Multi-processor Scheduling

Objectives

1. Discuss the impact of multiple cores/processors on scheduling.

Notes

• SMP - Symmetric Multiprocessor
  • All cores/processors have equal access to all memory
  • But each processor has a dedicated cache.
  • But they also use this for Symmetric Multiprocessing.
• NUMA - Non Uniform Memory Access
  • Each processor/group of processors has a memory
  • Each processor has dedicated cache.
• Moving processes on these different architectures have different costs.
  • Staying on a core is best.
  • Staying on a processor is good
  • Moving between processors becomes expensive, with the cost depending on the architecture.
• A first thought:
  • Dedicate one thread of execution to the kernel
    • But this can lead to bottlenecks and slowdowns.
    • And it is not very efficient.
• So then, what do we do about scheduling?
  1. Have a single queue, and the next selected process goes on the next available core?
  2. Have a per-core queue
• In general, a critical section leads to inefficiency.
  • Removing a process from the queue would be critical.
  • As would placing a new process on the queue.
  • In the single queue case, this is expensive.
  • So choice A is not good.
• In addition, choice A does not promote processor affinity.
• Choice B, however, leads to balance issues
  • One core can become lightly loaded while others are heavily loaded.
• The author proposes two solutions to load balancing
  • Push migration: a periodic task reblanaces the load among processors/cores.
  • Pull migration: an idle core searches for processes to run.
  • The author notes that some OS's do both.
• Hardware thread scheduling is covered in 5.5.2
  • Hardware manufacturers have built mechanisms to allow hardware to handle multiple threads per core.
    • This is from studies indicating burst patterns
    • The core can support quickly swapping several threads
    • This is done by the hardware, not the OS.
  • On a system with n cores, the OS believes there are 2n (or even larger multiples) cores.
  • The OS schedules a thread on each core, then the hardware schedules between the the 2 threads on a single core.
  • But if the OS is aware of the hardware scheduling algorithm it can make some decisions to improve performance
    • The book's examples are weak.
    • See for a proposed patch.
  • Be aware that for CPU bound applications, hyper-threading can be slower.
• Processor Affinity
  • If a scheduler trys to keep threads on the same processor, but will allow migration, this is called dsoft affinity
  • If it does not allow absolute migration, but only within a processor it is called hard affinity.
  • Apparently linux will support either.
  • Affinity is very important in NUMA applications
  • Apparently there is real conflict between load balancing and processor affinity.
  • The author states "Thus scheduling algorithms for modern multicore NUMA systems have become quite complex."
    • The last issue they introduce here is heterogeneous multiprocessing
      • My phone has a 2x2.2 GHz Cortex A-76 and a 6x2.0 GHz Cortex-A55) CPU.
        • These are both ARM processors
        • But have different cache sizes.
        • And apparently different pipeline characteristics.
      • The A-55 is slow but low energy
      • The A-76 is fast by high energy.
      • The scheduler is (hopefully) aware of this and will run
        • Low priority, long term tasks on the A-55
        • High priority, short term tasks on the A-76
        • And can work without the A-76 if we need to preserve power.