Priority Enhanced Stride Scheduling

1
Priority Enhanced Stride Scheduling

• Damien Le Moal, Mineyoshi Masuda, Masahiro
  Goshima,
• Shin-ichiro Mori, Yasuhiko Kitamura,
• Toshiaki Kitamura and Shinji Tomita

Graduate School of Informatics, Kyoto University
2
Overview

• Introduction
• context and objectives
• Background
• Priority based scheduling
• Fair-share scheduling Stride scheduling
• Colonia local scheduling
• Priority enhanced stride scheduling
• Evaluation results
• Conclusion and future work

3
Introduction Context

• Computer Colony a cluster of workstations
• High-bandwidth, low-latency communications
• hardware support for cache coherent shared
  distributed memory
• User level communications
• Colonia distributed operating system
• Implement a massively parallel computer
• Fine-grain parallelism support
• Dynamic preemptive process migration

4
Introduction Objectives

• In such environment
• Various workload execution
• User interactive, I/O bound and compute bound
  processes
• Parallel applications
• More users execute applications on a single node

5
Introduction Objectives

• As a result, process scheduling needs
• Support for a wide range of applications
• In particular user interactive processes
• Fair-share allocation of CPU time
• Among users and processes
• System wide
• But
• Adaptive to the current workload on highly
  loaded nodes only

6
Colonia Distributed Scheduling

• Approach
• Local and global scheduling
• Fair-share scheduling locally to each node
• Load balancing implements global fairness

Global fair-share scheduling
7
Priority Scheduling

• UNIX systems
• Priority scheduling
• General
• Simple to implement, small overhead
• Dynamic priority calculation efficient execution
  of a wide range of applications
• CPU time allocation to users depends on the
  number of processes
• CPU starvation problem

8
Fair-share Scheduling

• Fair-share scheduling
• Share calculated with a currency (tickets)
• Hierarchical fair-share share defined for users
  and processes
• Priority-based fair-share scheduling
• Use existing method
• Priority calculation method difficult
• Tuning not easy
• Fairness realized over long periods

9
Stride Scheduling

• Proportional-share scheduling methods
• Execute processes depending only on their share
  (tickets)
• Stride scheduling

At each quanta, the process with the smallest
pass is chosen
10
Stride Scheduling

• Stride scheduling problems
• Simple algorithm
• Fairness realized over short intervals
• Overhead depends on the number of processes
• Basic implementation degrade response time
• Process behavior not considered
• Not preemptive

11
Colonia Local Scheduling

• Solution
• Combination of two methods
• Priority scheduling to control process execution
  order
• Stride scheduling to implement simply overall
  fairness

Policy needed to decide on the use of either
priority or stride scheduling
12
Enhanced Stride Scheduling

• Stride scheduling implementation
• Avoiding sorting the ready queue in pass order
• Idea replacing pass with an array index

• Processes are chosen from the head list

13
Priority Enhanced Stride Scheduling

• Priority and Stride scheduling combination use
• New processes or processes waking up
• Scheduled first with their priority
• Temporary unfair CPU time allocation tolerated
• Depending on share and CPU usage
• Stride scheduling used

14
Priority Enhanced Stride Scheduling

• Priority and stride scheduling use

CPU time allocation
Share
Time
15
Priority Enhanced Stride Scheduling

• Control CPU usage accounting
• Classically
• Over process life time
• Recently used share
• Over a smaller time interval (few seconds)

16
Priority Enhanced Stride Scheduling

• Ready queue
• Multi-level feedback queue
• Stride queue assigned a fixed priority
• Priority calculation
• Control stride queue usage
• Classical priority/time quantum calculation (SJF)

17
Evaluation results

• Compute bound processes

? Process 3 (500)
? Process 2 (300)
? Process 1 (200)
18
Evaluation results

• Mix of load

? User C (200)
? User B (300)
? User A (500)
19
Conclusion

• Flexible fair-share scheduling
• Precise Allocation of CPU time
• Response time minimized
• Scheme flexible adaptive to the load
• On lightly loaded systems only priority
  scheduling
• Future work
• Tuning
• System wide extension of fairness

20
Evaluation results

• Stride scheduling overhead

? Simple get
? Colonia get
? Simple put
? Colonia put
21
Evaluation results

• Scheduling overhead

? stride scheduling only
? priority scheduling only
22
Evaluation results

• Basic behavior

Allocated share
? Process 2 (500)
Time (sec)