| 摘要 |
A relational database system has a non-volatile memory, a volatile memory for temporarily storing a set of data records, and a query processor. The volatile memory has an amount of available space for query processing that is segmented into multiple memory pages. Initially, these memory pages are empty and available in a pool for use by the query processor. The query processor establishes a partition table that defines multiple partitions. The query processor partitions incoming data records into the partitions according to a hashing function and stores the data records in memory pages associated with the partitions. As a new data record placed into a particular partition, the query processor attempts to aggregate the new data record with any like data record that already exists in the particular partition. If no like data record exists, the data record is stored separately on the memory page within the partition. In the event that a memory page of the partition becomes filled, the query processor retrieves an empty memory page from the free pool and assigns that empty memory page to the needy partition. In the event that no free memory pages are left in the pool, the query processor selects a memory page from any one of the partitions and writes the data records on the selected memory page to the non-volatile memory to free the memory page. The query processor selects the memory page according to selection criteria that favors output of full memory pages over partially filled memory pages and that favors memory pages with a low absorption rate. Data records with low activity are written to non-volatile memory in the interest of preserving data records with high absorption rates on the memory pages with the hope of absorbing future data records.
|
