25 Jan 98
Memory is allocated dynamically. It is not possible to fix portions of memory for specific databases. The SQL Server process space does not change size. Only the buffer pool will grow and shrink.
Buffer management is being done by a clock algorithm rather than by a least recently used algorithm. This is reported to improve performance and increase scalability.
The scanning thread directs Read Ahead. There are no configuration parameters. The absence of separate read-ahead threads reduces context switching.
Index reads can be issued from intermediate nodes rather than needing to go to the leaf level in all cases.
There is full row-level locking. It is not necessary to configure the number of locks. The lock escalation configuration options are gone.
Key range locking is used in indexes. This means that there is more concurrent update and also reduces the amount of locks required for page splits.
The lock granularity is determined at run time based on the lowest cost strategy. This is based on input from the Query Processor. Several different lock granularities may be used for a single query, for example, there may be page locks for index scans and row locks for reading from the table.
Many thanks to Sharon for providing these notes - drop her a note at email@example.com and tell her thanks!