Dev Blog

Summary ​

Conclusions ​

Data parallelism applies one operation to parts of a large dataset and scales with data volume. A loop can be parallelized if its iterations are independent; dependencies through shared variables are eliminated by reduction with thread-local state. The Parallel class provides For and ForEach loops with ParallelOptions settings, early termination through Break/Stop, and exception collection in AggregateException, plus ForEachAsync and ForAsync for asynchronous bodies. Performance depends on data partitioning: range partitioning has the least overhead, chunk and dynamic partitioning balance the load, and Partitioner.Create with ranges is needed for small loop bodies. PLINQ parallelizes LINQ queries with a single AsParallel call, but does not preserve order without AsOrdered and is not always faster than LINQ. Parallel aggregation is correct only for associative operations (and, without ordering, commutative ones). A reduction tree and the Blelloch algorithm have logarithmic depth, while parallel sorting algorithms (merge sort, quicksort, odd-even sort, sample sort) combine data partitioning with tasks and a threshold for switching to a sequential algorithm. Evaluate results using speedup and efficiency tables for strong and weak scaling.

Self-check questions ​

  1. How does data parallelism differ from task parallelism?
  2. Under what conditions can loop iterations run in parallel?
  3. What settings does ParallelOptions provide?
  4. How do Break and Stop differ? What does ParallelLoopResult contain?
  5. How are exceptions handled in parallel loops and PLINQ?
  6. What are the localInit and localFinally delegates for?
  7. When do you need Parallel.ForEachAsync and Parallel.ForAsync?
  8. Compare range, chunk, and dynamic data partitioning.
  9. Why use Partitioner.Create(from, to, rangeSize)?
  10. What merge modes does PLINQ provide, and how do they differ?
  11. How does ordering affect PLINQ results and performance?
  12. When is a PLINQ query slower than LINQ?
  13. Why must an aggregation operation be associative? When is commutativity also required?
  14. Describe a reduction tree and the Blelloch prefix sum algorithm.
  15. How does parallel merge sort work, and what limits its speedup?
  16. How do you make a parallel Monte Carlo method reproducible?