Designing Parallel Operating Systems using Modern Interconnects

1
Designing Parallel Operating Systems using Modern
Interconnects
Process Scheduling for the Parallel Desktop
Eitan Frachtenberg Los Alamos National
Laboratory 07 December 2005
Computer and Computational Sciences Division Los
Alamos National Laboratory
Ideas that change the world
2
Problem Outline

• Commodity processors are becoming increasingly
  parallel
• Demanding software will necessarily follow suit
  and be written for parallelism
• Parallel programs and complex workloads will
  require specialized OS support
• Focus on schduling
• Case studies and lessons learned from the HPC
  world

3
Hardware Trends

• Moores law cannot continue forever
• Annual single-core performance growth down from
  50-60 to 15
• Power and efficiency considerations
• More efficient to have multiple, slower threads
• Resulting in an alphabet soup of technologies
• CMP, SMT, SMP, HT,

4
Chip Parallelism
5
Software Trends

• Development of commodity parallel software
  stifled
• No incentives (single-core performance growth)
• Difficult to develop parallel code
• However, with parallel commodity hardware,
  parallel code will be the only way to extract
  additional performance
• Software makers traditionally vied to exploit
  additional performance and hardware advances
• Parallel paradigm ? Object Oriented paradigm?
• Modularity, responsiveness, resource overlapping

6
Commodity Parallel Software

• Engineering
• Games
• Complex desktop/database searches
• Content creation, smart video processing
• Science
• Finance
• User interfaces
• Multitasking

7
Challenges for the OS

• Scheduling in commodity OSs based on 30-year old
  principles. In particular, little support for
  parallel applications and workloads beyond mere
  resource allocation.
• Already small SMPs and SMTs expose many
  difficulties.
• As the degree of parallelism in the underlying
  hardware will increase, so will the scheduling
  requirements.
• Commodity schedulers are challenged at all levels
  of parallel abstraction
• Thread
• SMP / CMP
• Cluster

8
Parallel Scheduling

• Parallel scheduling is a complex and active field
  of research with supercomputers and clusters.
• Many algorithms and implementations exist to deal
  with the main challenge synchronization.
• However, parallel scheduling techniques are not
  directly applicable to desktops
• Interactive programs, responsiveness demands
• Mixed, heterogeneous workloads

9
Scheduling Examples

• We devised a simple synthetic benchmark to
  demonstrate parallelism effects
• Various contemporary parallel desktops
• Parallel programs BSP-style compute and synch
• Serial programs (stressors) only compute
• All programs launched together, experiment ends
  when parallel programs finish
• Two schedulers default Linux and Gang Scheduling

10
HP (Compaq) ES40

• 4-way Alpha (Linux 2.4)

11
HP (Compaq) ES45

• 4-way Alpha (Linux 2.6)

12
IBM / Intel

• 4-way Pentium III (Linux 2.6)

13
Xeon SMP

• 4-way Xeon MP with HT (Linux 2.6)

14
Interactive Processes

• Scenario foreground interactive process (e.g.,
  video playback) interrupted by batch background
  process (e.g., virus scan or download client)
• Even when computing resources are adequate for
  all processes, mis-scheduling the interactive
  process causes unnecessary interruptions

15
Unloaded system
16
Loaded System
17
Scheduler Challenges Summary

• Processes competing over the same resources
  suffer when coscheduled
• Collaborating processes suffer when not
  coscheduled
• Interactive processes require specially-timed
  scheduling
• The challenge identifying and servicing
  conflicting scheduling requirements

18
Conclusions

• Commodity computers and their workloads continue
  to evolve, and we seem to be heading for an era
  of ubiquitous parallel computing
• Complex architectures and workloads pose a new
  challenge to commodity schedulers, that remained
  largely unchanged in the last 30 years
• To face this rising complexity, schedulers will
  also have to adapt, potentially with more
  sophisticated heuristics.

19
Future Work

• We are developing scheduling solutions for the
  parallel desktop based on the following
  principles
• Classification Identify the scheduling
  requirements of processes by monitoring OS
  activity
• Cooperation Coschedule collaborating processes
• Separation Schedule apart interfering processes
• Adaptivity and automatic tuning Dynamically try
  out different combinations and choose the best