Pages

Tuesday, August 31, 2010

DEV305 Parallelize Your Microsoft .NET Framework-Based Applications with Parallel Extensions

Speaker: Ivan Towlson

Tech.Ed Online

Take advantage or newer processor architecture.

.NET Framework 4 has inbuilt functionality for using this new functionality.

Like LINQ, but faster.

Querying a large dataset.

Traditional approach requires large number of considerations for treading.



ParExtSamples

Works for anu IEnumerable
- Optimizations for other tyoes(T[], IList)

Writing a PLINQ query
- .AsParallel()
- ParallelEnumerable (Does the actual work)

Partitioning
- Operators are replicated across partitions for (ideally) complete isolation

Operator Fusion
- Avoids excessive merging and partitioning steps

Partitioning Algorithms
- Chunk - potential for a lot of locking when doing small tasks
  Progressively hands out more work out to each thread to separate both small and large jobs.
- Range - For IList equally divide work among threads
- Stripe - For IList
  Elements handed out round-robin to each other
  Less wasted effort in the case of exception
- Hash - For IEnumerable
  Elements assined to partition based on hash code.
- Custom - Partitioner

Merging
- Pipelined: Each result becomes available as they are completed.
- Stop-and-go: For OrderBy, ToArray, ToList, etc...
  Higher latency and more memory usage.
- Inverted: no merging needed
  ForAll extension method - no buffering required.

Parallelism Blockers
- Ordering not guaranteed (consider OrderBy)
- Exceptions (Stop other threads?) - System.AggregateException
- Thread affinity - Web form controls
- Operations with < 1.0 speedup - Overhead can make the entire operation slower.
- Side effects and mutability are serious issues
  Avoid shared state, side effects  

Task Parallel Library (TPL)
- Parallel.For(o, N, i => { work(i);});
- Parallel.ForEach

Regions
- Parallel.Invoke();

Task Concepts
- FromAync

Coordination Data Structures
- Thread-safe scalable collections
- Task and data management
- Synchronisation types
    SpinLock and SpinWait for very quick waiting periods.

No comments:

Post a Comment