sdbm - Substitute DBM or Berkeley ndbm for Every UN*X[1] Made Simple Ozan (oz) Yigit The Guild of PD Software Toolmakers Toronto - Canada oz@nexus.yorku.ca Implementation is the sincerest form of flattery. - L. Peter Deutsch A The Clone of the ndbm library The sources accompanying this notice - sdbm - consti- tute the first public release (Dec. 1990) of a complete clone of the Berkeley UN*X ndbm library. The sdbm library is meant to clone the proven functionality of ndbm as closely as possible, including a few improvements. It is practical, easy to understand, and compatible. The sdbm library is not derived from any licensed, proprietary or copyrighted software. The sdbm implementation is based on a 1978 algorithm [Lar78] by P.-A. (Paul) Larson known as "Dynamic Hashing". In the course of searching for a substitute for ndbm, I pro- totyped three different external-hashing algorithms [Lar78, Fag79, Lit80] and ultimately chose Larson's algorithm as a basis of the sdbm implementation. The Bell Labs dbm (and therefore ndbm) is based on an algorithm invented by Ken Thompson, [Tho90, Tor87] and predates Larson's work. The sdbm programming interface is totally compatible with ndbm and includes a slight improvement in database ini- tialization. It is also expected to be binary-compatible under most UN*X versions that support the ndbm library. The sdbm implementation shares the shortcomings of the ndbm library, as a side effect of various simplifications to the original Larson algorithm. It does produce holes in the page file as it writes pages past the end of file. (Larson's paper include a clever solution to this problem that is a result of using the hash value directly as a block address.) On the other hand, extensive tests seem to indicate that sdbm creates fewer holes in general, and the resulting page- files are smaller. The sdbm implementation is also faster than ndbm in database creation. Unlike the ndbm, the sdbm _________________________ [1] UN*X is not a trademark of any (dis)organization. - 2 - store operation will not "wander away" trying to split its data pages to insert a datum that cannot (due to elaborate worst-case situations) be inserted. (It will fail after a pre-defined number of attempts.) Important Compatibility Warning The sdbm and ndbm libraries cannot share databases: one cannot read the (dir/pag) database created by the other. This is due to the differences between the ndbm and sdbm algorithms[2], and the hash functions used. It is easy to convert between the dbm/ndbm databases and sdbm by ignoring the index completely: see dbd, dbu etc. Notice of Intellectual Property The entire sdbm library package, as authored by me, Ozan S. Yigit, is hereby placed in the public domain. As such, the author is not responsible for the consequences of use of this software, no matter how awful, even if they arise from defects in it. There is no expressed or implied warranty for the sdbm library. Since the sdbm library package is in the public domain, this original release or any additional public-domain releases of the modified original cannot possibly (by defin- ition) be withheld from you. Also by definition, You (singu- lar) have all the rights to this code (including the right to sell without permission, the right to hoard[3] and the right to do other icky things as you see fit) but those rights are also granted to everyone else. Please note that all previous distributions of this software contained a copyright (which is now dropped) to protect its origins and its current public domain status against any possible claims and/or challenges. Acknowledgments Many people have been very helpful and supportive. A partial list would necessarily include Rayan Zacherissen (who contributed the man page, and also hacked a MMAP _________________________ [2] Torek's discussion [Tor87] indicates that dbm/ndbm implementations use the hash value to traverse the radix trie differently than sdbm and as a result, the page indexes are generated in different order. For more information, send e-mail to the author. [3] You cannot really hoard something that is avail- able to the public at large, but try if it makes you feel any better. - 3 - version of sdbm), Arnold Robbins, Chris Lewis, Bill David- sen, Henry Spencer, Geoff Collyer, Rich Salz (who got me started in the first place), Johannes Ruschein (who did the minix port) and David Tilbrook. I thank you all. Distribution Manifest and Notes This distribution of sdbm includes (at least) the following: CHANGES change log README this file. biblio a small bibliography on external hashing dba.c a crude (n/s)dbm page file analyzer dbd.c a crude (n/s)dbm page file dumper (for conversion) dbe.1 man page for dbe.c dbe.c Janick's database editor dbm.c a dbm library emulation wrapper for ndbm/sdbm dbm.h header file for the above dbu.c a crude db management utility hash.c hashing function makefile guess. pair.c page-level routines (posted earlier) pair.h header file for the above readme.ms troff source for the README file sdbm.3 man page sdbm.c the real thing sdbm.h header file for the above tune.h place for tuning & portability thingies util.c miscellaneous dbu is a simple database manipulation program[4] that tries to look like Bell Labs' cbt utility. It is currently incomplete in functionality. I use dbu to test out the rou- tines: it takes (from stdin) tab separated key/value pairs for commands like build or insert or takes keys for commands like delete or look. dbu dbmfile dba is a crude analyzer of dbm/sdbm/ndbm page files. It scans the entire page file, reporting page level statistics, and totals at the end. dbd is a crude dump program for dbm/ndbm/sdbm data- bases. It ignores the bitmap, and dumps the data pages in sequence. It can be used to create input for the dbu util- ity. Note that dbd will skip any NULLs in the key and data fields, thus is unsuitable to convert some peculiar _________________________ [4] The dbd, dba, dbu utilities are quick hacks and are not fit for production use. They were developed late one night, just to test out sdbm, and convert some databases. - 4 - databases that insist in including the terminating null. I have also included a copy of the dbe (ndbm DataBase Editor) by Janick Bergeron [janick@bnr.ca] for your pleas- ure. You may find it more useful than the little dbu util- ity. dbm.[ch] is a dbm library emulation on top of ndbm (and hence suitable for sdbm). Written by Robert Elz. The sdbm library has been around in beta test for quite a long time, and from whatever little feedback I received (maybe no news is good news), I believe it has been func- tioning without any significant problems. I would, of course, appreciate all fixes and/or improvements. Portabil- ity enhancements would especially be useful. Implementation Issues Hash functions: The algorithm behind sdbm implementa- tion needs a good bit-scrambling hash function to be effec- tive. I ran into a set of constants for a simple hash func- tion that seem to help sdbm perform better than ndbm for various inputs: /* * polynomial conversion ignoring overflows * 65599 nice. 65587 even better. */ long dbm_hash(char *str, int len) { register unsigned long n = 0; while (len--) n = n * 65599 + *str++; return n; } There may be better hash functions for the purposes of dynamic hashing. Try your favorite, and check the pagefile. If it contains too many pages with too many holes, (in rela- tion to this one for example) or if sdbm simply stops work- ing (fails after SPLTMAX attempts to split) when you feed your NEWS history file to it, you probably do not have a good hashing function. If you do better (for different types of input), I would like to know about the function you use. Block sizes: It seems (from various tests on a few machines) that a page file block size PBLKSIZ of 1024 is by far the best for performance, but this also happens to limit the size of a key/value pair. Depending on your needs, you may wish to increase the page size, and also adjust PAIRMAX (the maximum size of a key/value pair allowed: should always - 5 - be at least three words smaller than PBLKSIZ.) accordingly. The system-wide version of the library should probably be configured with 1024 (distribution default), as this appears to be sufficient for most common uses of sdbm. Portability This package has been tested in many different UN*Xes even including minix, and appears to be reasonably portable. This does not mean it will port easily to non-UN*X systems. Notes and Miscellaneous The sdbm is not a very complicated package, at least not after you familiarize yourself with the literature on external hashing. There are other interesting algorithms in existence that ensure (approximately) single-read access to a data value associated with any key. These are directory- less schemes such as linear hashing [Lit80] (+ Larson varia- tions), spiral storage [Mar79] or directory schemes such as extensible hashing [Fag79] by Fagin et al. I do hope these sources provide a reasonable playground for experimentation with other algorithms. See the June 1988 issue of ACM Com- puting Surveys [Enb88] for an excellent overview of the field. References [Lar78] P.-A. Larson, "Dynamic Hashing", BIT, vol. 18, pp. 184-201, 1978. [Tho90] Ken Thompson, private communication, Nov. 1990 [Lit80] W. Litwin, "Linear Hashing: A new tool for file and table addressing", Proceedings of the 6th Conference on Very Large Dabatases (Montreal), pp. 212-223, Very Large Database Foundation, Saratoga, Calif., 1980. [Fag79] R. Fagin, J. Nievergelt, N. Pippinger, and H. R. Strong, "Extendible Hashing - A Fast Access Method for Dynamic Files", ACM Trans. Database Syst., vol. 4, no.3, pp. 315-344, Sept. 1979. [Wal84] Rich Wales, "Discussion of 'dbm' data base system", USENET newsgroup unix.wizards, Jan. 1984. [Tor87] Chris Torek, "Re: dbm.a and ndbm.a archives", - 6 - USENET newsgroup comp.unix, 1987. [Mar79] G. N. Martin, "Spiral Storage: Incrementally Augment- able Hash Addressed Storage", Technical Report #27, University of Varwick, Coventry, U.K., 1979. [Enb88] R. J. Enbody and H. C. Du, "Dynamic Hashing Schemes",ACM Computing Surveys, vol. 20, no. 2, pp. 85-113, June 1988.