Patent Application
20090228680
Data is very important to individuals and businesses. Many businesses regularly back up data stored on computer systems to avoid loss of data should a storage device or system fail or become damaged. One current data backup trend is to backup data to disks and use tapes for long term retention only. The amount of disk space needed to store a month's backup can be very large, such as around 70 terabytes in some examples. The amount of data will likely only be increasing going forward.
One strategy for backing up data involves trying to back up data that has changed, as opposed to all of the data, and then using prior backups of unchanged data to reconstruct the new backup. In one approach, data may be divided into fixed size blocks. An MD5 hash or a SHA256 hash may be calculated on the data belonging to the fixed size block of data, resulting in an MD5 signature for each block of data. The MD5 signature may be searched against an in memory database or an embedded database of previous MD5 signatures. In this approach any insertion of new data would cause blocks to shift and hence fixed size chunking and MD5 calculations on those fixed size chunks will not help.
The next time the file is backed up, signatures are generated for the blocks and searched against the database of signatures to find duplicates if any data has changed. Only the changed blocks need be saved during the backup.
A method comprises obtaining a plurality of variable size chunks of data, calculating a first hash for each variable size chunk of data, determining a length representation for each variable size chunk of data, determining a second hash of a subset of each variable size chunk of data, and forming a signature of each variable size chunk of data that is comprised of the first hash, the length representation and the second hash.
In a second embodiment, a method comprises obtaining a plurality of variable size chunks of data, calculating a first sequential hash for each variable size chunk of data, determining a length representation for each variable size chunk of data, determining a second Adler hash of a subset of each variable size chunk of data, and forming a signature of each variable size chunk of data that is comprised of the first hash, the length and the second hash.
In yet a further embodiment, a computer readable medium has instructions for causing a computer to execute a method, the method comprising obtaining a plurality of variable size chunks of data, calculating a first sequential hash for each variable size chunk of data, determining a length representation for each variable size chunk of data, determining a second Adler hash of a subset of each variable size chunk of data, and forming a signature of each variable size chunk of data that is comprised of the first hash, the length and the second hash.
In still a further embodiment, the device comprises means for obtaining a plurality of variable size chunks of data, means for calculating a first hash for each variable size chunk of data, means for determining a length representation for each variable size chunk of data; means for determining a second hash of a subset of each variable size chunk of data, and means for forming a signature of each variable size chunk of data that is comprised of the first hash, the length representation and the second hash.
In the following description, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific embodiments which may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention, and it is to be understood that other embodiments may be utilized and that structural, logical and electrical changes may be made without departing from the scope of the present invention. The following description of example embodiments is, therefore, not to be taken in a limited sense, and the scope of the present invention is defined by the appended claims.
The functions or algorithms described herein may be implemented in software or a combination of software and human implemented procedures in one embodiment. The software may consist of computer executable instructions stored on computer readable media such as memory or other type of storage devices. The term “computer readable media” is also used to represent any means by which the computer readable instructions may be received by the computer, such as by different forms of wired or wireless transmissions. Further, such functions correspond to modules, which are software, hardware, firmware or any combination thereof. Multiple functions may be performed in one or more modules as desired, and the embodiments described are merely examples. The software may be executed on a digital signal processor, ASIC, microprocessor, or other type of processor operating on a computer system, such as a personal computer, server or other computer system.
In some prior backup systems, a hash is calculated on fixed length blocks of data. If data in a block changes, it may result in further blocks also changing, and corresponding changes in hashes. As a result, it becomes difficult to determine if data following the first block changed, resulting in backing up more data than may be really needed.
A hashing mechanism may be used to find some “special boundaries” on data. A hash, such as an MD5 hash is calculated on the data between those special boundaries. Such data is referred to as a chunk of data. In one embodiment, a full hash may then be created which consists of the MD5 hash, the size of the data chunk between the boundaries, and a weak hash of data around the special boundary.
The hashes may be stored in a desired format of a hash database. The next time the backups are done, the special boundaries are calculated on the data, resulting in many of the exact same chunks. The MD5 hash is calculated on the data chunks between the special boundaries along with the full hash on data between the special boundaries, and the hashes are compared against the hash database. If the hashes match, the data corresponding to the hash does not need to be stored again. In various embodiments, this may lead to huge savings on disk storage.
In one embodiment of a method 100 of determining data boundaries is illustrated in
Consider a data set as consisting of a string of characters: “AAA BBB CCC DDD EEE FFF GGG HHH III JJJ KKK LLL MMM NNN OOO PPP QQQ RRR SSTTT UUU”. A character may be a desired amount of data, such as one or more bits of data that may be used to identify characters, including one or more bytes of data in some embodiments. A character as used herein need not be limited to an amount corresponding to the amount used to represent an alphanumeric character.
In one embodiment, a special boundary determining hash is 32 characters in length. One such special boundary determining hash used may be an Adler hash. Such hash is first calculated on the first 16 characters at 110, referred to as a subset, of the data set: “AAA BBB CCC DDD”
This calculated hash is then compared against a fixed boundary indicator value at 115. If the calculated hash equals boundary indicator value at 120, it is determined that a “natural boundary” in the data set has been reached as indicated at 125. If the calculated hash does NOT equal boundary indicator value, a next subset is found at 130 by shifting one character to the right in the data set. Control is returned to 110, where a new boundary determining hash of 32 characters is then calculated on the new subset: “AA BBB CCC DDD E”
The new boundary determining hash is then compared with the boundary indicator value at 115. The process of shifting the subset of data through the set of data and calculating and comparing the boundary determining hash with the boundary indicator value is repeated until the hash and value match. The point in the data set at which they match is the “natural boundary” for the chunk of data.
Assume for example that a natural boundary is found at the data subset of: “MM NNN OOO P”. The first “chunk” is then determined to be: “AAA BBB CCC DDD EEE FFF GGG HHH III JJJ KKK LLL MMM NNN OOO P”. This chunk is then fed to a module or device that performs an MD5 hash. An MD5 hash is calculated on the first chunk and stored in the hash database at 135.
The Adler hash is then calculated on the next subset of data, the 16 bytes starting with: “PP QQQ RRR SST”. If the last set has not been processed as indicated at 140, the process is repeated for all sets to determine all of the chunks in the data to be backed up. This result is chunks that are of variable size, wherein the size is of each chunk is dictated by the Adler hash algorithm. Each chunk has a corresponding MD5 hash in the hash database. These calculations and the process may be performed at any time, including during a backup of the data. Since the Adler hash is likely to identify the same data boundaries as before, even a change at the beginning of the data should result in many identical chunks and corresponding hashes being found. This may significantly reduce the amount of data that is written to disk during a backup.
In one embodiment, between backups of data, portions of the data may be changed. In one simple example, data may be inserted into the above example data set, resulting in a new data set: “ZZZ AAA BBB CCC DDD EEE FFF GGG HHH III JJJ KKK LLL MMM NNN OOO PPP QQQ RRR SSTTT UUU”. Note that the characters “ZZZ” has been added to the beginning of the data set. The Adler hash algorithm is applied on 32 byte subsets of data from the data set until the hash equals the boundary indicator value.
It is possible that a “natural boundary” may or may not be found prior to the previous natural boundary. Encountering the natural boundary earlier depends on the data. However, the same natural boundary will be found again at “MM NNN OOO P”, because this data has remained the same.
The MD5 hash is calculated from start of data set until the natural boundary “MM NNN OOO P”. This MD5 hash will not match with the previous one because of the change in data from previous one. However, the characters following “MM NNN OOO P” would remain the same, and the MD5 hash will match from this point. This ensures that similarities in data are found even if some parts of the file have changed.
In one embodiment, it is desired that the hashes are on chunks of data having a size between 8K and 32 K. Processing characters to find a “natural boundary” may be started after 8K of a file or set of data. In this embodiment, each subset is identified at 110 and the boundary identifying hash is performed on subsets starting after 8K Thus processing is avoided in the first 8K of each resulting chunk of data. The 8K size may be referred to as a “minBoundaryLength”. The minBoundaryLength may have different values in further embodiments, which may be dependent on the type of hash used and other desired efficiencies.
Processing continues as before with calculating a hash on 32 characters, checking the hash against a constant boundary indicator value. If it doesn't match, a window for the data subset is slid over by one character. Processing continues until the boundary indicator value is found.
If the boundary indicator value is not found by the time the sliding window defining the subset of data has reached 32K at 120, the corresponding 32K chunk is defined as the next chunk size, and an MD5 hash is calculated on the 32 K chunk. The 32 K size is called a “maxBoundaryLength”, which may vary in further embodiments as desired.
In one embodiment, a method 200 may be varied from method 100 to reduce the chances of hitting “maxBoundaryLength too often. At each position where a hash is calculated on a trailing 32 characters, the hash is compared with 2 different boundary indicating numbers at 220 and 230. A first boundary indicating number around 10K, and a second boundary indicating number around 20K.
Thus, the Adler hash is compared with two different boundary indicating values. A “natural boundary” may be hit at two different places by comparing against the two natural boundary values. Once the lower boundary position is hit, the algorithm continues processing characters hoping to hit a higher “natural boundary”. If the natural boundary corresponding to the divisor 20K is not found, the natural boundary corresponding to the lower 10K divisor is considered at 240. In a worst case where neither boundary indicator value is found for the lower or higher divisors, 32 K is considered as the chunk at 250. In one embodiment, natural boundaries are found up to approximately 95% of the time.
In yet a further embodiment, an efficient method 300 of calculating the Adler hash is illustrated in
If there are N characters, the Adler hash is performed on N characters. At each successive character the hash of the trailing 32 characters is calculated. This would be very costly. In one embodiment, the Adler hashing algorithm is a rolling algorithm.
For example, consider the characters “ABCDEFGH”. The hash is calculated as ‘A’+‘B’+ . . . +‘H’. When calculating the hash by shifting one character, all characters do not need to be added again. In one embodiment, ‘I’ is added to the previous sum. Then ‘A’ is subtracted at 330. This results in a new hash for the next subset and is referred to as a rolling hash algorithm. If a boundary does not result at 340, 320 and 330 are repeated. If a boundary is found at 340, a hash on the resulting chunk of data is calculated at 350. At 360, if no more data is found, the process ends at 370. Otherwise, a new subset is determined following the last natural boundary at 310 and the method continues.
Hash on Chunks
A hash, such an MD5 hash may not be guaranteed to generate a different hash for different chunks of data. The probability of MD5 clashes is very low. But our objective is to reduce the probability of clashes even further. In one embodiment, the hash on the chunk has additional data added to it. The length of the data chunk over which the hash is created is also be included with the overall hash. The last boundary identifying hash for the data chunk is also be included. If all three match, it is assumed that the data chunk is also matching.
In one embodiment, an MD5 algorithm is used to create the hash on a data chunk. An MD5 hashing algorithm may be used to generate a 128-bit hash of a data set of any length. The hash generated by this algorithm depends on the content of the data and the length of the data. The hash generated by the MD5 algorithm may be very sensitive to any change in the content of the data and exhibits a relatively large change in the hash output even if there is a very small change in the data content. An MD5 hash function processes the data set sequentially to calculate the hash output. The length of data set affects the generated hash output. Whenever processing variable length data sets, such as the data chunks herein, in addition to the hash of the data, the length of data set may also be stored. In case of data sets of different length, the hash comparison of the data set alone may not be meaningful.
In one embodiment, a combined signature 400 in
In one embodiment, signature 400 consists of not just the MD5 hash, but has the size of the Adler chunk, the Adler hash and the MD5 hash. This ensures that even if there is an MD5 clash, the chances of us getting affected because of MD5 clash is very low because not just the MD5 has is being relied upon. Instead, the size of the Adler chunk and an Adler signature are being used, which themselves may dictate the size of the chunk that is used to calculate the MD5.
In one embodiment, the Adler hash of 32 bits for a selected subset of data, such as the last subset before a boundary was determined, may be used as the checksum function. The Adler 32 bit hash is sensitive to the contents of data chunk and relative positioning of bytes in the data set. The checksum may be computed with a very high speed making it very efficient for this purpose. Since the probability of finding two data chunks of the same length with matching MD5 hash and Adler-32 checksum is infinitesimally small, the possibility of collision in signatures is highly unlikely.
By combining all the three attributes (length, checksum, and hash) of a data chunk in a single signature, signatures of the data chunk may be determined and compared in a single compare operation at a very high speed. By using a combination signature as above, the possibility of collision vastly reduced if not eliminated without any significant impact on processing speed.
A block diagram of a computer system that executes programming for performing the above methods is shown in
Computer-readable instructions stored on a computer-readable medium are executable by the processing unit 502 of the computer 510. A hard drive, CD-ROM, and RAM are some examples of articles including a computer-readable medium.
The Abstract is provided to comply with 37 C.F.R. § 1.72(b) to allow the reader to quickly ascertain the nature and gist of the technical disclosure. The Abstract is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims.
| Number | Date | Country | |
|---|---|---|---|
| 61033920 | Mar 2008 | US |
