Acceleration of Join Operations by a Relational Database Processor, RINDA

Fast join methods implemented in a relational database processor, RINDA, are described. RINDA performs complex queries including sorts and joins with specialized hardware. Join operations by RINDA are executed in three phases: filtering phase, sorting phase and merge-join phase. In the filtering phase, unjoinable tuples are removed with hashed-bit-arrays. Remaining tuples are sorted in the sorting phase. Sorted tuples are merged and connected together in the merge-join phase. Iterating operations in the filtering and sorting phases are rapidly executed by RINDA’s specialized hardware. Especially in the filtering phase, a new multiplication-folding method is used as a hashing function to set and refer hashed-bitarrays. It strongly reduces collisions for any type and length of keys. Three kinds of join algorithms, nestedloop, single-table filtering and dual-table filtering algorithms, are dynamically selected according to the number of tuples to be joined. Performance evaluation shows RINDA accelerates join operations about ten times compared with conventional software systems.