On Tuesday 13 May 2008, Dan Pascu wrote: > I have to agree with Juha here. While you may know that I'm not a fanatic > about performance and most of the time I would favor more functionality > or consistency and simplified code even if they incur a small performance > penalty, you may also know that I always advocated that performance > optimizations should be made where they really matter, not where they > make little to no difference. And I think we can all agree that the > database is one of the major bottlenecks in openser. Which means that > making it even slower than it is, doesn't make for a good move. > > I do not have any benchmarks here, but I would expect that userspace > functions to be slower. Especially if they are interpreted.
Hi Dan, i don't think that the difference is that big, that was the reason that i've applied Iouries patch to usrloc in the first place. But i was not aware that the function gets executed every second. > > why is the emulation function defined in mysql and not in oracle? i > > would prefer it to be defined in oracle, because so far majority of > > openser users is using mysql. > > Unfortunately this is not that simple. The BIT_AND function you mentioned > from mysql is a group (aggregate) function. You cannot use it to make a > bitwise and between 2 values, only between values in different rows for > the same column. It is used in conjunction with GROUP BY to make a > bitwise and over the values of a column in multiple rows. > > MySQL however, has an operator for this, which is like in C: & > But Oracle doesn't have the operator and it can't define one. > > However given that mysql and postgres are the most used databases, we > shouldn't cripple their abilities just to support other less used > databases. Maybe a different approach is needed here. Like have the > database backends provide functions that would encode certain operations > in the appropriate way for each database. Or do the bitwise operations in > C instead of the database (if possible). For postgres in the stable branch a similar approach is used. The emulation function (basically a stored procedure) defined for mysql is just a small wrapper around the '&' operator. Using a user defined procedure could be faster, but its not that easy to deploy them, as its necessary to implement them as object files [1]. I found a nice benchmark around the different function implementation, stored procedures are about 50% slower then using an internal operator (see 'adition benchmark', simplifing somewhat..) [2]. But even then its possible to execute over 50k of them every second. The fastest solution would be to implement this bitand in C in the usrloc. But i think that the overhead we get for this case is managable. Cheers, Henning [1] http://dev.mysql.com/doc/refman/5.0/en/adding-functions.html [2] http://rpbouman.blogspot.com/2008/03/udfs-at-mysql-users-conference.html _______________________________________________ Devel mailing list Devel@lists.openser.org http://lists.openser.org/cgi-bin/mailman/listinfo/devel
