slide 8 of 19

Scheduler Issues

• How many processes to make?
  • Lowest overhead when <= number of processors
  • More can overlap compute with I/O, but not much with one-at-a-time kernel entry
• How to cause a process to be scheduled?
  • Kernel scheduler can't tell who is waiting for shared memory interactions
  • Signals, blocking pipe/socket operations, etc. work
• How to bind processes to processors?
  • Lightly loaded SMP... it will just happen
  • Heavily loaded SMP... bump priority on the processes of the parallel program (see renice)