Ctrl+K

Hardware Path Optimizations

Hardware Path Optimizations#

The following section contains information about internal optimizations for hardware path supported by Intel® Query Processing Library (Intel® QPL).

Multi-Descriptor Processing#

The library can split qpl_op_scan_* operation onto several sub-tasks and perform them in paralleled mode. In other words, if the following conditions are met:

  • This is qpl_op_scan_* operation.

  • Nominal bit array output, i.e. no output modifications.

  • Input data size is at least 32kB large.

  • Input is in Little- or Big-Endian format (no encoded input support).

  • Aggregates and checksums are disabled via appropriate flags.

  • No more than 1 work queue per 1 device configured.

  • Operation should be executed with synchronous interface qpl_execute_job().

Note

The last limitation will be removed in some of the future releases.

Then the operation is auto-paralleled on the library level during execution on hardware path. This is intended to receive better performance when processing large streams.

Contents