Distributed Process Scheduling: A System Performance Model

1
Distributed Process Scheduling A System
Performance Model

• Vijay Jain
• CSc 8320, Spring 2007

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
Overview

• Before execution, processes need to be scheduled
  and allocated with resources
• Objective
• Enhance overall system performance metric
• Process completion time and processor utilization
• In distributed systems location and performance
  transparency
• In distributed systems
• Local scheduling (on each node) global
  scheduling
• Communication overhead
• Effect of underlying architecture

4
Process Interaction Models

• Precedence process model Directed Acyclic Graph
  (DAG)
• Represent precedence relationships between
  processes
• Minimize total completion time of task
  (computation communication)
• Communication process model
• Represent the need for communication between
  processes

5
Process Interaction Models

• Optimize the total cost of communication and
  computation
• Disjoint process model
• Processes can run independently and completed in
  finite time
• Maximize utilization of processors and minimize
  turnaround time of processes

6
Process Models
Partition 4 processes onto two nodes
7
System Performance Model
Attempt to minimize the total completion time of
(makespan) of a set of interacting processes
8
System Performance Model (Cont.)

• Related parameters
• 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 the actual time
  achieved on a n-processor system with the
  concurrent algorithm and a specific scheduling
  method being considered
• OCPTideal optimal concurrent processing time on
  an ideal system the best time that can achieved
  with the concurrent algorithm being

9
System Performance Model (Cont.)

• considered on an ideal n-processor system (no
  interprocessor communication overhead) and
  scheduled by an optimal scheduling policy
• 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

10
System Performance Model (Cont.)
Pi the computation time ofthe concurrent
algorithm onnode i
(RP ? 1)
11
System Performance Model (Cont.)
(The smaller, the better)
12
System Performance Model (Cont.)

• RP Relative processing
• Shows how much loss of speedup is due to the
  substitution of the best sequential algorithm by
  an algorithm better adapted for concurrent
  implementation but which may have a greater total
  processing need
• Sd
• Degradation of parallelism due to algorithm
  implementation

13
System Performance Model (Cont.)

• RC Relative concurrency
• How far from optimal the usage of the n-processor
  is
• RC1 ? best use of the processors
• ? Efficiency Loss is loss of parallelism when
  implemented on a real machine.
• ? can be decomposed into two terms
• ? ?sched ?syst

14
Efficiency Loss ?
Impact factors scheduling, system, and
communication
15
Efficiency Loss ? (Cont.)
16
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

17
Workload Distribution

• Performance of systems described as queuing
  models can be computed using queuing theory. An
  X/Y/c queue is one where
• X Arrival Process, Y Service time distribution,
  c Numbers of servers
• ? arrival rate ? service rate ? migration
  rate
• ? depends on channel bandwidth, migration
  protocol, context and state information of the
  process being transferred.

18
Processor-Pool and Workstation Queueing Models
Static Load Sharing
Dynamic Load Balancing
M for Markovian distribution
19
Comparison of Performance for Workload Sharing
(Communication overhead)
(Negligible Communication overhead)
?0 ? M/M/1 ???M/M/2
20
References

• Distributed Operating Systems and Algorithms by
  Randy Chow and Theodore Johnson
• Opearting System Concepts by Silberschatz,
  Galvin and Gagne
• Time Comparative Simulator for Distributed
  Process Scheduling Algorithms, Transactions on
  Engineering, Computing and Technology Volume 13
  May 2006 ISSN 1305-5313, Nazleeni Samiha Haron,
  Anang Hudaya Muhamad Amin, Mohd Hilmi Hasan,
  Izzatdin Abdul Aziz,and Wirdhayu Mohd Wahid