cluster output function is activated. When the cluster output function of the cluster static memory (B-SRAM) is activated, the output data of the traditional memory is the entry address of the cluster static memory (B-SRAM); the comparator changes the cluster static The data continuously output from the memory (B-SRAM) is compared with the remaining (32-n) bits one by one, and the result is output to the selector. Because several comparators simultaneously output the search results, the selector must be based on ), Select the result that meets the rules, and then stop the search action. Design Example: The present invention designs a structure as shown in FIG. 2 according to the existing memory specifications. Divide the IP address into 18, 14 bits for searching, and use the traditional memory of 256K X 16 bit s to search for layers 1 to 18, deducting (MSB), and the remaining 15 bits exactly point to two 32K X 36-bit cluster static memory (B-SRAM). These two cluster static memory (B-SRAM) are connected in parallel to form a line.

Page 13 508925 V. Description of the Invention (ίο) The width is 3 2 K X 7 2 b i t s memory block. Three comparators can be used for parallel comparison of valid nodes in the layers 1 to 3 2 *. A total of 800 K b y t e s memory capacity is used. In terms of transfer table capacity, each cluster static memory (B-SR AM) physical address can store three pieces of data on request, and theoretically can record up to 96K pieces of data. Actually, it is necessary to deduct the address reserved by the node in order to facilitate the update, so the effective capacity will be less than 96 K, depending on the reserve ratio. Performance analysis: «According to the statistics of the I PMA, as shown in Figure 3, the longest prefix between 8 and 24 digits accounts for more than 99% of the total available data, and most of them are concentrated in the 24th layer. In view of this property, the present invention assumes that the maximum number of searches occurs within the search of 18 to 24 layers to perform the analysis of the worst performance. Pushing from a node of the 18th layer to the 24th layer can grow into a complete seven-layer binary tree, which contains 64 valid nodes. In the worst case, the three sets of comparators must compare these data to The result is that the B-SRAM must be output 22 times. If a cluster static memory (B-SRAM) with an access time of 10 ns is used, plus the memory access time of the first stage search, the total time is estimated to be within 300 ns, which is the worst. In the case, it can still reach 3.3 million table lookups per second. However, the distribution of valid nodes in these layers is actually scattered.

Page 14 508925 V. Description of the invention (11), because there are only about 40,000 effective nodes in total, a densely distributed structure like this has a very low probability of appearing. In most cases, the results can be obtained after several comparisons, and the average number of table lookups per second will be much higher than 3.3 million times.

Page 15 508925 Schematic illustrations Schematic descriptions: Figure 1 Functional block diagram of the present invention Figure 2 Hardware architecture diagram of the present invention Figure Three prefix length distribution (Prefix Daily Length Distribution, which 47 The first row of bold numbers represents the prefix length, and the contents of the table represent the number of data on request for each prefix length, network name. MAE-EAST, MAE-WEST, ADDS, PacBell) «Drawing number description: l · · · ·. IP address 2 · · · · · Traditional memory (RAM) 3 · · · Cluster static memory (Burst mode SRAM, B-SRAM) 4 · · · · Comparator (comparator) 5 ... · · And data selector (selector) 6 · · · · · Memory (2 5 6K * 2 RAM) 7 · · · · · Cluster static memory (32K * 9 B-SRAM)

8 ···· Special application integrated circuit 9 ··· Control signal 1 0 ···· Next trip information (next hop inf or mat i on) ll · ···· Comparator 12… Comparer (comparat〇r) 508925 Brief description of the diagram 13 ... Comparator 15… Burst mode control line 16… .Perform table lookup Special application integrated circuit (table lookup ASIC)

Page 17

Claims (1)

508925 6. Scope of patent application 1. A hardware look-up table architecture for IP packet transfer, which includes traditional memory, clustered static memory (B-SRAM), comparator and data selector. Comparator and selector As a feature; the traditional memory is used to store the data structure corresponding to the first n bits of the IP value, and its data output is connected to the cluster static memory (B-SRAM) address line and data selector, which is used to output The search result or indicates the starting address of the second-stage search; the cluster static memory is used to store the lean material structure corresponding to the last (32-η) bits of the IP value, and its tributary output is connected to the comparative origin. The end of IP (32-η) value is compared; a set of comparators is used to search the end of IP value (32-η) bit (LSB) and B-SRAM output data in parallel to improve the lookup table speed, compare The number of registers can be determined according to the actual application requirements. A selector is used to select the result that meets the longest prefix principle from traditional memory or comparator, and judge whether the search ends or not. It is stored in traditional memory and cluster type. Static memory (B-SRAM) The special data structure, so that memory with high efficiency. Page 18