Data Type

Integer

Default Value

1+2+4=7 for most machines

Description

This parameter specifies how the compute and I/O threads are bound to processors. The following values are bit flags. To obtain the value for NUMAFX, add the selected options. The following options are available:

  • 0, Do not bind threads.
  • +1, Bind compute threads to individual processing cores.
  • +2, Bind I/O threads to individual processing cores.
  • +4, Place the parent on the target processor before starting the thread.
    On some systems, this is required to locate the stack and data segments of the new thread on the target CPU.
  • +8, Request is mandatory (SGI only).
  • +16, Bind the compute threads to their appropriate node.
  • +32, Bind the I/O threads to their appropriate node.

These options provide an essential part of the NUMA technology. They provide the link between what memory is addressed by each thread and where the memory is located.

Two choices are available for binding threads. One choice binds the threads to specific processing cores. This has the advantage of assuring that maximum reuse of cache will occur, even if a thread is interrupted by an I/O process.

The second choice binds the threads to the node. This allows the threads to be scheduled on any processing core on the node. This allows the I/O and CPU threads to make maximum use of the compute resources on the node. However, if the thread looses its affinity for the processing core and is rescheduled on another processing core, additional cache loads will result.

FMS distributes the I/O threads uniformly among the nodes.