Task Parallel Library. Data Parallelism Patterns презентация

Task Parallel Library Data Parallelism Patterns

Introduction to Parallel Programming Introduction to Parallel Programming Parallel Loops Parallel Aggregation

Introduction to Parallel Programming Introduction to Parallel Programming Parallel Loops Parallel Aggregation

Hardware trends predict more cores instead of faster clock speeds Multicore system features

Some parallel applications can be written for specific hardware Potential parallelism

Decomposition Parallel programming patterns aspects

Tasks are sequential operations that work together to perform a larger operation Decomposition

Tasks that are independent of one another can run in parallel Coordination

Tasks often need to share data Scalable sharing of data

Understand your problem or application and look for potential parallelism across the entire application as a whole Parallel programming design approaches

Concurrency is a concept related to multitasking and asynchronous input-output (I/O) Concurrency & parallelism

With parallelism, concurrent threads execute at the same time on multiple cores Concurrency & parallelism

Amdahl’s law says that no matter how many cores you have, the maximum speedup you can ever achieve is (1 / percent of time spent in sequential processing) The limits of parallelism

Whenever possible, stay at the highest possible level of abstraction and use constructs or a library that does the parallel work for you Parallel programming tips

Use patterns Parallel programming tips

Based on the .NET Framework 4 Code examples of this presentation

Introduction to Parallel Programming Introduction to Parallel Programming Parallel Loops Parallel Aggregation

Parallel programming patterns

Use the Parallel Loop pattern when you need to perform the same independent operation for each element of a collection or for a fixed number of iterations Parallel Loops

Parallel.For

Parallel.ForEach

Almost all LINQ-to-Objects expressions can easily be converted to their parallel counterpart by adding a call to the AsParallel extension method Parallel LINQ (PLINQ)

Use PLINQ’s ForAll extension method in cases where you want to iterate over the input values but you don’t want to select output values to return PLINQ ForAll

The .NET implementation of the Parallel Loop pattern ensures that exceptions that are thrown during the execution of a loop body are not lost Exceptions

Parallel loops Parallel loops variations

Writing to shared variables Dependencies between loop iterations

Referencing data types that are not thread safe Dependencies between loop iterations

Sequential iteration Breaking out of loops early

Use Break to exit a loop early while ensuring that lower-indexed steps complete Parallel Break

Calling Break doesn’t stop other steps that might have already started running Parallel Break

ParallelLoopResult

Use Stop to exit a loop early when you don’t need all lower-indexed iterations to run before terminating the loop Parallel Stop

External Loop Cancellation

Special handling of small loop bodies

The number of ranges that will be created by a Partitioner object depends on the number of cores in your computer Special handling of small loop bodies

You usually let the system manage how iterations of a parallel loop are mapped to your computer’s cores, in some cases, you may want additional control Controlling the degree of parallelism

The PLINQ query in the code example will run with a maximum of eight tasks at any one time Controlling the degree of parallelism

Sometimes you need to maintain thread-local state during the execution of a parallel loop Task-local state in a loop body

Random initialization of the large array

Calling the default Random constructor twice in short succession may use the same random seed Random class in parallel

You can substitute custom task scheduling logic for the default task scheduler that uses ThreadPool worker threads Using a custom task scheduler

Step size other than one Anti-Patterns

Adaptive partitioning Parallel loops design notes

Introduction to Parallel Programming Introduction to Parallel Programming Parallel Loops Parallel Aggregation

Parallel programming patterns

The pattern is more general than calculating a sum The Parallel Aggregation pattern

Sequential version Calculating a sum

PLINQ Calculating a sum

PLINQ is usually the recommended approach Parallel aggregation pattern in .NET

The PLINQ Aggregate extension method includes an overloaded version that allows a very general application of the parallel aggregation pattern Using PLINQ aggregation with range selection

Aggregation using Parallel For and ForEach Design notes

Aggregation in PLINQ does not require the developer to use locks Design notes

Task Parallel Library Data Parallelism Patterns