One embodiment of the invention relates to communications and computer systems, especially networked routers, packet switching systems, and other devices; and more particularly, one embodiment relates to tree bitmap data structures and their use in performing lookup operations (e.g., longest prefix matching, etc.).
This invention relates to a method for storing and managing data. More particularly this invention relates to a data structure representing a set of prefixes and a method for searching the data structure. The data structure and the search method are particularly advantageous for implementation in digital computer hardware. The primary application of current interest is to semiconductor integrated circuits used for packet classification in high-speed, multiservice Internet routers. However, the technique may be useful in a variety of applications involving data that needs to be prioritized or wherein structure in the data needs to be determined and then to be classified. As a result of classification of the data, action on data can be taken more quickly and efficiently.
In the past, in the packet-switched networking field, Internet Protocol (IP) traffic handling processes have required relatively complex searches and analyses of packet header information to develop routing and processing instructions. Recently, IP destination address lookup issues relating to techniques for routing a packet to a destination have received a great deal of attention from the networking community. The following references may be consulted on the state of the art related to the invention herein described:
[1] M. Degermark, A. Brodnik, S. Carlsson, S. Pink, “Small Forwarding Tables for Fast Routing Lookups” Proc. ACM SIGCOMM '97, Cannes (14-18 Sep. 1997) (describing the Lulea algorithm).
[2] V. Srinivasan, G. Varghese, “Faster IP Lookups using Controlled Prefix Expansion,” Proc. SIGMETRICS '98.
A number of patents of interest on lookup algorithms are known in the art.
It is helpful to understand the background of the closest known method for searching tree structures of data. Note the following definitions:
A trie is a tree structure organized for searching.
A trie element is that portion of a tree structure at a single node.
A prefix is a string of characters that appears at the beginning of a longer string of characters.
A stride is a number of levels in a tree accessed in a single read operation.
A target string is a string of characters to be classified.
In many cases of practical interest the characters in a prefix are binary digits, i.e., ones and zeroes. In the example prefix database 10 in
A prefix database can be used to classify strings of characters into subsets according to which prefix a particular string matches. The prefix database structure can include exact matches with a target string, i.e., the data to be analyzed can appear in the database.
The matching process is ambiguous unless the classification requires that the matching prefix be the longest possible match in the database. For example, in
A database of binary prefixes can be represented by a binary tree 20 specialized for searching. This search tree 20 is an example of a data structure called a trie and is convenient for finding the longest prefix match. Nodes (shown as small circles in
The search for a match to a target string progresses by examining one bit of the target at a time. Before examining the target's first bit, a match with the Prefix Node 21 for P1 is known to exist, since P1 matches all strings. If the first bit is a 0, take the left branch 22 of the trie. If the first bit is a 1 take the right branch 23 of the trie. In either case, determine if a Prefix Node is encountered. In the example of
The process proceeds by taking the left or right branch according to whether the next bit is a 0 or a 1, noting the Prefix Nodes encountered and tracing a path down the trie until either a final node 26-30 is reached (a leaf) or there is no longer a path on the trie (since the node is not in the database). In either case, the last Prefix Node found represents the longest prefix match.
The Lulea algorithm is the algorithm known to encode prefix tables most compactly. The Lulea algorithm is a scheme that uses multibit trie elements containing bitmaps to minimize wasted storage. The Lulea scheme does well in terms of worst case storage, using only 200 Kbytes of memory to store the current MAE East database in a fully structured form. (The MAE East database is a database that contains over 40,000 Internet prefixes for IP addresses and requires over 900 Kbytes of ASCII character storage in its uncompressed form.) One way to describe the Lulea scheme is to consider the database 10 of
The various possible target strings are shown in
This remapped trie 100 is called a controlled prefix expansion and is subject, in
Conceptually the Lulea scheme starts with a leaf-pushed trie 101 and replaces, with a single value, all the consecutive entries in a trie element that have the same value. This can greatly reduce the amount of storage required in a trie element. To allow trie indexing to take place even after the elements have been compressed, a bit map with 1's corresponding to retained positions and with 0's corresponding to the removed positions is associated with each compressed element.
For example consider the partially compressed root element 102′ in
The search of a trie element now consists of using a number of bits specified by the stride (e.g., 3 in this case) to index into each trie element starting with the root and continuing until a prefix label or a 0 is encountered. In the latter case, it is necessary to determine the appropriate action. For example, assume the data structure shown in
The Lulea scheme specifies a trie search that uses three strides of 16, 8 and 8. Without compression, the initial 16 bit array would require 64K entries of at least 16 bits each, and the remaining strides would require the use of large 256 entry trie elements. With the compression specified in the Lulea scheme, the MAE East prefix database requires only 200 Kbytes of storage. However, the use of leaf pushing makes insertion inherently slow in the Lulea scheme, since a single update can require the modification of many elements of the trie.
In the absence of deadlines and for small prefix databases, known software methods are sufficient to search for the longest prefix match. However, for many practical applications, such a packet classification in large dynamically-updated databases, there are deadlines or time limitations. Furthermore, the prefix database can contain many tens of thousands, or soon many hundreds of thousands, of entries which can slow a search. The resulting binary trie is exceedingly large and populated only sparsely with Prefix Nodes. What is needed is a method that facilitates a rapid search and minimizes the storage required.
Disclosed are, inter alia, methods, apparatus, data structures, computer-readable media, mechanisms, and means for defining, creating and using tree bitmap data structures, such as for, but not limited to their use in performing lookup operations (e.g., longest prefix matching, etc.).
One embodiment includes a data structure which includes a tree bitmap for identifying for each node of multiple nodes within a stride of a number of tree levels greater than one whether each node is a prefix or vacant node, the multiple nodes representing multiple tree levels, a lowest level subset of the multiple nodes corresponding to a lowest level of the tree levels in the stride, the lowest level subset of the multiple nodes including two or more nodes; a child bitmap for identifying which trie paths emanate and which trie paths do not emanate from the lowest level subset of the multiple nodes; a result array; and a trie element, the trie element including the child bitmap, a pointer to the tree bitmap, and a pointer to the result array. In one embodiment, the data structure includes a child array; and wherein trie element includes a pointer to the child array.
One embodiment includes a data structure which includes a first segmented tree bitmap and a second segmented tree bitmap for identifying for each node of multiple nodes within a stride of a number of tree levels greater than one whether said each node is a prefix or vacant node, the multiple nodes representing multiple tree levels, a lowest level subset of the multiple nodes corresponding to a lowest level of the tree levels in the stride, the lowest level subset of the multiple nodes including two or more nodes; a first child bitmap; a second child bitmap; a first result array; a second result array; a first segmented trie element including the first segmented tree bitmap, a pointer to the first child bitmap and a pointer to the first result array; and a second segmented trie element including the second segmented tree bitmap, a pointer to the second child bitmap, and a pointer to the second result array. The first child bitmap and the second child bitmaps identify which trie paths emanate and which trie paths do not emanate from the lowest level subset of the multiple nodes for the first segmented tree bitmap and the second segmented tree bitmap, respectively.
One embodiment performs lookup operations on a data structure. The data structure typically includes a first trie element, with the first trie element including a tree bitmap for identifying for each node of multiple nodes within a stride of a number of tree levels greater than one whether said each node is a prefix or vacant node, the multiple nodes representing multiple tree levels, a lowest-level subset of the multiple nodes corresponding to a lowest level of the multiple tree levels in the stride, the lowest-level subset of the multiple nodes including two or more nodes; a child bitmap for identifying which trie paths emanate and which trie paths do not emanate from the lowest-level subset of the multiple nodes; and a pointer for identifying a link to a next-level trie element.
The appended claims set forth the features of the invention with particularity. The invention, together with its advantages, may be best understood from the following detailed description taken in conjunction with the accompanying drawings of which:
According to the invention, a means is provided for rapid search for the longest matching prefix from a prefix database. First, with reference to the trie 20′ of
The trie elements 52, 54, 56; and 58, 60 in the lower two trie levels of the illustrated data structure are each linked to their parent trie element by a single Child Pointer 61, 62, 64 which locates all children associated with a parent. This is made possible by storing all of the parent's children in a contiguous block, herein called a Child Array 66, 68, 70. In some cases (those trie elements 54, 58, 60 containing prefix labels P4, P9, P8), there are no children. In other cases in the example there are one or more children. In general, there may be up to as many as 256 children for a stride of eight.
Thereafter, an action must be taken if a longest matching prefix is found, the action being based on the particular prefix found. In the case of packet classification the result sought is often the next hop address, but may be the disposition of the packet or some modification of the packet header. In general, the results (each being a value that indicates what action is to be taken) are also stored in the Child Array. In the example of
Next, according to the invention, a trie element is coded into a compact bitmap as illustrated by
The locations of children that are linked to the parent must also be coded in the trie element 360. In the example of the root trie element 50, paths 270, 271, 272 to the children occur for target strings 100, 101 and 111. (These 3-bit binary strings correspond to the decimal numbers 4, 5 and 7.) Numbering the bits in the child bitmap 352 from left to right as 01234567, the children are represented by 1's in the positions 4, 5 and 7, i.e., 00001101. In the general case, for a stride of s bits, the length of the child bitmap is 2s bits.
Finally, the child pointer 354 must be coded into the trie element 360. Assume b is the number of bits required in the child pointer 354. The storage for each trie element 360 is b−1+2(s+1) bits. In the example shown in
It is desirable to have each trie element stored in a fixed number of bytes. This will be the case for a fixed stride and fixed size child pointer. However, the results (i.e., actions to be taken) also need to be stored adjacent to a block of child Child Elements. For memory management purposes, such results are placed in one or more storage units whose size is equal to that of a trie element. Assume that n results can fit into a single storage unit of memory, where n depends upon the size of a result field.
Required Computations
After a trie element has been retrieved from memory there are several required computations in order to proceed along the search path to subsequent trie elements. The first is indicated in Table 1 where the first stride of the target string is assumed to be 100 (binary). Tree Level 0 (
In Line 2, these binary coded strings are transformed into unary coded strings. For example, in the entry for the child bitmap the binary coded string in Line 1 represents the integer four. The unary coded version in Line 2 numbers the bits from left to right 01234567, sets the bit numbered four to 1 and then sets the remaining bits to 0 resulting in the code 00001000. In Line 3 the tree structure 350 in
The second step in the required computations is shown in Table 2. Line 1 corresponds to Line 3 in Table 1 except that all the tree levels are merged into a single tree bitmap 350′. In Line 2 each bit that is a 1 in Line 4 of Table 1 is extended all the way to the left, i.e., to the limit of the size of the field. As a result the Extended AND is all 1's to the left of and including the longest match bit and is all 0's thereafter. A similar extension is applied to Line 4 of the child bitmap. Taking the bit-by-bit AND of the first two lines produces the Offset Maps of Line 3 of Table 2. Here it is only necessary to count the number of 1's to get the Match Offset m and the Path Offset p. Match Offset m is the number of 1's in the tree bitmap up to and including the longest prefix match to that point in the search path. The Path Offset p is the number of 1's in the child bitmap up to and including the one that connects to the next child in the search path. One alternative for the structure of the complete Child Array, is shown in
The memory address calculations in the Child Array are shown by an example in
Architecture for the Required Computations
The computations described in Tables 1 and 2 are carried out in a Finite State Machine (FSM) 500 and its associated combinational logic 502 shown in
The Combinational Logic 502 is based on a fixed stride s and produces the logic signals m and p shown in the last line of Table 2, where m is the Result Element Offset, p is the Child Element Offset. The technique for actual computation of the parameters m and p will be explained hereinafter.
Flow Chart Description of the Operation of the Finite State Machine
As can be seen in
After the data representing the Root Element appears in D, all information needed by the Combinational Logic is available, and the following decisions based on the values of m and p can be made while in state T=2 (Step D) There are four cases (labeled by paths 1, 2, 3, 4):
1) m>0; p>0: There is a match and a path to a Child Element.
2) m=0; p>0: There is no match, but there is a path to a Child Element.
3) m>0; p=0: There is a match, but there is no path to a Child Element.
4) m=0; p=0: There is no match and no path to a Child Element.
Analyzing these cases from left to right in the flow chart of
1) Load the Result Address RA(A,m) into L based on the Child Pointer in D and the Match Offset m (Step E). Also, concurrently, load the Child Address CA(D,p) into memory address register A based on the Child Pointer in D and the Path Offset p (Step F). These two steps, both taken in state T=3, save the longest match found in the current trie element in the L register and load the address of the next trie element to be searched into the memory address register A.
2) Since there is no match in this case it is only necessary to load the memory address register A (Step F) as in 1).
3) There is a match, but no child exists, so the Result Address RA(D,m) is loaded into the memory address register A (Step G). In state T=4 the result value in memory address register A (with a possible barrel shift) is fetched and loaded into the H register (Step J).
4) In this last case there is neither a match nor a child, so the longest match stored in L is loaded into the memory address register A (Step H). As in path 3) the resultant value of the subsequent memory access is loaded into the H resister (Step J).
For the paths 1 and 2, no result has been obtained so the target string must be searched at the next level (Step K). The S register is shifted left a stride to expose the next s bits of the target string (Step B) and the process is repeated at the next trie level. For paths 3 and 4, a result has been obtained, so the FSM 500 is re-initialized and loaded with the next target string in preparation for starting a new search at the Root Element (Step A).
Combinational Logic Design
The second level AND gates (x7612 and x8616) carry out a bit-by-bit AND yielding the Offset Maps in Line 3 of Table 2. The only step that remains is to count the 1's in the Offset Maps yielding m with the Count7 block 622 and p with the Count8 block 624. This step can be accomplished by a table lookup using techniques familiar in the implementation of Programmed Logic Arrays (PLAs). Both these steps can be augmented for large strides by multiple PLAs and an adder tree.
The remaining logic is familiar to those skilled in the art. For example the Child Pointer calculation requires a simple adder. The Result Address calculation requires an adder and a shift provided that n is limited to powers of 2. The S register 514 must be a shift register and the memory output logic must include a Barrel Shifter.
The Combinational Logic described in
Optimizations, Ramifications and Conclusions
The specific embodiment described hereinabove is one of many alternative designs for rapid prefix matching that employ the Tree Bitmap concept. It is, however, an excellent balance between simplicity and speed. Some of the other alternatives are described briefly below and others, in addition, can be easily developed by one of ordinary skill in the area of digital system design who follows the teachings of the previous section.
Other strides: The example shown above utilizes a stride of s=3. Other strides, both smaller and larger, can be used. The most practical values are in the range from s=2 to s=8. Given a prefix database of up to 500,000 prefixes, the Child Address pointer would have to contain 19 bits. Current databases are an order of magnitude smaller and could be implemented with a pointer of only 16 bits, but it is clear that the prefix database size will grow substantially within the next few years. For the various strides in this range and with a 19-bit pointer, the size of a trie element is shown below in the second column of Table 3.
As the stride increases, the memory bandwidth required first decreases slightly and then increases. Also as stride increases the number of memory accesses needed to examine all prefixes in a search path decreases as does the time required for a complete search. The Bandwidth-SearchTime product is minimized for strides between 3 and 6. However, other considerations such as advanced memory technology may recommend strides outside this range.
Mixed memory technology: It may be desirable to mix on-chip SRAM with off-chip DRAM in longest prefix matching systems. The on-chip SRAM is fast but expensive. The off-chip DRAM is slower but less expensive. By combining these two technologies and taking advantage of a storage structure and stride appropriate for each technology, it is possible to gain the advantages of both. One technique for accomplishing this combination is through use of a pipelined architecture with the first strides accomplished through on-chip SRAM in the first stage of processing, the next strides accomplished through off-chip DRAM in the next stage of processing and any remaining very long prefixes searched through on-chip SRAM in the final stage of processing. Each processing stage should have the storage structure, the number of bits in the stride and number of strides in the stage optimized for search speed and storage capacity.
In
Split Tree Bitmaps: Reducing memory bandwidth while keeping the stride constant is possible by storing separately the Tree and Child Bitmaps and thereby reducing the size of each random memory access.
For split tree bitmaps to function properly, each child must have a bit indicating if the Tree Bitmap associated with its parent contains a prefix that is a longest match so far. If there was a prefix in the path, the FSM records the location of the Tree Bitmap containing the longest matching prefix thus far. This location is known from the data structure of the last trie element. When the search terminates, the FSM must access the corresponding Tree Bitmap and Result Array. Notice that the first described algorithm accesses the result information only when actually needed; the split tree algorithm accesses even the Tree Bitmap only when needed. This is possible because any time a prefix P is stored in a trie element X, all children of X that match P can store a bit noting that the parent has a stored prefix on the search path.
The split tree algorithm requires some added complexity in the FSM in order to defer access to the Tree Bitmap until the search terminates, but can be accomplished with only minor modifications of the flow chart of
Segmented Bitmaps: After splitting the Tree and Child Bitmaps into separate parts, the size of the trie elements may still be too large for the optimal burst size associated with the system memory. The next step for reducing the element size is to segment the trie element itself. The goal behind segmented bitmaps is to maintain the desired stride while keeping the storage space per trie element constant, but reducing the memory data path width. The simplest case of segmented bitmaps is shown in
Because each segment of a trie element has its own pointers, the children and result pointers of other segmented trie elements are independent. Since the segments are stored contiguously, the high order bits (of the bits that would normally have been used to access the trie element) can be used to access only the required segment. Thus roughly only half the bitmap size needs to be accessed. For example, using 8-bit strides, this could reduce the bit map accessed from 256 bits to 128 bits.
Content Addressable Memory (CAM) Elements: Empirical results show that a significant number of trie elements contain only a few prefixes. In these cases the space normally occupied by a Tree Bitmap, a Child Bitmap and a pointer to a Child Array or Result Array can be replaced by a number of entries that implement a content addressing function either in software or hardware. There are well known algorithms to accomplish this since it corresponds to searching a list of prefix fragments for a match to the bits in the current stride of the target string. Commercial hardware technology is available that also has this functionality (see for example: “What is a CAM (Content Addressable Memory)?” Application Brief AB-N6, MUSIC Corporation, Hackettstown, N.J., (presented as a tutorial at http://www.music-ic.com/). Which elements of the trie data structure should be replaced by CAM elements is determined primarily by the storage saved by the replacement and can be determined empirically. Generally, elements with very few prefixes or children are candidates to be replaced by a CAM element, but as the number of prefixes or children grows diminishing returns are quickly reached.
Multi-Field Packet Classification: The direction of state-of-the-art router technology is toward packet handling functions that require header information that is increasingly deep within a packet. Thus, the direction is to move beyond looking up the destination address (Layer 2) toward more general packet classification (Layers 3 and 4) which is required by multicast, tunnels, virtual private networks (VPNs), QoS support and security filters.
The relevant information is contained in a series of header fields 946 (
The most common type of filter can be reformatted and recognized as either an exact match or a single prefix match (950-988). The remaining filters (990-998) are small in number and can be handled by CAM techniques that execute concurrently with the Tree Bitmap method or with any other implementation of a prefix search.
The Tree Bitmap method described hereinabove is the preferred implementation of a prefix search. It can be used to classify packets by applying filters based on information from the five header fields mentioned above. The method is to create a larger trie that contains a compact representation of these filters, but is still in either exact match or prefix match form and then use the process described hereinabove.
An example of how this method can be applied is shown schematically in
In
In conclusion, the data structure and search method of this invention provides for the rapid search of a database for the longest prefix match. The method is well suited to implementation in computer hardware, provides a compact storage structure, is scalable to large databases and can be applied to a pressing problem in the design of Internet multiservice routers.
There are a number of optimizations and ramifications suitable for adjusting the preferred embodiment to various specialized needs. Although the descriptions above contain many specific details, they should not be construed as limiting the scope of the invention, but merely providing illustrations of some of the presently preferred embodiments of this invention. Thus the scope of the invention should be determined by the appended claims rather than the example given.
This is a continuation of application Ser. No. 10/369,354, filed Feb. 18, 2003, (U.S. Pat. No. 6,728,732 issued Apr. 7, 2004), which is a continuation of application Ser. No. 09/371,907, filed Aug. 10, 1999, (U.S. Pat. No. 6,560,610 issued May 6, 2003), with all of these applications and patents hereby incorporated by reference.
The invention was performed in the course of work under sponsorship of DARPA, an agency of the Department of Defense, under Contract No. DABT-63-95-C-0083, and of the National Science Foundation, under NSF Grants NCR-9714698 and ANI-9318178 of the CISE Directorate. The U.S. Government retains certain nonexclusive rights in this invention.
Number | Name | Date | Kind |
---|---|---|---|
5088032 | Bosack | Feb 1992 | A |
5528701 | Aref | Jun 1996 | A |
5574910 | Bialkowski et al. | Nov 1996 | A |
5627748 | Baker et al. | May 1997 | A |
5651099 | Konsella | Jul 1997 | A |
5721899 | Namba | Feb 1998 | A |
5740171 | Mazzola et al. | Apr 1998 | A |
5781431 | Duret et al. | Jul 1998 | A |
5781722 | Wilkinson, III et al. | Jul 1998 | A |
5787430 | Doeringer et al. | Jul 1998 | A |
5798762 | Sfarti et al. | Aug 1998 | A |
5809501 | Noven | Sep 1998 | A |
5829004 | Au | Oct 1998 | A |
5848416 | Tikkanan | Dec 1998 | A |
5856829 | Gray et al. | Jan 1999 | A |
5857196 | Angle et al. | Jan 1999 | A |
5884297 | Noven | Mar 1999 | A |
5920857 | Rishe et al. | Jul 1999 | A |
5920886 | Feldmeier | Jul 1999 | A |
5995971 | Douceur et al. | Nov 1999 | A |
6011795 | Varghese et al. | Jan 2000 | A |
6014659 | Wilkinson et al. | Jan 2000 | A |
6018524 | Turner et al. | Jan 2000 | A |
6041053 | Douceur et al. | Mar 2000 | A |
6044408 | Engstrom et al. | Mar 2000 | A |
6067574 | Tzeng | May 2000 | A |
6115716 | Tikkanan et al. | Aug 2000 | A |
6141738 | Munter et al. | Oct 2000 | A |
6185559 | Brin et al. | Feb 2001 | B1 |
6236658 | Essbaum et al. | May 2001 | B1 |
6237061 | Srinivasan et al. | May 2001 | B1 |
6243667 | Kerr et al. | Jun 2001 | B1 |
6289414 | Feldmeier et al. | Sep 2001 | B1 |
6295576 | Ogura et al. | Sep 2001 | B1 |
6298339 | Bjornson | Oct 2001 | B1 |
6334153 | Boucher et al. | Dec 2001 | B2 |
6341346 | Benayoun et al. | Jan 2002 | B1 |
6356951 | Gentry, Jr. | Feb 2002 | B1 |
6560610 | Eatherton et al. | May 2003 | B1 |
6564211 | Andreev et al. | May 2003 | B1 |
6725326 | Patra et al. | Apr 2004 | B1 |
6728732 | Eatherton et al. | Apr 2004 | B1 |
6741985 | Green | May 2004 | B2 |
Number | Date | Country | |
---|---|---|---|
Parent | 10369354 | Feb 2003 | US |
Child | 10833532 | US | |
Parent | 09371907 | Aug 1999 | US |
Child | 10369354 | US |