CS837: ADV.

OPERATING SYSTEMS
MULTIPROCESSOR & REAL-TIME
SCHEDULING

Dr. Mian M.Hamayun

[email protected]
http://seecs.nust.edu.pk/faculty/mianhamayun.html
 TYPES OF MULTIPROCESSOR
SYSTEMS
 Distributed Systems: Loosely connected
relatively independent systems e.g. PCs
connected via internet.
 Non-uniform or Specialized Processor
Systems: Systems with specialized processors
e.g. I/O processors, Vector processors etc.
 Symmetric Multiprocessors(SMP): Similar
processors or chips generally connected by a
common bus or specialized network. Share
memory and other resources.

2
 GRANULARITY OF PARALLELISM
Another characteristic of MP systems is the frequency of
synchronization between processes in a system.
1) Independent Parallelism: No explicit synchronization among
processes e.g. separate applications i.e. time-sharing system
2) Very Course Grain Parallelism: Loosely coupled tasks that
require no or little synchronization. e.g. Large Application
compilation (with many sources files).
3) Coarse Grain Parallelism: Process level synchronization. e.g.
a process spawning multiple other processes.
4) Medium Grain Parallelism: Frequent synchronization required.
e.g. parallelism between multiple threads of the same process.
5) Fine Grain Parallelism: Parallelism down to the instruction
level e.g. DSP/VLIW applications, ILP (Instruction Level
Parallelism)

3
GRANULARITY OF PARALLELISM

4
DESIGN ISSUES
Scheduling on a multiprocessor system involves the following
issues:

1. Assignment of a Processor to a Processes: Decision

of where to run a process.
2. Use of multiprogramming on individual processors:
How to run multiple processes on an individual
processor.
3. Process Dispatching: Actual decision of which process
to run next.
The approach taken will depend on the degree of granularity of
applications and the number of processors available.

5
DESIGN ISSUES
Assignment of a Processor to a Processes:
 Assuming all processors are equal, it is
simplest to treat processors as a pooled
resource and assign process to processors on
demand.
 Should the assignment be dynamic or static?

6
DESIGN ISSUES
Assignment of a Processor to a Processes:
 Dynamic Assignment
 Threads are moved from a queue of one processor
to a queue of another processor
 Static Assignment
 Permanently assign process to a processor
 Dedicate short-term queue for each processor &
requires less scheduling overhead
 Allows the use of ‘group’ or ‘gang’ scheduling (later)
 May leave a processor idle with an empty queue
while others have a backlog!

7
DESIGN ISSUES:
SCHEDULING ON AN SMP
Master/Slave Model: The kernel routines are run on
only one of the cores. The other cores are assigned
tasks by the master.
 Advantages
 Simpler to implement
 Simple conflict resolution
 Disadvantages
 Failure of the master may bring down the whole system
 The master may become a performance bottleneck

8
DESIGN ISSUES:
SCHEDULING ON AN SMP
Peer Model: The scheduling algorithm is run on any
core that is idle and needs a new job.
 Advantages
 Can tolerate System Failures
 OS execution load is distributed on the entire system
 Disadvantages
 Difficult to implement
 OS must ensure that two processors don’t select the
same process or some processes disappear.
 Shared system data structures may become a
bottleneck.

9
COMPARISON: ONE VS. TWO
PROCESSORS
Dual-processor system has ½ of the
processing rate of a single
processor system.
Coefficient of Variation Cs = 0;
No variability in service times of
processes i.e. services times are
equal for all processes.

A single process with long service

time is much less disruptive in the
FCFS case.
General Conclusion:
Specific scheduling discipline is
much less important with two or
more processors.

10
THREAD SCHEDULING ON
MULTIPROCESSORS
 An application can be a set of threads that
cooperate and execute concurrently in the
same address space.
 Threads serve as structuring aid and to
overlap I/O and processing on uniprocessor
systems.
 True parallelism on multi-processor systems.
 Dramatic gains in performance are possible as
compared to uniprocessor systems.

11
APPROACHES TO THREAD
SCHEDULING ON MULTIPROCESSORS
Many proposals exist but four general approaches
stand out:
1) Load Sharing
2) Gang Scheduling
3) Dedicated processor assignment
4) Dynamic scheduling

12
THREAD SCHEDULING:
LOAD SHARING
Usually threads are not dedicated to processors!
 Whenever a processor is idle it pulls a job from a
single shared queue or multiple queues are
used for priorities.
 Load evenly distributed on all processors.
 No central scheduler needed.
 The system could be viewed as a multi-server
queuing architecture.
 The central queue(s) needs mutual exclusion and
may become a bottleneck!

13
THREAD SCHEDULING:
GANG SCHEDULING

 A set of related threads is run simultaneously

on a set of processors on a one-by-one basis.
 Synchronization, blocking time may be reduced
and performance may increase. Applications
may make faster progress.
 Scheduling overhead may decrease as a
single decision affects a number of threads.
 Performance usually superior to Load Sharing.

14
THREAD SCHEDULING:
GANG SCHEDULING – EXAMPLE
Suppose we have a four processor system and we
have two applications to run. One with 4 threads and
the other with just one thread.
 System is shared equally among the two applications
(Uniform Division)
Utilization
Group 1 Group 2 Individual System
PE 1
T1 T1 100% 25.0%
PE 2 Utilization: 62.5%
T2 Idle 50% 12.5% Wastage: 37.5 %
PE 3 T3 Idle 50% 12.5%
PE 4
T4 Idle 50% 12.5%

15
THREAD SCHEDULING:
GANG SCHEDULING – EXAMPLE
We have 5 threads we may give 4/5 to application 1
and 1/5 of the system share to application 2.

Group 1 Group 2 Utilization

(Share: 4/5) (Share 1/5) Individual System
PE 1
T1 T1 100% 25.0%
Utilization: 85.0%
PE 2 80% 20.0% Wastage: 15.0%
T2 Idle
PE 3 T3 Idle 80% 20.0%
PE 4
T4 Idle 80% 20.0%

16
THREAD SCHEDULING: DEDICATED
PROCESSOR ASSIGNMENT
 Each program is assigned a number of processors equal
to the threads in the program.
 The processors remain dedicated to the program till
its completion.
 Some processors may be idle: there is no
multiprogramming of processors!
 Reverse of Load Sharing.
 Extreme form of Gang Scheduling.
 May result in good performance for systems with a large
number of processors by avoiding process/thread
switching during the lifetime of a program.

17
THREAD SCHEDULING: DEDICATED
PROCESSOR ASSIGNMENT

18
THREAD SCHEDULING: DYNAMIC
SCHEDULING
 Number of threads of the program can be
altered at runtime on system demand.
 Special Runtime support needed for this kind of
system.
 Simulations suggest that this form of scheduling
is superior to both Gang Scheduling and
Dedicated Assignment.
 Experience with actual systems is needed to
prove the worth of dynamic scheduling.

19