Speculative Parallelization of Partially Parallel Loops

1
Speculative Parallelization of Partially Parallel
Loops

• Francis Dang and Dr. Lawrence Rauchwerger
• Department of Computer Science,
• Texas AM University
• http//www.cs.tamu.edu/people/fhd4244
• http//www.cs.tamu.edu/faculty/rwerger

2
Motivation

• Static compiler methods cannot always extract all
  the parallelism in loops because
• Access patterns are too complex.
• Required information is not available at
  compile-time.
• Can use run-time methods to parallelize more
  loops.

3
Partially Parallel Loops

• Loops can be
• Fully parallel (doall)
• Fully sequential
• Partially parallel
• Partially parallel loops
• Not all iterations can be executed independently
• May still have enough parallelism to exploit

4
Partially Parallel Loops Example

• do i 1, 8
• z AKi
• ALi z Ci
• enddo
• K15 1,2,3,1,4,2,1,1
• L15 4,5,5,4,3,5,3,3

5
Related Work

• Inspector/Executor (Saltz, Zhu and Yew,
  Rauchwerger, Mellor-Crummey, et al.)
• Advantage Works well for partially parallel
  loops
• Disadvantages
• A proper side-effect free inspector is not always
  available.
• Can require large additional data structures
• LRPD Test (Rauchwerger)
• Advantage Works well for fully parallel loops
• Disadvantage Slowdown is proportional to the
  speculative parallel execution time.

6
Recursive LRPD

• Main Idea
• Transform a partially parallel loop into a
  sequence of fully parallel loops.
• Iterations before the first data dependence are
  correct and committed.
• Reapply the LRPD test on the remaining
  iterations.
• Blocked scheduling
• Worst case
• Sequential time plus testing overhead.

7
Recursive LRPD Algorithm
8
Recursive LRPD Implementation

• Implemented with our run-time pass in Polaris and
  hand-inserted code.
• Privatization for arrays under test
• Replicated buffers for reductions.
• Checkpoint shared arrays.
• Record memory references during execution.

9
Recursive LRPD Example

• do i 1, 8
• z AKi
• ALi z Ci
• enddo

K15 1,2,3,1,4,2,1,1 L15
4,5,5,4,2,5,3,3
10
Work Redistribution

• Redistribute remaining iterations across all
  processors.
• Advantage
• Execution time for each stage will decrease.
• Disadvantages
• May uncover new dependences across processors.
• May have remote cache misses from data
  redistribution.

11
Work Redistribution Illustration
With Redistribution
Without Redistribution
12
Work Redistribution Example

• do i 1, 8
• z AKi
• ALi z Ci
• enddo

K15 1,2,3,1,4,2,1,1 L15
4,5,5,4,2,5,3,3
13
Redistribution Model

• Redistribution may not always be beneficial.
• Stop redistribution if
• The cost of data redistribution outweighs the
  benefit from redistribution.
• Used a synthetic loop to model this adaptive
  method.

14
Redistribution Model
15
Experiments

• Setup
• 16 processor HP V-Class
• 4 GB memory
• HP-UX 11.0
• Codes and Loops

16
Input Profile
17
Experimental Results
18
Experimental Results
19
Experimental Results
20
Experimental Results
21
Issues

• May have load imbalance due to blocked
  scheduling.
• Checkpointing can be expensive.
• Work redistribution
• May uncover more dependencies.
• May cause remote cache misses from data
  redistribution.

22
Feedback Guided Block Scheduling

• Use the timing information from the previous
  instantiation (Bull, EuroPar 98).
• Estimate the processor chunk sizes for minimal
  load imbalance.

23
Conclusion

• Contributions
• Can speculatively parallelize any loop.
• Concern is now not when to parallelize but
  optimizing the parallelization.
• Future Work
• Use feedback-guided block scheduling to minimize
  load imbalance.
• Decrease run-time overhead.
• Use dependence distribution information for
  adaptive redistribution and scheduling.