A System Performance Model

1
Distributed Process Scheduling

• A System Performance Model

2
Outline

• Overview
• Process Interaction Models
• A System Performance Model
• Efficiency Loss
• Processor Pool and Workstation Queuing Models
• Comparison of Performance for Workload Sharing
• References

3
For concurrent execution of interacting
processes-

• Communication and
• Synchronization between processes
• are the two essential system
  components
• Before processes can execute, they need to be-
• Scheduled and
• Allocated with resources.

4
Why scheduling?

• 1.To enhance overall system performance metrices
  like
• Process completion time and
• Processor utilization.
• 2. To achieve location and performance
  transparencies by distributed process scheduling.

5
Why scheduling in distributed systems is of
special interest

• This is so because of the issues that are
  different from those in traditional
  multiprocessor systems
• The communication overhead is significant.
• The effect of underlying architecture cannot be
  ignored.
• And the dynamic behaviour of the system must be
  addressed.

6
Process Models(in brief)

• Precedence Process Model
• Processes are represented by a DAG.
• Nodes- sequential processes
• Arcs- eg i to j requires that process I
  completes before j can start executing.

7
Communication Process Model

• Processes are created to coexist and communicate
  synchronously.
• So we have undirected edges.

8
Disjoint Process Model

• We assume that processes can be run independently
  of each other.
• So order in which processes are executed is not
  important.

9
System Performance

• Speedup
• -What are the factors on which it depends
• How to calculate speedup when we apply these
  factors

10
Speedup depends on three factors

• The design of the algorithm
• The efficiency of the scheduling algorithm
• The underlying system architecture.
• So if we take S as the speedup factor then the
  above dependencies can be represented as
• S F(Algorithm, System, Schedule)

11

• Where
• OSPT optimal sequential processing time the
  best time that can be achieved on a single
  processor using the best sequential algorithm.
• CPT concurrent processing time actual time
  achieved with the concurrent algorithm on an
  ideal n-processor system using an optimal
  scheduling policy.
• OCPTideal optimal concurrent processing time on
  an ideal system
• Si ideal speedup obtained by using a multiple
  processor system over the best sequential time
• Sd the degradation of the system due to actual
  implementation compared to an ideal system

12
Refined formula for speedup..

• n number of processors
• RP- Relative Processing requirement,
• RC- Relative Concurrency

13
. Refined formula for speedup

• Sd- degradation of parallelism due to algorithm
  implementation.

14
Final formula for speedup

• ?- Efficiency Loss, loss of parallelism when
  implemented on a real machine.
• ? can be decomposed into two terms
• ? ?sched ?syst

15
Efficiency Loss ?
16
Efficiency Loss ? (Cont.)
17
Workload Distribution

• Performance can be further improved by workload
  distribution
• Load sharing static workload distribution
• Dispatch process to the idle processors
  statically upon arrival
• Corresponding to processor pool model
• Load balancing dynamic workload distribution
• Migrate processes dynamically from heavily loaded
  processors to lightly loaded processors
• Corresponding to migration workstation model

18
Processor-Pool and Workstation Queueing Models
Static Load Sharing
Dynamic Load Balancing
M for Markovian distribution
19
Comparison of Performance for Workload Sharing
20
References

• Distributed Operating Systems and Algorithms by
  Randy Chow and Theodore Johnson
• Operating System Concepts by Silberschatz,
  Galvin and Gagne