The Application of Virtual Machine Technology to Petascale Systems

1
The Application of Virtual Machine Technology to
Peta-scale Systems
Sun Microsystems Proprietary

• Christopher Vick
• Senior Staff Engineer
• Sun Microsystems, Inc.

2
High Productivity Computing Systems
65
3
DARPA's Goals For Productivity

• Performance (time-to-solution) speedup
  critical
• national security applications by a factor
  of 10X to 40X
• Programmability (idea-to-first-solution)
  reduce
• cost and time of developing application
  solutions by 10X
• Portability (transparency) insulate research
  and
• operational application software from
  system
• Robustness (reliability) apply all known
  techniques
• to protect against outside attacks,
  hardware faults,
• programming errors

Productivity
Machine Performance
Human Performance
4
Overview

• Virtualization
• Safety
• Integration of Static and Dynamic Operations

5
Virtualization

• Peta-scale hardware will be very complex
• Abstraction hides hardware complexity from
  programmers
• Reduction in complexity improves both the speed
  of coding and the robustness of the resulting code

6
Verification and Validation

• Verify that code correctly implements the
  mathematics
• Validate that the code accurately predicts nature
• Simplified machine model eases burden on
  scientist for verification
• Optimization technologies keep solid performance
  for validation

7
Portability

• A Key issue in making code portable is the
  presence of machine dependencies
• A virtual machine can significantly reduce these
  while maintaining performance
• Language independence

8
Observability

• Abstraction layer hides machine behaviors
• Complicates both debugging and performance
  analysis
• Challenge is to maintain observability while
  reducing complexity burden

9
Safety

• A key lesson of JavaTM is that improved safety
  improves productivity
• Pointer safety
• Type safety
• Bounds checking
• Safety ! Poor Performance

10
Safety Examples

• Array Bounds Checking
• Generally has minimal performance impact
• Will not handle pointer increment styles
• Data Race Detection
• Key issue on highly parallel systems
• Data races are not always bad
• May impact performance

11
Numeric Safety

• Floating point precision problems can heavily
  impact productivity
• Runtime checks can detect loss of precision or
  indeterminate results
• Automate computation of error bounds
• Ease burden of verification/validation

12
Default Behaviors

• Safety features should be on by default
• Encourages vendors to improve accuracy and
  performance of features
• Provides maximum safety during development cycle
• Provide switches to turn off where performance
  impact dictates

13
Language Support

• Add safety features at the language level
• Tag blocks of code where data races are benign
• Permits more precise application than
  compiler/system switches
• Further exploration is needed

14
Static/Dynamic Mechanisms

• Combine static analysis with dynamic optimization
• Limit runtime impact of expensive analysis
  techniques
• Benefit from runtime profiling and checking
• Optimistic Optimizations
• Deoptimization
• Runtime profiling

15
Speculative Parallelism

• Increased parallelism is requirement for
  peta-scale systems
• Apply optimistic optimization to parallelization
• Combine expensive static analysis with runtime
  checks for correctness
• Deoptimization may present some special
  challenges in this context

16
Performance Analysis

• Layered profiling
• Runtime profiling for dynamic optimization
• Maintain profiles across runs to improve
  resolution and detect long-term or phased
  behaviors
• Key issues of when to ignore/reset profile data

17
Dynamic Instrumentation

• Runtime recompilation allows dynamic
  instrumentation of programs
• Debugging
• Allow program to run for a while, and then
  recompile to add breakpoints, etc.
• Key feature for debugging long running codes
• Profiling
• JFluid

18
The Application of Virtual Machine Technology to
Peta-scale Systems
Sun Proprietary