Scheduling For Distributed Systems

1
Scheduling For Distributed Systems

• Classification degree of coupling
• Classification granularity
• Local vs centralized scheduling
• Methods

2
Motivation

• Additional computing capability can be added more
  easily by adding than enlarging processors
• Users and tasks are distributed, but in different
  ways
• Many applications and languages are now threaded

3
Classification degree of coupling
4
Classification granularity
5
What scheduling involves

• Assigning processes to processors
• static assignment process runs there until
  completion
• Common queue time slot is the unit of
  assignment
• Peer architecture each processor selects
  process to run
• Using multiprogramming on individual processors
• Primary issue is how threads are handled
• Goal is best performance rather than 100
  utilization
• Actual dispatching of a process
• Coefficient of variance idea variability of
  service times
• Simpler scheduling can work for multiprocessors
  without causing bottlenecks

6
Thread scheduling

• Load sharing
• Global queue of ready threads this is not load
  balancing
• Gang scheduling
• Related threads are scheduled to run on a set of
  processors at the same time
• Dedicated processor assignment
• Application is alloted a processor per thread
  until done
• Dedication of processors reduces thread-switching
  overhead
• Dynamic scheduling
• Number of threads in a process can change while
  it runs
• Scheduler allots processors on a demand basis (it
  can take them away)

7
Real-time Scheduling

• Classification hard vs soft real-time
• Classification aperiodic/periodic
• Responsiveness requirements
• Reliability requirements

8
Real-time OS characteristics
9
RT scheduling approaches
Tacit assumption tasks are identified with
deadlines and durations
10
Deadline scheduling explicitly
11
Rate monotonic scheduling (RMS)

• Basic idea
• Suboptimal algorithm
• High arrival rate high priority
• Empirical observations
• RMS is almost as good as earliest-deadline
  scheduling
• Soft-RT parts of tasks can be scheduled to use
  the CPU time not used by the hard-RT work
• Stability is better than earliest-deadline

12
UNIX SVR4 scheduling

• Single priority queue
• Three priority classes
• Real-time 100-159
• Kernel 60-99
• Time-shared 0-59
• Each level is actually a priority queue
• Preemption is done when a high-priority RT
  process arrives
• Time-sharing class
• Penalty box priority is reduced if process uses
  all its time
• Time allowed varies from
• 10 ms priority 59
• 100 ms priority 0
• Observations system has much adaptability

13
Linux RT scheduling

• Basic approach
• Like non-RT Linux except RT processes have
  highest priority
• Three task queues
• SCHED_FIFO
• SCHED_RR also has time quota
• SCHED_OTHER executes only if no RT is waiting
• Multiple priorities within each class
• This description differs from present methods
• Continuous revision area of Linux

14
Windows 2K Scheduling

• Priority-driven preemptive scheduler
• 2 bands
• Real-time
• All threads have fixed priority
• All threads in a given level are in a RR queue
• 16 levels (31-16)
• Other
• All threads have variable priority
• FIFO at each level behavior-based priority
  modification
• 16 levels (15-0)
• Single vs multiple processor
• Single highest priority thread
• Multiple n-1 highest priority threads each get
  own processor
• Last processor is shared by all others