Multi-Processor Scheduling

Notes

• Chapter 5.5
• Remember
  • Cores share memory (possibly private L1/L2, but everything else is shared)
  • Processors that share memory are called symmetric multiprocessors
  • Non Uniform Memory Access (NUMA) is where each processor has its own memory
  • Distributed/Heterogeneous systems have independent
• Each of these has different needs for a scheduler
  • Process migration is the biggest thing
  • When a core/cpu/machine is busy, can I ship the work to another core/cpu/machine?
    • No real problem between cores
    • Very little problem in a SMP
    • Slow data transfer in a NUMA
    • Problem or impossible in a Distributed system.
• This is related to the memory hierarchy
  • Moving between cores is a cache flush, and a compulsory miss, but possibly on l1 and l2
  • Moving between processors in SMP is possibly a full cache flush
  • On NUMA it is transferring memory between locations
  • In Distributed ...
• In core/smp a common ready queue or a per thread/cpu ready queue could work.
• In NUMA and beyond a per cpu/machine queue is more likely.
• With more than one queue, load balancing is desirable
  • Different meanings of balance include
    • Same sized queue
    • Similar distribution of priorities
    • And others
  • Pull/push question again
    • An idle unit may request work
    • Or an overloaded unit may push work.
  • It appears that in linux an idle processor may pull work from a busy processor (but I can't find a good reference)
• Trying to keep a task on a processor is know as processor affinity