System and method for efficient filtering of data set addresses in a web crawler

[0001] The present invention relates to a system and method for filtering document addresses in a web crawler and, more particularly, to a method for efficiently filtering the addresses of downloaded documents to avoid downloading from duplicate addresses.

[0002] Documents on interconnected computer networks are typically stored on numerous host computers that are connected over the networks. For example, so-called “web pages” may be stored on the global computer network known as the Internet, which includes the world wide web. Web pages can also be stored on Intranets, which are typically private networks maintained by corporations, government entities, and other groups. Each web page, whether on the world wide web or an Intranet, has a distinct address called its uniform resource locator (URL), which at least in part identifies the location or host computer of the web page. Many of the documents on Intranets and the world wide web are written in standard document description languages (e.g., HTML, XML). Theses languages allow an author of a document to create hypertext links to other documents. Hypertext links allow a reader of a web page to access other web pages by clicking on links to the other pages. These links are typically highlighted in the original web page. A web page containing hypertext links to other web pages generally refers to those pages by their URL's. A URL may be referred to more generally as a data set address, which corresponds to a web page, or data set. Links in a web page may refer to web pages that are stored in the same or different host computers.

[0003] A web crawler is a program that automatically finds and downloads documents from host computers in an Intranet or the world wide web. A computer with a web crawler installed on it may also be referred to as a web crawler. When a web crawler is given a set of starting URL's, the web crawler downloads the corresponding documents. The web crawler then extracts any URL's contained in those downloaded documents. Before the web crawler downloads the documents associated with the newly discovered URL's, the web crawler needs to find out whether these documents have already been downloaded. If the documents associated with the newly discovered URL's have not been downloaded, the web crawler downloads the documents and extracts any URL's contained in them. This process repeats indefinitely or until a predetermined stop condition occurs.

[0004] Typically, to find out whether the documents associated with a set of discovered URL's have already been downloaded or are scheduled to be downloaded, the web crawler checks a directory of document addresses. These document addresses are URL's that correspond to documents which have either already been downloaded or are scheduled to be downloaded; for convenience, these documents will be referred to as downloaded documents. The directory stores the URL's of the downloaded documents, or representations of the URL's. The set of URL's in downloaded documents could potentially contain addresses of every document on the world wide web. As of 1999 there were approximately 800 million web pages on the world wide web and the number is continuously growing. Even Intranets can store millions of web pages. Thus, web crawlers need efficient data structures to keep track of downloaded documents and any discovered addresses of documents to be downloaded. Such data structures are needed to facilitate fast data checking and to avoid downloading a document multiple times.

[0005] Typically, the set of downloaded document addresses is stored in disk storage, which has relatively slow access time. One example of a method designed to facilitate fast data checking and to avoid downloading a document multiple times is disclosed in U.S. patent application Ser. No. 09/433,008, filed Nov. 2, 1999. That document discloses storing address representations on disk, and using an efficient address representation to facilitate fast look-up of document addresses stored on disk. The present invention provides improved storage methods, decreasing the frequency with which disk storage must be accessed.

[0006] The present invention provides efficient address filtering operations for a web crawler by storing representations of recently-identified URL'S, typically a checksum or “fingerprint,” in a buffer in main memory, rather than on disk. A representation of a URL is added to the buffer only when it is not already stored in the buffer. The buffer is merged with a disk file only when the buffer is full. Thus, the disk is accessed only when the buffer is full, and not every time a URL is identified by the web crawler while processing a downloaded document.

[0007] During the merge, items in the buffer not found in the disk file are identified and the corresponding URL's are queued for downloading by the web crawler.

[0008] The present application is applicable to both Internet and Intranet web crawlers.

[0009] In one embodiment, a cache for storing “popular” URL's is maintained in memory in addition to the buffer. Each identified URL is compared with entries in the cache and entries in the buffer, and is added to the buffer only when it is not in either the cache or buffer. Use of the cache reduces the rate at which the buffer is filled with fingerprints, which reduces the frequency of merges with the disk file.

[0010] In some embodiments the disk file is a sparse file, having empty entries interspersed among the non-empty entries, using open-addressing. Thus, during a merge, the number of writes to the disk is one for each representation in the buffer that is not already present in the disk file, which increases the efficiency of the merge operation. In some embodiments, the disk file is divided into sparse sub-files, which further increases the efficiency of the merge operation. Efficient address representations may be used, increasing the efficiency of the merge operation, particularly in the case where URLs on a given web page have a high degree of locality.

[0011] Additional features of various embodiments of the invention are described in the detailed description of the invention provided below.

[0023] FIG. 1 shows an exemplary embodiment of a distributed computer system 100. The distributed computer system 100 includes a web crawler 115 connected to a network 120 through a network interconnection 116. The network 120 may be a global communication network, such as the Internet, or a private network, sometimes called an Intranet. Examples of network interconnection 116 includes switches, routers, etc.

[0024] The Internet network 120 includes web servers 121 and a service known as a domain name system 122. It may also optionally include a web page indexing system 123. The web servers 121 store web pages. The domain name system 122 is a distributed database that provides the mapping between Internet Protocol (IP) addresses and host names. The domain name system 122 is a distributed system because no single site on the Internet has the domain name mapping information for all the web servers in the network. Each site participating in the domain name system 122 maintains its own database of information and runs a server program that other systems across the Intranet or Internet can query. The domain name system provides the protocol that allows clients and servers to communicate with each other. Any application may look up the IP address (or addresses) corresponding to a given host name or the host name corresponding to a given IP address in the domain name system 122. An application accesses the domain name system 122 through a resolver. The resolver contacts one or more name servers to perform a mapping of a host name to the corresponding IP address, or vice versa. A given host name may be associated with more than one IP address because an Intranet or Internet host may have multiple interfaces, with each interface of the host having a unique IP address.

[0025] The web page indexing system 123 includes an index of words used on the world wide web and addresses of the web pages that use each word. Such indexing systems are maintained by various search engines, such as the AltaVista search engine. The domain name system 122 and the web page indexing system 123 may be accessed by the web crawler 115 in the process of downloading web pages from the world wide web.

[0026] The web crawler 115 includes a communications interface, or network connection, 102, one or more CPUs 101, an operator interface 103 (which may be remotely located on another computer), primary or main memory 104 and secondary (e.g. disk) memory 112. In an exemplary embodiment, the network connection 102 is able to handle overlapping communication requests. The memory 104 includes:

• • [0027] a multitasking operating system 105;
  • [0028] an Intranet/Internet access procedure 106 for fetching web pages as well as communicating with the domain name system 122;
  • [0029] a fixed size buffer B 107 for storing recently added numerical representations of URL's, and optionally the URL's themselves;
  • [0030] an optional second fixed size buffer B′ (not shown) for storing recently-added numerical representations of URL's, and optionally the URL's as well, when buffer B 107 is being merged with disk file 113;
  • [0031] an optional cache C (not shown) for storing “popular” URL's;
  • [0032] threads 108 for downloading web pages from the servers 121, and processing the downloaded web pages;
  • [0033] a main web crawler procedure 109 executed by each of the threads 108; and
  • [0034] a URL address filtering procedure 110 executed by each of the threads 108 to filter the URL's identified in a downloaded web page, and determine which URL's are new and therefore should be scheduled for downloading.

[0035] Buffer B 107, optional buffer B′ and optional cache C are preferably implemented in main memory as hash tables to facilitate fast lookup operations.

[0036] A fingerprint disk file 113, which is not bounded in size, is stored in secondary disk storage 112. A second URL disk file, not shown, may also be stored in disk storage 112. The buffers B and B′, cache C and the fingerprint disk file 113 store a numerical representation, or “fingerprint,” of each URL known to the web crawler. The fingerprint of a URL may be generated using an appropriate function, such as a function that employs one or more checksum functions. The buffers may additionally store the actual address, or URL, corresponding to each stored fingerprint. Alternatively, in a preferred embodiment, the addresses corresponding to the fingerprints in a buffer may be stored in the optional URL disk file instead of in a buffer.

[0037] In the exemplary embodiment the web crawler uses multiple threads to download and process documents. The web crawler 115 is given a set of initial URL's and begins downloading documents using those URL's. Various data structures may be used to keep track of which documents (web pages) the threads should download and process, but those particular data structures are not the subject of the present document. Rather, the present invention concerns the data structures used to keep track of the URL's of documents that have been downloaded or may be scheduled for downloading and to ensure that duplicate URL's are not scheduled for downloading.

[0038] Each thread executes a main web crawler procedure 109, which will be now described with reference to the diagram shown in FIG. 2 and the flow chart shown in FIG. 3. The web crawler thread determines the data set address, URL-i, of the next document, or data set, to be downloaded, typically by retrieving it from a queue data structure 140 (step 150). The thread then downloads the document 141 corresponding to the URL (step 151), and processes the document. The processing may include indexing the words in the document so as to make the document accessible via a search engine. However, the only processing of the document that is relevant to the present discussion is that the main procedure identifies URL's in the downloaded document 141 (step 152) that are candidates for downloading and processing. Typically, these URL's are found in hypertext links in the document 141 being processed.

[0039] Each identified URL is passed to the address filtering procedure in step 153 to determine if it is the URL for a page already known to the web crawler, or is a “new URL” for a document not previously known to the web crawler. If that procedure determines that the identified URL is already known to the web crawler, the URL is discarded and is not scheduled for downloading. If the identified URL is not known, it is ultimately added to the queue 140 of URL's scheduled for downloading. In the meantime, the main procedure continues to identify URL's and pass URL's to the address filtering procedure until all URL's from the downloaded page have been processed (steps 154–155). When all URL's have been passed to the address filtering procedure, the next address in the queue 140 is downloaded (step 150).

[0040] An exemplary embodiment of the address filtering procedure (also herein called the address filter) invoked by the main web crawler procedure is described next with reference to the diagram shown in FIG. 2 and the flow chart shown in FIGS. 4 and 5. The first step of the URL processing procedure is to convert a specified URL into a fixed length numeric representation, or “fingerprint,” (step 160), by applying a predefined checksum function (for example, without limitation, a checksum fingerprint function), or set of checksum functions, to the URL. The fingerprint disk file 113 stores fingerprints of URL's instead of storing the corresponding URL's. The buffer B 107 also stores fingerprints, and additionally stores the corresponding URL's as well. Many suitable fingerprint functions are known to those skilled in the art, and thus are not described here.

[0041] After the specified URL U has been converted into a numeric representation N, a lookup is performed to see if N is already stored in buffer B 107 (step 161), which stores the fingerprints of recently added URL's, as well as the corresponding URL's. If N is already stored in buffer B 107, it has already been scheduled for downloading, and therefore no further processing of the URL U is needed. If N is not stored in buffer B 107, it may be new (although it may already be stored in the fingerprint disk file 113), and therefore N and the corresponding URL U are added to buffer B 107 (step 163). The address filter continues to process URL's (step 167) until buffer B 107 is full.

[0042] If the buffer B 107 is full (step 164), it must be merged with the fingerprint disk file 113 (step 166). The buffer B 107 is first sorted (step 165) so that the fingerprints are in order. After ordering, the buffer B 107 is then merged with the fingerprint disk file 113, where only fingerprints corresponding to new URL's are added to the disk file 113. As shown in FIG. 5, during the merge, the fingerprint disk file 113 is searched to determine whether a fingerprint N_k is located in the disk file (step 171). If so, fingerprint N_k is ignored. If not, the fingerprint is added to the disk file 113, and the corresponding URL is added to the queue for downloading of the corresponding web page document in step 174. If there are more fingerprints in buffer B (step 175), the next fingerprint in buffer B 107 is processed (step 176). When the merge process is completed, the contents of buffer B 107 are deleted (step 177). The process of merging buffer B 107 into the disk file 113 is a relatively time-consuming operation. Therefore buffer B 107 is typically made fairly large so as to minimize the frequency of such merge operations.

[0043] During the merge process, which is an ordered merge, fingerprint N_k must be inserted in the fingerprint disk file 113 in the proper location, as illustrated in FIG. 6, so that the disk file 113 remains ordered. This requires the disk file to be completely re-written. To avoid this lengthy rewrite process, in a preferred embodiment, the fingerprint disk file may be sparsely-filled, using open addressing. For this embodiment, the fingerprint disk file represents a hash table, with a substantial proportion of the table, for example 50% or 75%, being empty entries, or “holes.”

[0044] In this embodiment, in order to determine whether a particular fingerprint N_k is in the disk file, the hash of the fingerprint is computed. In one embodiment, only a prefix of the fingerprint is used for the hash value. The hash value is the starting position for searching through the fingerprint disk file. The disk file is searched sequentially, starting at the starting position, for either a match or a hole. If a hole is found, the fingerprint N_k is stored in that hole; if a match is found, N_k is discarded. Thus, there is only one write to the disk file for each fingerprint not already present in the disk file, and the size of the disk file is not a factor in the merge time. When the disk file becomes too full—for example, when only 25% of the slots in the disk file are holes—the file must be completely rewritten into a new, larger file. For example, the new file may be doubled in size, in which case the amortized cost of maintaining the file is constant per fingerprint in the hash table. It will be appreciated that the use of open addressing a sparsely-filled disk file drastically reduces the disk re-writing required during a merge.

[0045] In one embodiment, the disk file may be divided into sparse sub-files, with open-addressing used for each sub-file. An index may be used to identify the range of fingerprint hash values located in each sub-file, or an additional hash table may be used to map fingerprints to the various sub-files. When a sub-file becomes too full, it may be re-written into a new, larger file, but the entire disk file need not be re-written.

[0046] In another aspect of the present invention, an efficient addressing scheme may be used for either a sparse disk file, or a disk file consisting of a set of sparse sub-files. In this addressing scheme, discussed in U.S. patent application Ser. No. 09/433,008, filed Nov. 2, 1999 (hereby incorporated by reference in its entirety), each fingerprint is composed of two components: a first fingerprint component based only upon the host component of the URL, and a second fingerprint component based upon the entire URL. A canonical representation of the host component of each URL may be used instead of the actual host component. Each fingerprint component is generated using a fingerprint function, which may differ for the two components. The fingerprint for a URL is the concatenation of these two components.

[0047] The disk file (or each sub-file) is ordered, so that all fingerprints from the same host will be likely stored in either the same sub-file, or at contiguous memory locations within the disk file. Because URL's located on a given web page often have the same host, this efficient addressing scheme can significantly improve the speed of the merge process. When a new fingerprint is to be added to the disk file during the merge, often the buffer used for the merge will already contain the disk page on which that fingerprint will be stored. This will occur when the previous fingerprint added to the disk file was from the same host as the new fingerprint, and consequently would likely have been stored on the same disk page as the new fingerprint.

[0048] In an alternative embodiment, illustrated in FIG. 7, in addition to buffer B 107, a cache C 117 of “popular” URL's is included. When a URL is removed from the queue 140 and the corresponding web page 141 is downloaded, the fingerprint for each URL extracted from the web page 141 is first compared to fingerprints in cache C 117, to see whether the URL is a “popular” URL that has already been downloaded and therefore need not be processed any further. If the fingerprint is not in cache C 117, the address filter then goes on to compare the fingerprint to the fingerprints in buffer B 107, as discussed previously. Use of the cache C 117 significantly reduces the rate at which the buffer B 107 is filled with fingerprints, which reduces the frequency of merges with the disk file and thus improves the efficiency of the address filter.

[0049] The cache C is updated according to an update policy. When, according to the update policy, a new fingerprint must be added to cache C, and cache C 117 is full, a fingerprint is evicted from cache C 117. The fingerprint to be evicted at any one time is determined using an appropriate eviction policy, such as a least recently used eviction policy a round robin eviction policy or a clock replacement eviction policy.

[0050] A new fingerprint is added to cache C 117 according to one of the following update policies. A fingerprint corresponding to a URL may be added to cache C 117, following the comparison of the fingerprint to the contents of cache C 117, whenever that fingerprint is not already contained in cache C 117. Thus, under this update policy, cache C 117 is simply a cache of fingerprints for recently-identified URL's. Alternatively, a fingerprint may be added to cache C 117 following the comparison of a fingerprint to the contents of buffer B 107 whenever that fingerprint is not contained in cache C 117, but is contained in buffer B 107. In this case, cache C 117 is a cache of fingerprints for URL's which have been identified more than once recently. Or, a fingerprint may be added to cache C whenever that fingerprint is not contained in cache C 117 and is not contained in buffer B 107. Cache C 117 in this case is a cache of fingerprints for URL's which have been identified only once recently. Finally, a fingerprint may be added to cache C 117 during the merge when the fingerprint is found to have previously existed in the fingerprint disk file 113. For this policy, cache C 117 is a cache of fingerprints for URL's which have been identified more than once during the entire web crawl.

[0051] Experiments conducted using these various update policies indicate that the policies yield comparable results, with the last-mentioned policy providing marginally better results than the other three. Each policy has a corresponding “miss rate”—a fraction of fingerprints compared against buffer B 107 following the comparison with cache C 117—of close to 30%. That is, about 70% of all the fingerprints that would have been compared against the buffer B 107 were instead discarded after a comparison with cache C 117 indicated that the fingerprints were old. Obviously, this reduction in fingerprints compared against, and potentially stored in, buffer B 107 reduces the frequency with which buffer B 107 must be merged 142 with the fingerprint disk file 113, and thus increases the efficiency of the address filter.

[0052] Another alternative embodiment is illustrated in FIG. 8. When buffer B 107 is full, it must be merged 142 with the fingerprint disk file 113, and cannot accept any newly-identified URL's. Hence, no new address cache accepted for filtering while the time-consuming merge process is in progress. To avoid delaying the address filtering procedure during the merge process, the address filter uses a second buffer B′ 180 in addition to the buffer B 107. While buffer B 107 is being merged with the disk file 113, buffer B′ 180 takes over the role of buffer B 107. Fingerprints for newly-identified URL's are compared with the contents of buffer B′ 180 and then stored in buffer B′ 180 if it is determined that the fingerprints are not already stored in buffer B′ 180. When buffer B′ 180 is full, the two buffers change roles again.

[0053] The cache C 117 of the embodiment in FIG. 7 can also be used in the embodiment shown in FIG. 8. The addition of cache C makes use of the buffer B 107/180 more efficient, while the use of two buffer B's 107, 180 greatly reduces the “cost” of the buffer merge operations in terms of the web crawler's ability to process discovered URL's that may or may not be new URL'S.

[0054] Yet another alternative embodiment is illustrated in FIG. 9, with the process performed by this embodiment shown in FIG. 10. In this embodiment, instead of storing URL's in buffer B 190, along with their corresponding fingerprints, the URL's are stored in a separate URL disk file 192. In this embodiment, each fingerprint in buffer B 190 includes a pointer to the corresponding URL in the URL disk file 192. The pointer may, for instance, indicate the relative character position of the start of the URL from the start of the URL disk file 192. Alternatively, the pointer may simply indicate the count of the URL in the disk file—the first URL gets pointer 1, the second pointer 2 and so on. In addition, each fingerprint in buffer B 190 will be associated with a flag D.

[0055] When the fingerprints are sorted (step 200, FIG. 10) by fingerprint value, at the beginning of the merge 191 (FIG. 9), the pointers must remain associated with their corresponding fingerprints. Pointers and flags may be stored, along with their associated fingerprints, in a table 190a, to facilitate preserving this association. In other words, during the buffer sort operation (step 200), each row of the buffer 190 (i.e., table 190a) is moved as a whole within the buffer to its new position.

[0056] As in other embodiments described above, for each fingerprint, the merge process 191 determines whether the fingerprint N_k is in the fingerprint disk file 113 (step 201). If so, the corresponding URL U_k is old, and the fingerprint and URL are discarded, and the flag D_k is set to a first predefined value (e.g., “0”) to indicate that the fingerprint is already located in the fingerprint disk file (step 202). If the fingerprint N_k is not in the fingerprint disk file 113, it is new, is added to the fingerprint disk file 113 in step 204, and the flag D_k is set to a second predefined value (e.g., “1”) to indicate that the URL U_k needs to be added to the list of URL in the queue 140 (FIG. 9).

[0057] After the merge process 191 is complete, buffer B 190 is sorted a second time (step 210), using the pointers as the sort key, and sorting the pointers in, for example, increasing order. Again, each row, consisting of a fingerprint, a pointer, and a flag, is moved as a whole during the sort. After the sort, URL's from the URL disk file are added to the queue. More specifically, for every fingerprint FP_k in the buffer B whose flag D_k is set to the second predefined value (“1”), the URL is added to the queue for downloading (step 212). If flag D_k is not set to the second predefined value, the corresponding URL is not added to the queue. Because the entries in buffer B are sorted in the same order as URL's are stored in the URL disk file, the URL disk file is read sequentially from beginning to end during step 212, which is much more efficient than reading the URL's in the URL disk file in random order. After step 212 is completed, and thus all new URL's in the URL disk file have been added to the queue, the URL disk file is either discarded, in which case a new URL disk file is used for storing the URL's while the buffer is filled again with fingerprints of URL's, or the URL disk file 192 may be cleared for reuse (step 214). In addition, the contents of buffer B are deleted to make buffer B ready for reuse (step 214).

[0058] Another aspect of the present invention is illustrated in FIG. 11, where the web crawler has just been initialized and no fingerprints have been stored in the fingerprint disk file. Since there are no “old” URL's, the address filter procedure simply bypasses the merge operation 142 to save processing time. The buffer B 107, which is implemented as a hash table, is searched to see if a newly-identified URL is in buffer B 107. If not, the URL is loaded directly into the queue 140, and the corresponding fingerprint is stored in buffer B 107. When buffer B 107 fills up, all of its fingerprints are written to the fingerprint disk file 113. After the first buffer B of fingerprints has been written to the fingerprint disk file, normal operation of the web crawler, as described above, begins.

[0059] The present invention can be implemented as a computer program product that includes a computer program mechanism embedded in a computer readable storage medium. For instance, the computer program product could contain the program modules shown in FIG. 1. These program modules may be stored on a CD-ROM, magnetic disk storage product, or any other computer readable data or program storage product. The software modules in the computer program product may also be distributed electronically, via the Internet or otherwise, by transmission of a computer data signal (in which the software modules are embedded) on a carrier wave.

[0060] The foregoing examples illustrate certain exemplary embodiments of the invention from which other embodiments, variations and modifications will be apparent to those skilled in the art. The invention should therefore not be limited to the particular exemplary embodiments discussed herein, but rather defined by the claims appended hereto.