I/O Isn’t the RDBMS Bottleneck - It’s CPU

• Search Algorithms are Based on Disk Addresses

  • Multiple look-ups before you get to a memory pointer

• Data is Frequently Copied

  • Buffers must be available to be used by other pages

• Tables and Indices are Turned Into Random Pages

  • Serial table scans require a separate lookup sequence for each page
  • An index search invokes several page look-ups, then a data page search

• Optimization is Biased Toward I/O-Avoidance

  • No attention to instruction path length to perform memory operations
  • Optimizer must assume a page is on disk - it’s usually not