Threading in C#

Joseph Albahari

Last updated: 2011-4-27

Translations: Chinese | Czech | Persian | Russian | Japanese

Download PDF

Part 5: Parallel Programming

 

Acknowledgements

Huge thanks to Stephen Toub, Jon Skeet and Mitch Wheat for their feedback — particularly Stephen Toub whose input shaped the entire threading article and the concurrency chapters in C# 4.0 in a Nutshell.

In this section, we cover the multithreading APIs new to Framework 4.0 for leveraging multicore processors:

These APIs are collectively known (loosely) as PFX (Parallel Framework). The Parallel class together with the task parallelism constructs is called the Task Parallel Library or TPL.

Framework 4.0 also adds a number of lower-level threading constructs that are aimed equally at traditional multithreading. We covered these previously:

You’ll need to be comfortable with the fundamentals in Parts 1-4 before continuing — particularly locking and thread safety.

All the code listings in the parallel programming sections are available as interactive samples in LINQPad. LINQPad is a C# code scratchpad and is ideal for testing code snippets without having to create a surrounding class, project or solution. To access the samples, click Download More Samples in LINQPad's Samples tab in the bottom left, and select C# 4.0 in a Nutshell: More Chapters.

Why PFX?

In recent times, CPU clock speeds have stagnated and manufacturers have shifted their focus to increasing core counts. This is problematic for us as programmers because our standard single-threaded code will not automatically run faster as a result of those extra cores.

Leveraging multiple cores is easy for most server applications, where each thread can independently handle a separate client request, but is harder on the desktop — because it typically requires that you take your computationally intensive code and do the following:

  1. Partition it into small chunks.
  2. Execute those chunks in parallel via multithreading.
  3. Collate the results as they become available, in a thread-safe and performant manner.

Although you can do all of this with the classic multithreading constructs, it’s awkward — particularly the steps of partitioning and collating. A further problem is that the usual strategy of locking for thread safety causes a lot of contention when many threads work on the same data at once.

The PFX libraries have been designed specifically to help in these scenarios.

Programming to leverage multicores or multiple processors is called parallel programming. This is a subset of the broader concept of multithreading.

PFX Concepts

There are two strategies for partitioning work among threads: data parallelism and task parallelism.

When a set of tasks must be performed on many data values, we can parallelize by having each thread perform the (same) set of tasks on a subset of values. This is called data parallelism because we are partitioning the data between threads. In contrast, with task parallelism we partition the tasks; in other words, we have each thread perform a different task.

In general, data parallelism is easier and scales better to highly parallel hardware, because it reduces or eliminates shared data (thereby reducing contention and thread-safety issues). Also, data parallelism leverages the fact that there are often more data values than discrete tasks, increasing the parallelism potential.

Data parallelism is also conducive to structured parallelism, which means that parallel work units start and finish in the same place in your program. In contrast, task parallelism tends to be unstructured, meaning that parallel work units may start and finish in places scattered across your program. Structured parallelism is simpler and less error-prone and allows you to farm the difficult job of partitioning and thread coordination (and even result collation) out to libraries.

PFX Components

PFX comprises two layers of functionality. The higher layer consists of two structured data parallelism APIs: PLINQ and the Parallel class. The lower layer contains the task parallelism classes — plus a set of additional constructs to help with parallel programming activities.

Parallel Programming Components

PLINQ offers the richest functionality: it automates all the steps of parallelization — including partitioning the work into tasks, executing those tasks on threads, and collating the results into a single output sequence. It’s called declarative — because you simply declare that you want to parallelize your work (which you structure as a LINQ query), and let the Framework take care of the implementation details. In contrast, the other approaches are imperative, in that you need to explicitly write code to partition or collate. In the case of the Parallel class, you must collate results yourself; with the task parallelism constructs, you must partition the work yourself, too:

Partitions work Collates results
PLINQ Yes Yes
The Parallel class Yes No
PFX’s task parallelism No No

The concurrent collections and spinning primitives help you with lower-level parallel programming activities. These are important because PFX has been designed to work not only with today’s hardware, but also with future generations of processors with far more cores. If you want to move a pile of chopped wood and you have 32 workers to do the job, the biggest challenge is moving the wood without the workers getting in each other's way. It’s the same with dividing an algorithm among 32 cores: if ordinary locks are used to protect common resources, the resultant blocking may mean that only a fraction of those cores are ever actually busy at once. The concurrent collections are tuned specifically for highly concurrent access, with the focus on minimizing or eliminating blocking. PLINQ and the Parallel class themselves rely on the concurrent collections and on spinning primitives for efficient management of work.

When to Use PFX

The primary use case for PFX is parallel programming: leveraging multicore processors to speed up computationally intensive code.

A challenge in leveraging multicores is Amdahl's law, which states that the maximum performance improvement from parallelization is governed by the portion of the code that must execute sequentially. For instance, if only two-thirds of an algorithm’s execution time is parallelizable, you can never exceed a threefold performance gain — even with an infinite number of cores.

So, before proceeding, it’s worth verifying that the bottleneck is in parallelizable code. It’s also worth considering whether your code needs to be computationally intensive — optimization is often the easiest and most effective approach. There’s a trade-off, though, in that some optimization techniques can make it harder to parallelize code.

The easiest gains come with what’s called embarrassingly parallel problems — where a job can be divided easily into tasks that execute efficiently on their own (structured parallelism is very well suited to such problems). Examples include many image processing tasks, ray tracing, and brute force approaches in mathematics or cryptography. An example of a nonembarrassingly parallel problem is implementing an optimized version of the quicksort algorithm — a good result takes some thought and may require unstructured parallelism.

PLINQ

PLINQ automatically parallelizes local LINQ queries. PLINQ has the advantage of being easy to use in that it offloads the burden of both work partitioning and result collation to the Framework.

To use PLINQ, simply call AsParallel() on the input sequence and then continue the LINQ query as usual. The following query calculates the prime numbers between 3 and 100,000 — making full use of all cores on the target machine:

// Calculate prime numbers using a simple (unoptimized) algorithm.
//
// NB: All code listings in this chapter are available as interactive code snippets in LINQPad.
// To activate these samples, click Download More Samples in LINQPad's Samples tab in the 
// bottom left, and select C# 4.0 in a Nutshell: More Chapters.
 
IEnumerable<int> numbers = Enumerable.Range (3, 100000-3);
 
var parallelQuery = 
  from n in numbers.AsParallel()
  where Enumerable.Range (2, (int) Math.Sqrt (n)).All (i => n % i > 0)
  select n;
 
int[] primes = parallelQuery.ToArray();

AsParallel is an extension method in System.Linq.ParallelEnumerable. It wraps the input in a sequence based on ParallelQuery<TSource>, which causes the LINQ query operators that you subsequently call to bind to an alternate set of extension methods defined in ParallelEnumerable. These provide parallel implementations of each of the standard query operators. Essentially, they work by partitioning the input sequence into chunks that execute on different threads, collating the results back into a single output sequence for consumption:

PLINQ Execution

Calling AsSequential() unwraps a ParallelQuery sequence so that subsequent query operators bind to the standard query operators and execute sequentially. This is necessary before calling methods that have side effects or are not thread-safe.

For query operators that accept two input sequences (Join, GroupJoin, Concat, Union, Intersect, Except, and Zip), you must apply AsParallel() to both input sequences (otherwise, an exception is thrown). You don’t, however, need to keep applying AsParallel to a query as it progresses, because PLINQ’s query operators output another ParallelQuery sequence. In fact, calling AsParallel again introduces inefficiency in that it forces merging and repartitioning of the query:

mySequence.AsParallel()           // Wraps sequence in ParallelQuery<int>
          .Where (n => n > 100)   // Outputs another ParallelQuery<int>
          .AsParallel()           // Unnecessary - and inefficient!
          .Select (n => n * n)

Not all query operators can be effectively parallelized. For those that cannot, PLINQ implements the operator sequentially instead. PLINQ may also operate sequentially if it suspects that the overhead of parallelization will actually slow a particular query.

PLINQ is only for local collections: it doesn’t work with LINQ to SQL or Entity Framework because in those cases the LINQ translates into SQL which then executes on a database server. However, you can use PLINQ to perform additional local querying on the result sets obtained from database queries.

If a PLINQ query throws an exception, it’s rethrown as an AggregateException whose InnerExceptions property contains the real exception (or exceptions). See Working with AggregateException for details.

Parallel Execution Ballistics

Like ordinary LINQ queries, PLINQ queries are lazily evaluated. This means that execution is triggered only when you begin consuming the results — typically via a foreach loop (although it may also be via a conversion operator such as ToArray or an operator that returns a single element or value).

As you enumerate the results, though, execution proceeds somewhat differently from that of an ordinary sequential query. A sequential query is powered entirely by the consumer in a “pull” fashion: each element from the input sequence is fetched exactly when required by the consumer. A parallel query ordinarily uses independent threads to fetch elements from the input sequence slightly ahead of when they’re needed by the consumer (rather like a teleprompter for newsreaders, or an antiskip buffer in CD players). It then processes the elements in parallel through the query chain, holding the results in a small buffer so that they’re ready for the consumer on demand. If the consumer pauses or breaks out of the enumeration early, the query processor also pauses or stops so as not to waste CPU time or memory.

You can tweak PLINQ’s buffering behavior by calling WithMergeOptions after AsParallel. The default value of AutoBuffered generally gives the best overall results. NotBuffered disables the buffer and is useful if you want to see results as soon as possible; FullyBuffered caches the entire result set before presenting it to the consumer (the OrderBy and Reverse operators naturally work this way, as do the element, aggregation, and conversion operators).

PLINQ and Ordering

A side effect of parallelizing the query operators is that when the results are collated, it’s not necessarily in the same order that they were submitted, as illustrated in the previous diagram. In other words, LINQ’s normal order-preservation guarantee for sequences no longer holds.

If you need order preservation, you can force it by calling AsOrdered() after AsParallel():

myCollection.AsParallel().AsOrdered()...

Calling AsOrdered incurs a performance hit with large numbers of elements because PLINQ must keep track of each element’s original position.

You can negate the effect of AsOrdered later in a query by calling AsUnordered: this introduces a “random shuffle point” which allows the query to execute more efficiently from that point on. So if you wanted to preserve input-sequence ordering for just the first two query operators, you’d do this:

inputSequence.AsParallel().AsOrdered()
  .QueryOperator1()
  .QueryOperator2()
  .AsUnordered()       // From here on, ordering doesn’t matter
  .QueryOperator3()
  ...

AsOrdered is not the default because for most queries, the original input ordering doesn’t matter. In other words, if AsOrdered was the default, you’d have to apply AsUnordered to the majority of your parallel queries to get the best performance, which would be burdensome.

PLINQ Limitations

There are currently some practical limitations on what PLINQ can parallelize. These limitations may loosen with subsequent service packs and Framework versions.

The following query operators prevent a query from being parallelized, unless the source elements are in their original indexing position:

Most query operators change the indexing position of elements (including those that remove elements, such as Where). This means that if you want to use the preceding operators, they’ll usually need to be at the start of the query.

The following query operators are parallelizable, but use an expensive partitioning strategy that can sometimes be slower than sequential processing:

The Aggregate operator’s seeded overloads in their standard incarnations are not parallelizable — PLINQ provides special overloads to deal with this.

All other operators are parallelizable, although use of these operators doesn’t guarantee that your query will be parallelized. PLINQ may run your query sequentially if it suspects that the overhead of parallelization will slow down that particular query. You can override this behavior and force parallelism by calling the following after AsParallel():

.WithExecutionMode (ParallelExecutionMode.ForceParallelism)

Example: Parallel Spellchecker

Suppose we want to write a spellchecker that runs quickly with very large documents by leveraging all available cores. By formulating our algorithm into a LINQ query, we can very easily parallelize it.

The first step is to download a dictionary of English words into a HashSet for efficient lookup:

if (!File.Exists ("WordLookup.txt"))    // Contains about 150,000 words
  new WebClient().DownloadFile (
    "http://www.albahari.com/ispell/allwords.txt", "WordLookup.txt");
 
var wordLookup = new HashSet<string> (
  File.ReadAllLines ("WordLookup.txt"),
  StringComparer.InvariantCultureIgnoreCase);

We’ll then use our word lookup to create a test “document” comprising an array of a million random words. After building the array, we’ll introduce a couple of spelling mistakes:

var random = new Random();
string[] wordList = wordLookup.ToArray();
 
string[] wordsToTest = Enumerable.Range (0, 1000000)
  .Select (i => wordList [random.Next (0, wordList.Length)])
  .ToArray();
 
wordsToTest [12345] = "woozsh";     // Introduce a couple
wordsToTest [23456] = "wubsie";     // of spelling mistakes.

Now we can perform our parallel spellcheck by testing wordsToTest against wordLookup. PLINQ makes this very easy:

var query = wordsToTest
  .AsParallel()
  .Select  ((word, index) => new IndexedWord { Word=word, Index=index })
  .Where   (iword => !wordLookup.Contains (iword.Word))
  .OrderBy (iword => iword.Index);
 
query.Dump();     // Display output in LINQPad

Here's the output, as displayed in LINQPad:

OrderedParallelQuery<IndexedWord> (2 items)
Word Index
woozsh 12345
wubsie 23456

IndexedWord is a custom struct that we define as follows:

struct IndexedWord { public string Word; public int Index; }

The wordLookup.Contains method in the predicate gives the query some “meat” and makes it worth parallelizing.

We could simplify the query slightly by using an anonymous type instead of the IndexedWord struct. However, this would degrade performance because anonymous types (being classes and therefore reference types) incur the cost of heap-based allocation and subsequent garbage collection.

The difference might not be enough to matter with sequential queries, but with parallel queries, favoring stack-based allocation can be quite advantageous. This is because stack-based allocation is highly parallelizable (as each thread has its own stack), whereas all threads must compete for the same heap — managed by a single memory manager and garbage collector.

Using ThreadLocal<T>

Let’s extend our example by parallelizing the creation of the random test-word list itself. We structured this as a LINQ query, so it should be easy. Here’s the sequential version:

string[] wordsToTest = Enumerable.Range (0, 1000000)
  .Select (i => wordList [random.Next (0, wordList.Length)])
  .ToArray();

Unfortunately, the call to random.Next is not thread-safe, so it’s not as simple as inserting AsParallel() into the query. A potential solution is to write a function that locks around random.Next; however, this would limit concurrency. The better option is to use ThreadLocal<Random> to create a separate Random object for each thread. We can then parallelize the query as follows:

var localRandom = new ThreadLocal<Random>
 ( () => new Random (Guid.NewGuid().GetHashCode()) );
 
string[] wordsToTest = Enumerable.Range (0, 1000000).AsParallel()
  .Select (i => wordList [localRandom.Value.Next (0, wordList.Length)])
  .ToArray();

In our factory function for instantiating a Random object, we pass in a Guid’s hashcode to ensure that if two Random objects are created within a short period of time, they’ll yield different random number sequences.

Functional Purity

Because PLINQ runs your query on parallel threads, you must be careful not to perform thread-unsafe operations. In particular, writing to variables is side-effecting and therefore thread-unsafe:

// The following query multiplies each element by its position.
// Given an input of Enumerable.Range(0,999), it should output squares.
int i = 0;
var query = from n in Enumerable.Range(0,999).AsParallel() select n * i++;

We could make incrementing i thread-safe by using locks or Interlocked, but the problem would still remain that i won’t necessarily correspond to the position of the input element. And adding AsOrdered to the query wouldn’t fix the latter problem, because AsOrdered ensures only that the elements are output in an order consistent with them having been processed sequentially — it doesn’t actually process them sequentially.

Instead, this query should be rewritten to use the indexed version of Select:

var query = Enumerable.Range(0,999).AsParallel().Select ((n, i) => n * i);

For best performance, any methods called from query operators should be thread-safe by virtue of not writing to fields or properties (non-side-effecting, or functionally pure). If they’re thread-safe by virtue of locking, the query’s parallelism potential will be limited — by the duration of the lock divided by the total time spent in that function.

Calling Blocking or I/O-Intensive Functions

Sometimes a query is long-running not because it’s CPU-intensive, but because it waits on something — such as a web page to download or some hardware to respond. PLINQ can effectively parallelize such queries, providing that you hint it by calling WithDegreeOfParallelism after AsParallel. For instance, suppose we want to ping six websites simultaneously. Rather than using clumsy asynchronous delegates or manually spinning up six threads, we can accomplish this effortlessly with a PLINQ query:

from site in new[]
{
  "www.albahari.com",
  "www.linqpad.net",
  "www.oreilly.com",
  "www.takeonit.com",
  "stackoverflow.com",
  "www.rebeccarey.com"  
}
.AsParallel().WithDegreeOfParallelism(6)
let p = new Ping().Send (site)
select new
{
  site,
  Result = p.Status,
  Time = p.RoundtripTime
}

WithDegreeOfParallelism forces PLINQ to run the specified number of tasks simultaneously. This is necessary when calling blocking functions such as Ping.Send because PLINQ otherwise assumes that the query is CPU-intensive and allocates tasks accordingly. On a two-core machine, for instance, PLINQ may default to running only two tasks at once, which is clearly undesirable in this situation.

PLINQ typically serves each task with a thread, subject to allocation by the thread pool. You can accelerate the initial ramping up of threads by calling ThreadPool.SetMinThreads.

To give another example, suppose we were writing a surveillance system and wanted to repeatedly combine images from four security cameras into a single composite image for display on a CCTV. We’ll represent a camera with the following class:

class Camera
{
  public readonly int CameraID;
  public Camera (int cameraID) { CameraID = cameraID; }
 
  // Get image from camera: return a simple string rather than an image
  public string GetNextFrame()
  {
    Thread.Sleep (123);       // Simulate time taken to get snapshot
    return "Frame from camera " + CameraID;
  }
}

To obtain a composite image, we must call GetNextFrame on each of four camera objects. Assuming the operation is I/O-bound, we can quadruple our frame rate with parallelization — even on a single-core machine. PLINQ makes this possible with minimal programming effort:

Camera[] cameras = Enumerable.Range (0, 4)    // Create 4 camera objects.
  .Select (i => new Camera (i))
  .ToArray();
 
while (true)
{
  string[] data = cameras
    .AsParallel().AsOrdered().WithDegreeOfParallelism (4)
    .Select (c => c.GetNextFrame()).ToArray();
 
  Console.WriteLine (string.Join (", ", data));   // Display data...
}

GetNextFrame is a blocking method, so we used WithDegreeOfParallelism to get the desired concurrency. In our example, the blocking happens when we call Sleep; in real life it would block because fetching an image from a camera is I/O- rather than CPU-intensive.

Calling AsOrdered ensures the images are displayed in a consistent order. Because there are only four elements in the sequence, this would have a negligible effect on performance.

Changing the degree of parallelism

You can call WithDegreeOfParallelism only once within a PLINQ query. If you need to call it again, you must force merging and repartitioning of the query by calling AsParallel() again within the query:

"The Quick Brown Fox"
  .AsParallel().WithDegreeOfParallelism (2)
  .Where (c => !char.IsWhiteSpace (c))
  .AsParallel().WithDegreeOfParallelism (3)   // Forces Merge + Partition
  .Select (c => char.ToUpper (c))

Cancellation

Canceling a PLINQ query whose results you’re consuming in a foreach loop is easy: simply break out of the foreach and the query will be automatically canceled as the enumerator is implicitly disposed.

For a query that terminates with a conversion, element, or aggregation operator, you can cancel it from another thread via a cancellation token. To insert a token, call WithCancellation after calling AsParallel, passing in the Token property of a CancellationTokenSource object. Another thread can then call Cancel on the token source, which throws an OperationCanceledException on the query’s consumer:

IEnumerable<int> million = Enumerable.Range (3, 1000000);
 
var cancelSource = new CancellationTokenSource(); 
var primeNumberQuery = 
  from n in million.AsParallel().WithCancellation (cancelSource.Token)
  where Enumerable.Range (2, (int) Math.Sqrt (n)).All (i => n % i > 0)
  select n;
 
new Thread (() => {
                    Thread.Sleep (100);      // Cancel query after
                    cancelSource.Cancel();   // 100 milliseconds.
                  }
           ).Start();
try 
{
  // Start query running:
  int[] primes = primeNumberQuery.ToArray();
  // We'll never get here because the other thread will cancel us.
}
catch (OperationCanceledException)
{
  Console.WriteLine ("Query canceled");
}

PLINQ doesn’t preemptively abort threads, because of the danger of doing so. Instead, upon cancellation it waits for each worker thread to finish with its current element before ending the query. This means that any external methods that the query calls will run to completion.

Optimizing PLINQ

Output-side optimization

One of PLINQ’s advantages is that it conveniently collates the results from parallelized work into a single output sequence. Sometimes, though, all that you end up doing with that sequence is running some function once over each element:

foreach (int n in parallelQuery)
  DoSomething (n);

If this is the case — and you don’t care about the order in which the elements are processed — you can improve efficiency with PLINQ’s ForAll method.

The ForAll method runs a delegate over every output element of a ParallelQuery. It hooks right into PLINQ’s internals, bypassing the steps of collating and enumerating the results. To give a trivial example:

"abcdef".AsParallel().Select (c => char.ToUpper(c)).ForAll (Console.Write);
PLINQ ForAll

Collating and enumerating results is not a massively expensive operation, so the ForAll optimization yields the greatest gains when there are large numbers of quickly executing input elements.

Input-side optimization

PLINQ has three partitioning strategies for assigning input elements to threads:

Strategy Element allocation Relative performance
Chunk partitioning Dynamic Average
Range partitioning Static Poor to excellent
Hash partitioning Static Poor

For query operators that require comparing elements (GroupBy, Join, GroupJoin, Intersect, Except, Union, and Distinct), you have no choice: PLINQ always uses hash partitioning. Hash partitioning is relatively inefficient in that it must precalculate the hashcode of every element (so that elements with identical hashcodes can be processed on the same thread). If you find this too slow, your only option is to call AsSequential to disable parallelization.

For all other query operators, you have a choice as to whether to use range or chunk partitioning. By default:

In a nutshell, range partitioning is faster with long sequences for which every element takes a similar amount of CPU time to process. Otherwise, chunk partitioning is usually faster.

To force range partitioning:

ParallelEnumerable.Range is not simply a shortcut for calling Enumerable.Range().AsParallel(). It changes the performance of the query by activating range partitioning.

To force chunk partitioning, wrap the input sequence in a call to Partitioner.Create (in System.Collection.Concurrent) as follows:

int[] numbers = { 3, 4, 5, 6, 7, 8, 9 };
var parallelQuery =
  Partitioner.Create (numbers, true).AsParallel()
  .Where (...)

The second argument to Partitioner.Create indicates that you want to load-balance the query, which is another way of saying that you want chunk partitioning.

Chunk partitioning works by having each worker thread periodically grab small “chunks” of elements from the input sequence to process. PLINQ starts by allocating very small chunks (one or two elements at a time), then increases the chunk size as the query progresses: this ensures that small sequences are effectively parallelized and large sequences don’t cause excessive round-tripping. If a worker happens to get “easy” elements (that process quickly) it will end up getting more chunks. This system keeps every thread equally busy (and the cores “balanced”); the only downside is that fetching elements from the shared input sequence requires synchronization (typically an exclusive lock) — and this can result in some overhead and contention.

Chunk vs Range Partitioning

Range partitioning bypasses the normal input-side enumeration and preallocates an equal number of elements to each worker, avoiding contention on the input sequence. But if some threads happen to get easy elements and finish early, they sit idle while the remaining threads continue working. Our earlier prime number calculator might perform poorly with range partitioning. An example of when range partitioning would do well is in calculating the sum of the square roots of the first 10 million integers:

ParallelEnumerable.Range (1, 10000000).Sum (i => Math.Sqrt (i))

ParallelEnumerable.Range returns a ParallelQuery<T>, so you don’t need to subsequently call AsParallel.

Range partitioning doesn’t necessarily allocate element ranges in contiguous blocks — it might instead choose a “striping” strategy. For instance, if there are two workers, one worker might process odd-numbered elements while the other processes even-numbered elements. The TakeWhile operator is almost certain to trigger a striping strategy to avoid unnecessarily processing elements later in the sequence.

Parallelizing Custom Aggregations

PLINQ parallelizes the Sum, Average, Min, and Max operators efficiently without additional intervention. The Aggregate operator, though, presents special challenges for PLINQ.

If you’re unfamiliar with this operator, you can think of Aggregate as a generalized version of Sum, Average, Min, and Max — in other words, an operator that lets you plug in a custom accumulation algorithm for implementing unusual aggregations. The following demonstrates how Aggregate can do the work of Sum:

int[] numbers = { 2, 3, 4 };
int sum = numbers.Aggregate (0, (total, n) => total + n);   // 9

The first argument to Aggregate is the seed, from which accumulation starts. The second argument is an expression to update the accumulated value, given a fresh element. You can optionally supply a third argument to project the final result value from the accumulated value.

Most problems for which Aggregate has been designed can be solved as easily with a foreach loop — and with more familiar syntax. The advantage of Aggregate is precisely that large or complex aggregations can be parallelized declaratively with PLINQ.

Unseeded aggregations

You can omit the seed value when calling Aggregate, in which case the first element becomes the implicit seed, and aggregation proceeds from the second element. Here’s the preceding example, unseeded:

int[] numbers = { 1, 2, 3 };
int sum = numbers.Aggregate ((total, n) => total + n);   // 6

This gives the same result as before, but we’re actually doing a different calculation. Before, we were calculating 0+1+2+3; now we’re calculating 1+2+3. We can better illustrate the difference by multiplying instead of adding:

int[] numbers = { 1, 2, 3 };
int x = numbers.Aggregate (0, (prod, n) => prod * n);   // 0*1*2*3 = 0
int y = numbers.Aggregate (   (prod, n) => prod * n);   //   1*2*3 = 6

As we’ll see shortly, unseeded aggregations have the advantage of being parallelizable without requiring the use of special overloads. However, there is a trap with unseeded aggregations: the unseeded aggregation methods are intended for use with delegates that are commutative and associative. If used otherwise, the result is either unintuitive (with ordinary queries) or nondeterministic (in the case that you parallelize the query with PLINQ). For example, consider the following function:

(total, n) => total + n * n

This is neither commutative nor associative. (For example, 1+2*2 != 2+1*1). Let’s see what happens when we use it to sum the square of the numbers 2, 3, and 4:

int[] numbers = { 2, 3, 4 };
int sum = numbers.Aggregate ((total, n) => total + n * n);    // 27

Instead of calculating:

2*2 + 3*3 + 4*4    // 29

it calculates:

2 + 3*3 + 4*4      // 27

We can fix this in a number of ways. First, we could include 0 as the first element:

int[] numbers = { 0, 2, 3, 4 };

Not only is this inelegant, but it will still give incorrect results if parallelized — because PLINQ leverages the function’s assumed associativity by selecting multiple elements as seeds. To illustrate, if we denote our aggregation function as follows:

f(total, n) => total + n * n

then LINQ to Objects would calculate this:

f(f(f(0, 2),3),4)

whereas PLINQ may do this:

f(f(0,2),f(3,4))

with the following result:

First partition:   a = 0 + 2*2  (= 4)
Second partition:  b = 3 + 4*4  (= 19)
Final result:          a + b*b  (= 365)
OR EVEN:               b + a*a  (= 35)  

There are two good solutions. The first is to turn this into a seeded aggregation — with zero as the seed. The only complication is that with PLINQ, we’d need to use a special overload in order for the query not to execute sequentially (as we’ll see soon).

The second solution is to restructure the query such that the aggregation function is commutative and associative:

int sum = numbers.Select (n => n * n).Aggregate ((total, n) => total + n);

Of course, in such simple scenarios you can (and should) use the Sum operator instead of Aggregate:

int sum = numbers.Sum (n => n * n);

You can actually go quite far just with Sum and Average. For instance, you can use Average to calculate a root-mean-square:

Math.Sqrt (numbers.Average (n => n * n))

and even standard deviation:

double mean = numbers.Average();
double sdev = Math.Sqrt (numbers.Average (n =>
              {
                double dif = n - mean;
                return dif * dif;
              }));

Both are safe, efficient and fully parallelizable.

Parallelizing Aggregate

We just saw that for unseeded aggregations, the supplied delegate must be associative and commutative. PLINQ will give incorrect results if this rule is violated, because it draws multiple seeds from the input sequence in order to aggregate several partitions of the sequence simultaneously.

Explicitly seeded aggregations might seem like a safe option with PLINQ, but unfortunately these ordinarily execute sequentially because of the reliance on a single seed. To mitigate this, PLINQ provides another overload of Aggregate that lets you specify multiple seeds — or rather, a seed factory function. For each thread, it executes this function to generate a separate seed, which becomes a thread-local accumulator into which it locally aggregates elements.

You must also supply a function to indicate how to combine the local and main accumulators. Finally, this Aggregate overload (somewhat gratuitously) expects a delegate to perform any final transformation on the result (you can achieve this as easily by running some function on the result yourself afterward). So, here are the four delegates, in the order they are passed:

seedFactory
Returns a new local accumulator
updateAccumulatorFunc
Aggregates an element into a local accumulator
combineAccumulatorFunc
Combines a local accumulator with the main accumulator
resultSelector
Applies any final transformation on the end result

In simple scenarios, you can specify a seed value instead of a seed factory. This tactic fails when the seed is a reference type that you wish to mutate, because the same instance will then be shared by each thread.

To give a very simple example, the following sums the values in a numbers array:

numbers.AsParallel().Aggregate (
  () => 0,                                     // seedFactory
  (localTotal, n) => localTotal + n,           // updateAccumulatorFunc
  (mainTot, localTot) => mainTot + localTot,   // combineAccumulatorFunc
  finalResult => finalResult)                  // resultSelector

This example is contrived in that we could get the same answer just as efficiently using simpler approaches (such as an unseeded aggregate, or better, the Sum operator). To give a more realistic example, suppose we wanted to calculate the frequency of each letter in the English alphabet in a given string. A simple sequential solution might look like this:

string text = "Let’s suppose this is a really long string";
var letterFrequencies = new int[26];
foreach (char c in text)
{
  int index = char.ToUpper (c) - 'A';
  if (index >= 0 && index <= 26) letterFrequencies [index]++;
};

An example of when the input text might be very long is in gene sequencing. The “alphabet” would then consist of the letters a, c, g, and t.

To parallelize this, we could replace the foreach statement with a call to Parallel.ForEach (as we’ll cover in the following section), but this will leave us to deal with concurrency issues on the shared array. And locking around accessing that array would all but kill the potential for parallelization.

Aggregate offers a tidy solution. The accumulator, in this case, is an array just like the letterFrequencies array in our preceding example. Here’s a sequential version using Aggregate:

int[] result =
  text.Aggregate (
    new int[26],                // Create the "accumulator"
    (letterFrequencies, c) =>   // Aggregate a letter into the accumulator
    {
      int index = char.ToUpper (c) - 'A';
      if (index >= 0 && index <= 26) letterFrequencies [index]++;
      return letterFrequencies;
    });

And now the parallel version, using PLINQ’s special overload:

int[] result =
  text.AsParallel().Aggregate (
    () => new int[26],             // Create a new local accumulator
 
    (localFrequencies, c) =>       // Aggregate into the local accumulator
    {
      int index = char.ToUpper (c) - 'A';
      if (index >= 0 && index <= 26) localFrequencies [index]++;
      return localFrequencies;
    },
                                   // Aggregate local->main accumulator
    (mainFreq, localFreq) =>
      mainFreq.Zip (localFreq, (f1, f2) => f1 + f2).ToArray(),
 
    finalResult => finalResult     // Perform any final transformation
  );                               // on the end result.

Notice that the local accumulation function mutates the localFrequencies array. This ability to perform this optimization is important — and is legitimate because localFrequencies is local to each thread.

The Parallel Class

PFX provides a basic form of structured parallelism via three static methods in the Parallel class:

Parallel.Invoke
Executes an array of delegates in parallel
Parallel.For
Performs the parallel equivalent of a C# for loop
Parallel.ForEach
Performs the parallel equivalent of a C# foreach loop

All three methods block until all work is complete. As with PLINQ, after an unhandled exception, remaining workers are stopped after their current iteration and the exception (or exceptions) are thrown back to the caller — wrapped in an AggregateException.

Parallel.Invoke

Parallel.Invoke executes an array of Action delegates in parallel, and then waits for them to complete. The simplest version of the method is defined as follows:

public static void Invoke (params Action[] actions);

Here’s how we can use Parallel.Invoke to download two web pages at once:

Parallel.Invoke (
 () => new WebClient().DownloadFile ("http://www.linqpad.net", "lp.html"),
 () => new WebClient().DownloadFile ("http://www.jaoo.dk", "jaoo.html"));

On the surface, this seems like a convenient shortcut for creating and waiting on two Task objects (or asynchronous delegates). But there’s an important difference: Parallel.Invoke still works efficiently if you pass in an array of a million delegates. This is because it partitions large numbers of elements into batches which it assigns to a handful of underlying Tasks — rather than creating a separate Task for each delegate.

As with all of Parallel’s methods, you’re on your own when it comes to collating the results. This means you need to keep thread safety in mind. The following, for instance, is thread-unsafe:

var data = new List<string>();
Parallel.Invoke (
 () => data.Add (new WebClient().DownloadString ("http://www.foo.com")),
 () => data.Add (new WebClient().DownloadString ("http://www.far.com")));

Locking around adding to the list would resolve this, although locking would create a bottleneck if you had a much larger array of quickly executing delegates. A better solution is to use a thread-safe collection such as ConcurrentBag would be ideal in this case.

Parallel.Invoke is also overloaded to accept a ParallelOptions object:

public static void Invoke (ParallelOptions options,
                           params Action[] actions);

With ParallelOptions, you can insert a cancellation token, limit the maximum concurrency, and specify a custom task scheduler. A cancellation token is relevant when you’re executing (roughly) more tasks than you have cores: upon cancellation, any unstarted delegates will be abandoned. Any already-executing delegates will, however, continue to completion. See Cancellation for an example of how to use cancellation tokens.

Parallel.For and Parallel.ForEach

Parallel.For and Parallel.ForEach perform the equivalent of a C# for and foreach loop, but with each iteration executing in parallel instead of sequentially. Here are their (simplest) signatures:

public static ParallelLoopResult For (
  int fromInclusive, int toExclusive, Action<int> body)
 
public static ParallelLoopResult ForEach<TSource> (
  IEnumerable<TSource> source, Action<TSource> body)

The following sequential for loop:

for (int i = 0; i < 100; i++)
  Foo (i);

is parallelized like this:

Parallel.For (0, 100, i => Foo (i));

or more simply:

Parallel.For (0, 100, Foo);

And the following sequential foreach:

foreach (char c in "Hello, world")
  Foo (c);

is parallelized like this:

Parallel.ForEach ("Hello, world", Foo);

To give a practical example, if we import the System.Security.Cryptography namespace, we can generate six public/private key-pair strings in parallel as follows:

var keyPairs = new string[6];
 
Parallel.For (0, keyPairs.Length,
              i => keyPairs[i] = RSA.Create().ToXmlString (true));

As with Parallel.Invoke, we can feed Parallel.For and Parallel.ForEach a large number of work items and they’ll be efficiently partitioned onto a few tasks.

The latter query could also be done with PLINQ:

string[] keyPairs =
  ParallelEnumerable.Range (0, 6)
  .Select (i => RSA.Create().ToXmlString (true))
  .ToArray();

Outer versus inner loops

Parallel.For and Parallel.ForEach usually work best on outer rather than inner loops. This is because with the former, you’re offering larger chunks of work to parallelize, diluting the management overhead. Parallelizing both inner and outer loops is usually unnecessary. In the following example, we’d typically need more than 100 cores to benefit from the inner parallelization:

Parallel.For (0, 100, i =>
{
  Parallel.For (0, 50, j => Foo (i, j));   // Sequential would be better
});                                        // for the inner loop.

Indexed Parallel.ForEach

Sometimes it’s useful to know the loop iteration index. With a sequential foreach, it’s easy:

int i = 0;
foreach (char c in "Hello, world")
  Console.WriteLine (c.ToString() + i++);

Incrementing a shared variable, however, is not thread-safe in a parallel context. You must instead use the following version of ForEach:

public static ParallelLoopResult ForEach<TSource> (
  IEnumerable<TSource> source, Action<TSource,ParallelLoopState,long> body)

We’ll ignore ParallelLoopState (which we’ll cover in the following section). For now, we’re interested in Action’s third type parameter of type long, which indicates the loop index:

Parallel.ForEach ("Hello, world", (c, state, i) =>
{
   Console.WriteLine (c.ToString() + i);
});

To put this into a practical context, we’ll revisit the spellchecker that we wrote with PLINQ. The following code loads up a dictionary along with an array of a million words to test:

if (!File.Exists ("WordLookup.txt"))    // Contains about 150,000 words
  new WebClient().DownloadFile (
    "http://www.albahari.com/ispell/allwords.txt", "WordLookup.txt");
 
var wordLookup = new HashSet<string> (
  File.ReadAllLines ("WordLookup.txt"),
  StringComparer.InvariantCultureIgnoreCase);
 
var random = new Random();
string[] wordList = wordLookup.ToArray();
 
string[] wordsToTest = Enumerable.Range (0, 1000000)
  .Select (i => wordList [random.Next (0, wordList.Length)])
  .ToArray();
 
wordsToTest [12345] = "woozsh";     // Introduce a couple
wordsToTest [23456] = "wubsie";     // of spelling mistakes.

We can perform the spellcheck on our wordsToTest array using the indexed version of Parallel.ForEach as follows:

var misspellings = new ConcurrentBag<Tuple<int,string>>();
 
Parallel.ForEach (wordsToTest, (word, state, i) =>
{
  if (!wordLookup.Contains (word))
    misspellings.Add (Tuple.Create ((int) i, word));
});

Notice that we had to collate the results into a thread-safe collection: having to do this is the disadvantage when compared to using PLINQ. The advantage over PLINQ is that we avoid the cost of applying an indexed Select query operator — which is less efficient than an indexed ForEach.

ParallelLoopState: Breaking early out of loops

Because the loop body in a parallel For or ForEach is a delegate, you can’t exit the loop early with a break statement. Instead, you must call Break or Stop on a ParallelLoopState object:

public class ParallelLoopState
{
  public void Break();
  public void Stop();
 
  public bool IsExceptional { get; }
  public bool IsStopped { get; }
  public long? LowestBreakIteration { get; }
  public bool ShouldExitCurrentIteration { get; }
}

Obtaining a ParallelLoopState is easy: all versions of For and ForEach are overloaded to accept loop bodies of type Action<TSource,ParallelLoopState>. So, to parallelize this:

foreach (char c in "Hello, world")
  if (c == ',')
    break;
  else
    Console.Write (c);

do this:

Parallel.ForEach ("Hello, world", (c, loopState) =>
{
  if (c == ',')
    loopState.Break();
  else
    Console.Write (c);
});
Hlloe

You can see from the output that loop bodies may complete in a random order. Aside from this difference, calling Break yields at least the same elements as executing the loop sequentially: this example will always output at least the letters H, e, l, l, and o in some order. In contrast, calling Stop instead of Break forces all threads to finish right after their current iteration. In our example, calling Stop could give us a subset of the letters H, e, l, l, and o if another thread was lagging behind. Calling Stop is useful when you’ve found something that you’re looking for — or when something has gone wrong and you won’t be looking at the results.

The Parallel.For and Parallel.ForEach methods return a ParallelLoopResult object that exposes properties called IsCompleted and LowestBreakIteration. These tell you whether the loop ran to completion, and if not, at what cycle the loop was broken.

If LowestBreakIteration returns null, it means that you called Stop (rather than Break) on the loop.

If your loop body is long, you might want other threads to break partway through the method body in case of an early Break or Stop. You can do this by polling the ShouldExitCurrentIteration property at various places in your code; this property becomes true immediately after a Stop — or soon after a Break.

ShouldExitCurrentIteration also becomes true after a cancellation request — or if an exception is thrown in the loop.

IsExceptional lets you know whether an exception has occurred on another thread. Any unhandled exception will cause the loop to stop after each thread’s current iteration: to avoid this, you must explicitly handle exceptions in your code.

Optimization with local values

Parallel.For and Parallel.ForEach each offer a set of overloads that feature a generic type argument called TLocal. These overloads are designed to help you optimize the collation of data with iteration-intensive loops. The simplest is this:

public static ParallelLoopResult For <TLocal> (
  int fromInclusive,
  int toExclusive,
  Func <TLocal> localInit,  Func <int, ParallelLoopState, TLocal, TLocal> body,
  Action <TLocal> localFinally);

These methods are rarely needed in practice because their target scenarios are covered mostly by PLINQ (which is fortunate because these overloads are somewhat intimidating!).

Essentially, the problem is this: suppose we want to sum the square roots of the numbers 1 through 10,000,000. Calculating 10 million square roots is easily parallelizable, but summing their values is troublesome because we must lock around updating the total:

object locker = new object();
double total = 0;
Parallel.For (1, 10000000,
              i => { lock (locker) total += Math.Sqrt (i); });

The gain from parallelization is more than offset by the cost of obtaining 10 million locks — plus the resultant blocking.

The reality, though, is that we don’t actually need 10 million locks. Imagine a team of volunteers picking up a large volume of litter. If all workers shared a single trash can, the travel and contention would make the process extremely inefficient. The obvious solution is for each worker to have a private or “local” trash can, which is occasionally emptied into the main bin.

The TLocal versions of For and ForEach work in exactly this way. The volunteers are internal worker threads, and the local value represents a local trash can. In order for Parallel to do this job, you must feed it two additional delegates that indicate:

  1. How to initialize a new local value
  2. How to combine a local aggregation with the master value

Additionally, instead of the body delegate returning void, it should return the new aggregate for the local value. Here’s our example refactored:

object locker = new object();
double grandTotal = 0;
 
Parallel.For (1, 10000000,
 
  () => 0.0,                        // Initialize the local value.
 
  (i, state, localTotal) =>         // Body delegate. Notice that it
     localTotal + Math.Sqrt (i),    // returns the new local total.
 
  localTotal =>                                    // Add the local value
    { lock (locker) grandTotal += localTotal; }    // to the master value.
);

We must still lock, but only around aggregating the local value to the grand total. This makes the process dramatically more efficient.

As stated earlier, PLINQ is often a good fit in these scenarios. Our example could be parallelized with PLINQ simply like this:

ParallelEnumerable.Range(1, 10000000)
                  .Sum (i => Math.Sqrt (i))

(Notice that we used ParallelEnumerable to force range partitioning: this improves performance in this case because all numbers will take equally long to process.)

In more complex scenarios, you might use LINQ’s Aggregate operator instead of Sum. If you supplied a local seed factory, the situation would be somewhat analogous to providing a local value function with Parallel.For.

Task Parallelism

Task parallelism is the lowest-level approach to parallelization with PFX. The classes for working at this level are defined in the System.Threading.Tasks namespace and comprise the following:

Class Purpose
Task For managing a unit for work
Task<TResult> For managing a unit for work with a return value
TaskFactory For creating tasks
TaskFactory<TResult> For creating tasks and continuations with the same return type
TaskScheduler For managing the scheduling of tasks
TaskCompletionSource For manually controlling a task’s workflow

Essentially, a task is a lightweight object for managing a parallelizable unit of work. A task avoids the overhead of starting a dedicated thread by using the CLR’s thread pool: this is the same thread pool used by ThreadPool.QueueUserWorkItem, tweaked in CLR 4.0 to work more efficiently with Tasks (and more efficiently in general).

Tasks can be used whenever you want to execute something in parallel. However, they’re tuned for leveraging multicores: in fact, the Parallel class and PLINQ are internally built on the task parallelism constructs.

Tasks do more than just provide an easy and efficient way into the thread pool. They also provide some powerful features for managing units of work, including the ability to:

Tasks also implement local work queues, an optimization that allows you to efficiently create many quickly executing child tasks without incurring the contention overhead that would otherwise arise with a single work queue.

The Task Parallel Library lets you create hundreds (or even thousands) of tasks with minimal overhead. But if you want to create millions of tasks, you’ll need to partition those tasks into larger work units to maintain efficiency. The Parallel class and PLINQ do this automatically.

Visual Studio 2010 provides a new window for monitoring tasks (Debug | Window | Parallel Tasks). This is equivalent to the Threads window, but for tasks. The Parallel Stacks window also has a special mode for tasks.

Creating and Starting Tasks

As we described in Part 1 in our discussion of thread pooling, you can create and start a Task by calling Task.Factory.StartNew, passing in an Action delegate:

Task.Factory.StartNew (() => Console.WriteLine ("Hello from a task!"));

The generic version, Task<TResult> (a subclass of Task), lets you get data back from a task upon completion:

Task<string> task = Task.Factory.StartNew<string> (() =>    // Begin task
{
  using (var wc = new System.Net.WebClient())
    return wc.DownloadString ("http://www.linqpad.net");
});
 
RunSomeOtherMethod();         // We can do other work in parallel...
 
string result = task.Result;  // Wait for task to finish and fetch result.

Task.Factory.StartNew creates and starts a task in one step. You can decouple these operations by first instantiating a Task object, and then calling Start:

var task = new Task (() => Console.Write ("Hello"));
...
task.Start();

A task that you create in this manner can also be run synchronously (on the same thread) by calling RunSynchronously instead of Start.

You can track a task’s execution status via its Status property.

Specifying a state object

When instantiating a task or calling Task.Factory.StartNew, you can specify a state object, which is passed to the target method. This is useful should you want to call a method directly rather than using a lambda expression:

static void Main()
{
  var task = Task.Factory.StartNew (Greet, "Hello");
  task.Wait();  // Wait for task to complete.
}
 
static void Greet (object state) { Console.Write (state); }   // Hello

Given that we have lambda expressions in C#, we can put the state object to better use, which is to assign a meaningful name to the task. We can then use the AsyncState property to query its name:

static void Main()
{
  var task = Task.Factory.StartNew (state => Greet ("Hello"), "Greeting");
  Console.WriteLine (task.AsyncState);   // Greeting
  task.Wait();
}
 
static void Greet (string message) { Console.Write (message); }

Visual Studio displays each task’s AsyncState in the Parallel Tasks window, so having a meaningful name here can ease debugging considerably.

TaskCreationOptions

You can tune a task’s execution by specifying a TaskCreationOptions enum when calling StartNew (or instantiating a Task). TaskCreationOptions is a flags enum with the following (combinable) values:

LongRunning
PreferFairness
AttachedToParent

LongRunning suggests to the scheduler to dedicate a thread to the task. This is beneficial for long-running tasks because they might otherwise “hog” the queue, and force short-running tasks to wait an unreasonable amount of time before being scheduled. LongRunning is also good for blocking tasks.

The task queuing problem arises because the task scheduler ordinarily tries to keep just enough tasks active on threads at once to keep each CPU core busy. Not oversubscribing the CPU with too many active threads avoids the degradation in performance that would occur if the operating system was forced to perform a lot of expensive time slicing and context switching.

PreferFairness tells the scheduler to try to ensure that tasks are scheduled in the order they were started. It may ordinarily do otherwise, because it internally optimizes the scheduling of tasks using local work-stealing queues. This optimization is of practical benefit with very small (fine-grained) tasks.

AttachedToParent is for creating child tasks.

Child tasks

When one task starts another, you can optionally establish a parent-child relationship by specifying TaskCreationOptions.AttachedToParent:

Task parent = Task.Factory.StartNew (() =>
{
  Console.WriteLine ("I am a parent");
 
  Task.Factory.StartNew (() =>        // Detached task
  {
    Console.WriteLine ("I am detached");
  });
 
  Task.Factory.StartNew (() =>        // Child task
  {
    Console.WriteLine ("I am a child");
  }, TaskCreationOptions.AttachedToParent);
});

A child task is special in that when you wait for the parent task to complete, it waits for any children as well. This can be particularly useful when a child task is a continuation, as we’ll see shortly.

Waiting on Tasks

You can explicitly wait for a task to complete in two ways:

You can also wait on multiple tasks at once — via the static methods Task.WaitAll (waits for all the specified tasks to finish) and Task.WaitAny (waits for just one task to finish).

WaitAll is similar to waiting out each task in turn, but is more efficient in that it requires (at most) just one context switch. Also, if one or more of the tasks throw an unhandled exception, WaitAll still waits out every task — and then rethrows a single AggregateException that accumulates the exceptions from each faulted task. It’s equivalent to doing this:

// Assume t1, t2 and t3 are tasks:
var exceptions = new List<Exception>();
try { t1.Wait(); } catch (AggregateException ex) { exceptions.Add (ex); }
try { t2.Wait(); } catch (AggregateException ex) { exceptions.Add (ex); }
try { t3.Wait(); } catch (AggregateException ex) { exceptions.Add (ex); }
if (exceptions.Count > 0) throw new AggregateException (exceptions);

Calling WaitAny is equivalent to waiting on a ManualResetEventSlim that’s signaled by each task as it finishes.

As well as a timeout, you can also pass in a cancellation token to the Wait methods: this lets you cancel the wait — not the task itself.

Exception-Handling Tasks

When you wait for a task to complete (by calling its Wait method or accessing its Result property), any unhandled exceptions are conveniently rethrown to the caller, wrapped in an AggregateException object. This usually avoids the need to write code within task blocks to handle unexpected exceptions; instead we can do this:

int x = 0;
Task<int> calc = Task.Factory.StartNew (() => 7 / x);
try
{
  Console.WriteLine (calc.Result);
}
catch (AggregateException aex)
{
  Console.Write (aex.InnerException.Message);  // Attempted to divide by 0
}

You still need to exception-handle detached autonomous tasks (unparented tasks that are not waited upon) in order to prevent an unhandled exception taking down the application when the task drops out of scope and is garbage-collected (subject to the following note). The same applies for tasks waited upon with a timeout, because any exception thrown after the timeout interval will otherwise be “unhandled.”

The static TaskScheduler.UnobservedTaskException event provides a final last resort for dealing with unhandled task exceptions. By handling this event, you can intercept task exceptions that would otherwise end the application — and provide your own logic for dealing with them.

For parented tasks, waiting on the parent implicitly waits on the children — and any child exceptions then bubble up:

TaskCreationOptions atp = TaskCreationOptions.AttachedToParent;
var parent = Task.Factory.StartNew (() => 
{
  Task.Factory.StartNew (() =>   // Child
  {
    Task.Factory.StartNew (() => { throw null; }, atp);   // Grandchild
  }, atp);
});
 
// The following call throws a NullReferenceException (wrapped
// in nested AggregateExceptions):
parent.Wait();

Interestingly, if you check a task’s Exception property after it has thrown an exception, the act of reading that property will prevent the exception from subsequently taking down your application. The rationale is that PFX’s designers don’t want you ignoring exceptions — as long as you acknowledge them in some way, they won’t punish you by terminating your program.

An unhandled exception on a task doesn’t cause immediate application termination: instead, it’s delayed until the garbage collector catches up with the task and calls its finalizer. Termination is delayed because it can’t be known for certain that you don’t plan to call Wait or check its Result or Exception property until the task is garbage-collected. This delay can sometimes mislead you as to the original source of the error (although Visual Studio’s debugger can assist if you enable breaking on first-chance exceptions).

As we’ll see soon, an alternative strategy for dealing with exceptions is with continuations.

Canceling Tasks

You can optionally pass in a cancellation token when starting a task. This lets you cancel tasks via the cooperative cancellation pattern described previously:

var cancelSource = new CancellationTokenSource();
CancellationToken token = cancelSource.Token;
 
Task task = Task.Factory.StartNew (() => 
{
  // Do some stuff...
  token.ThrowIfCancellationRequested();  // Check for cancellation request
  // Do some stuff...
}, token);
...
cancelSource.Cancel();

To detect a canceled task, catch an AggregateException and check the inner exception as follows:

try 
{
  task.Wait();
}
catch (AggregateException ex)
{
  if (ex.InnerException is OperationCanceledException)
    Console.Write ("Task canceled!");
}

If you want to explicitly throw an OperationCanceledException (rather than calling token.ThrowIfCancellationRequested), you must pass the cancellation token into OperationCanceledException’s constructor. If you fail to do this, the task won’t end up with a TaskStatus.Canceled status and won’t trigger OnlyOnCanceled continuations.

If the task is canceled before it has started, it won’t get scheduled — an OperationCanceledException will instead be thrown on the task immediately.

Because cancellation tokens are recognized by other APIs, you can pass them into other constructs and cancellations will propagate seamlessly:

var cancelSource = new CancellationTokenSource();
CancellationToken token = cancelSource.Token;
 
Task task = Task.Factory.StartNew (() =>
{
  // Pass our cancellation token into a PLINQ query:
  var query = someSequence.AsParallel().WithCancellation (token)...
  ... enumerate query ...
});

Calling Cancel on cancelSource in this example will cancel the PLINQ query, which will throw an OperationCanceledException on the task body, which will then cancel the task.

The cancellation tokens that you can pass into methods such as Wait and CancelAndWait allow you to cancel the wait operation and not the task itself.

Continuations

Sometimes it’s useful to start a task right after another one completes (or fails). The ContinueWith method on the Task class does exactly this:

Task task1 = Task.Factory.StartNew (() => Console.Write ("antecedant.."));
Task task2 = task1.ContinueWith (ant => Console.Write ("..continuation"));

As soon as task1 (the antecedent) finishes, fails, or is canceled, task2 (the continuation) automatically starts. (If task1 had completed before the second line of code ran, task2 would be scheduled to execute right away.) The ant argument passed to the continuation’s lambda expression is a reference to the antecedent task.

Our example demonstrated the simplest kind of continuation, and is functionally similar to the following:

Task task = Task.Factory.StartNew (() =>
{
  Console.Write ("antecedent..");
  Console.Write ("..continuation");
});

The continuation-based approach, however, is more flexible in that you could first wait on task1, and then later wait on task2. This is particularly useful if task1 returns data.

Another (subtler) difference is that by default, antecedent and continuation tasks may execute on different threads. You can force them to execute on the same thread by specifying TaskContinuationOptions.ExecuteSynchronously when calling ContinueWith: this can improve performance in very fine-grained continuations by lessening indirection.

Continuations and Task<TResult>

Just like ordinary tasks, continuations can be of type Task<TResult> and return data. In the following example, we calculate Math.Sqrt(8*2) using a series of chained tasks and then write out the result:

Task.Factory.StartNew<int> (() => 8)
  .ContinueWith (ant => ant.Result * 2)
  .ContinueWith (ant => Math.Sqrt (ant.Result))
  .ContinueWith (ant => Console.WriteLine (ant.Result));   // 4

Our example is somewhat contrived for simplicity; in real life, these lambda expressions would call computationally intensive functions.

Continuations and exceptions

A continuation can find out if an exception was thrown by the antecedent via the antecedent task’s Exception property. The following writes the details of a NullReferenceException to the console:

Task task1 = Task.Factory.StartNew (() => { throw null; });
Task task2 = task1.ContinueWith (ant => Console.Write (ant.Exception));

If an antecedent throws and the continuation fails to query the antecedent’s Exception property (and the antecedent isn’t otherwise waited upon), the exception is considered unhandled and the application dies (unless handled by TaskScheduler.UnobservedTaskException).

A safe pattern is to rethrow antecedent exceptions. As long as the continuation is Waited upon, the exception will be propagated and rethrown to the Waiter:

Task continuation = Task.Factory.StartNew     (()  => { throw null; })
                                .ContinueWith (ant =>
  {
    if (ant.Exception != null) throw ant.Exception;    // Continue processing...
  });
 
continuation.Wait();    // Exception is now thrown back to caller.

Another way to deal with exceptions is to specify different continuations for exceptional versus nonexceptional outcomes. This is done with TaskContinuationOptions:

Task task1 = Task.Factory.StartNew (() => { throw null; });
 
Task error = task1.ContinueWith (ant => Console.Write (ant.Exception),
                                 TaskContinuationOptions.OnlyOnFaulted);
 
Task ok = task1.ContinueWith (ant => Console.Write ("Success!"),
                              TaskContinuationOptions.NotOnFaulted);

This pattern is particularly useful in conjunction with child tasks, as we’ll see very soon.

The following extension method “swallows” a task’s unhandled exceptions:

public static void IgnoreExceptions (this Task task)
{
  task.ContinueWith (t => { var ignore = t.Exception; },
    TaskContinuationOptions.OnlyOnFaulted);
} 

(This could be improved by adding code to log the exception.) Here’s how it would be used:

Task.Factory.StartNew (() => { throw null; }).IgnoreExceptions();

Continuations and child tasks

A powerful feature of continuations is that they kick off only when all child tasks have completed. At that point, any exceptions thrown by the children are marshaled to the continuation.

In the following example, we start three child tasks, each throwing a NullReferenceException. We then catch all of them in one fell swoop via a continuation on the parent:

TaskCreationOptions atp = TaskCreationOptions.AttachedToParent;
Task.Factory.StartNew (() =>
{
  Task.Factory.StartNew (() => { throw null; }, atp);
  Task.Factory.StartNew (() => { throw null; }, atp);
  Task.Factory.StartNew (() => { throw null; }, atp);
})
.ContinueWith (p => Console.WriteLine (p.Exception),
                    TaskContinuationOptions.OnlyOnFaulted);
Continuations

Conditional continuations

By default, a continuation is scheduled unconditionally — whether the antecedent completes, throws an exception, or is canceled. You can alter this behavior via a set of (combinable) flags included within the TaskContinuationOptions enum. The three core flags that control conditional continuation are:

NotOnRanToCompletion = 0x10000,
NotOnFaulted = 0x20000,
NotOnCanceled = 0x40000,

These flags are subtractive in the sense that the more you apply, the less likely the continuation is to execute. For convenience, there are also the following precombined values:

OnlyOnRanToCompletion = NotOnFaulted | NotOnCanceled,
OnlyOnFaulted = NotOnRanToCompletion | NotOnCanceled,
OnlyOnCanceled = NotOnRanToCompletion | NotOnFaulted

(Combining all the Not* flags [NotOnRanToCompletion, NotOnFaulted, NotOnCanceled] is nonsensical, as it would result in the continuation always being canceled.)

“RanToCompletion” means the antecedent succeeded — without cancellation or unhandled exceptions.

“Faulted” means an unhandled exception was thrown on the antecedent.

“Canceled” means one of two things:

It’s essential to grasp that when a continuation doesn’t execute by virtue of these flags, the continuation is not forgotten or abandoned — it’s canceled. This means that any continuations on the continuation itself will then run — unless you predicate them with NotOnCanceled. For example, consider this:

Task t1 = Task.Factory.StartNew (...);
 
Task fault = t1.ContinueWith (ant => Console.WriteLine ("fault"),
                              TaskContinuationOptions.OnlyOnFaulted);
 
Task t3 = fault.ContinueWith (ant => Console.WriteLine ("t3"));

As it stands, t3 will always get scheduled — even if t1 doesn’t throw an exception. This is because if t1 succeeds, the fault task will be canceled, and with no continuation restrictions placed on t3, t3 will then execute unconditionally.

Conditional Continuations

If we want t3 to execute only if fault actually runs, we must instead do this:

Task t3 = fault.ContinueWith (ant => Console.WriteLine ("t3"),
                              TaskContinuationOptions.NotOnCanceled);

(Alternatively, we could specify OnlyOnRanToCompletion; the difference is that t3 would not then execute if an exception was thrown within fault.)

Continuations with multiple antecedents

Another useful feature of continuations is that you can schedule them to execute based on the completion of multiple antecedents. ContinueWhenAll schedules execution when all antecedents have completed; ContinueWhenAny schedules execution when one antecedent completes. Both methods are defined in the TaskFactory class:

var task1 = Task.Factory.StartNew (() => Console.Write ("X"));
var task2 = Task.Factory.StartNew (() => Console.Write ("Y"));
 
var continuation = Task.Factory.ContinueWhenAll (
  new[] { task1, task2 }, tasks => Console.WriteLine ("Done"));

This writes “Done” after writing “XY” or “YX”. The tasks argument in the lambda expression gives you access to the array of completed tasks, which is useful when the antecedents return data. The following example adds together numbers returned from two antecedent tasks:

// task1 and task2 would call complex functions in real life:
Task<int> task1 = Task.Factory.StartNew (() => 123);
Task<int> task2 = Task.Factory.StartNew (() => 456);
 
Task<int> task3 = Task<int>.Factory.ContinueWhenAll (
  new[] { task1, task2 }, tasks => tasks.Sum (t => t.Result));
 
Console.WriteLine (task3.Result);           // 579

We’ve included the <int> type argument in our call to Task.Factory in this example to clarify that we’re obtaining a generic task factory. The type argument is unnecessary, though, as it will be inferred by the compiler.

Multiple continuations on a single antecedent

Calling ContinueWith more than once on the same task creates multiple continuations on a single antecedent. When the antecedent finishes, all continuations will start together (unless you specify TaskContinuationOptions.ExecuteSynchronously, in which case the continuations will execute sequentially).

The following waits for one second, and then writes either “XY” or “YX”:

var t = Task.Factory.StartNew (() => Thread.Sleep (1000));
t.ContinueWith (ant => Console.Write ("X"));
t.ContinueWith (ant => Console.Write ("Y"));

Task Schedulers and UIs

A task scheduler allocates tasks to threads. All tasks are associated with a task scheduler, which is represented by the abstract TaskScheduler class. The Framework provides two concrete implementations: the default scheduler that works in tandem with the CLR thread pool, and the synchronization context scheduler. The latter is designed (primarily) to help you with the threading model of WPF and Windows Forms, which requires that UI elements and controls are accessed only from the thread that created them. For example, suppose we wanted to fetch some data from a web service in the background, and then update a WPF label called lblResult with its result. We can divide this into two tasks:

  1. Call a method to get data from the web service (antecedent task).
  2. Update lblResult with the results (continuation task).

If, for a continuation task, we specify the synchronization context scheduler obtained when the window was constructed, we can safely update lblResult:

public partial class MyWindow : Window
{
  TaskScheduler _uiScheduler;   // Declare this as a field so we can use
                                // it throughout our class.
  public MyWindow()
  {    
    InitializeComponent();
 
    // Get the UI scheduler for the thread that created the form:
    _uiScheduler = TaskScheduler.FromCurrentSynchronizationContext();
 
    Task.Factory.StartNew<string> (SomeComplexWebService)
      .ContinueWith (ant => lblResult.Content = ant.Result, _uiScheduler);
  }
 
  string SomeComplexWebService() { ... }
}

It’s also possible to write our own task scheduler (by subclassing TaskScheduler), although this is something you’d do only in very specialized scenarios. For custom scheduling, you’d more commonly use TaskCompletionSource, which we’ll cover soon.

TaskFactory

When you call Task.Factory, you’re calling a static property on Task that returns a default TaskFactory object. The purpose of a task factory is to create tasks — specifically, three kinds of tasks:

Interestingly, TaskFactory is the only way to achieve the latter two goals. In the case of StartNew, TaskFactory is purely a convenience and technically redundant in that you can simply instantiate Task objects and call Start on them.

Creating your own task factories

TaskFactory is not an abstract factory: you can actually instantiate the class, and this is useful when you want to repeatedly create tasks using the same (nonstandard) values for TaskCreationOptions, TaskContinuationOptions, or TaskScheduler. For example, if we wanted to repeatedly create long-running parented tasks, we could create a custom factory as follows:

var factory = new TaskFactory (
  TaskCreationOptions.LongRunning | TaskCreationOptions.AttachedToParent,
  TaskContinuationOptions.None);

Creating tasks is then simply a matter of calling StartNew on the factory:

Task task1 = factory.StartNew (Method1);
Task task2 = factory.StartNew (Method2);
...

The custom continuation options are applied when calling ContinueWhenAll and ContinueWhenAny.

TaskCompletionSource

The Task class achieves two distinct things:

Interestingly, these two things are not joined at the hip: you can leverage a task’s features for managing work items without scheduling anything to run on the thread pool. The class that enables this pattern of use is called TaskCompletionSource.

To use TaskCompletionSource you simply instantiate the class. It exposes a Task property that returns a task upon which you can wait and attach continuations—just like any other task. The task, however, is entirely controlled by the TaskCompletionSource object via the following methods:

public class TaskCompletionSource<TResult>
{
  public void SetResult (TResult result);
  public void SetException (Exception exception);
  public void SetCanceled();
 
  public bool TrySetResult (TResult result);
  public bool TrySetException (Exception exception);
  public bool TrySetCanceled();
  ...
}

If called more than once, SetResult, SetException, or SetCanceled throws an exception; the Try* methods instead return false.

TResult corresponds to the task’s result type, so TaskCompletionSource<int> gives you a Task<int>. If you want a task with no result, create a TaskCompletionSource of object and pass in null when calling SetResult. You can then cast the Task<object> to Task.

The following example prints 123 after waiting for five seconds:

var source = new TaskCompletionSource<int>();
 
new Thread (() => { Thread.Sleep (5000); source.SetResult (123); })
  .Start();
 
Task<int> task = source.Task;      // Our "slave" task.
Console.WriteLine (task.Result);   // 123 

Later on, we'll show how BlockingCollection can be used to write a producer/consumer queue. We then demonstrate how TaskCompletionSource improves the solution by allowing queued work items to be waited upon and canceled.

Working with AggregateException

As we’ve seen, PLINQ, the Parallel class, and Tasks automatically marshal exceptions to the consumer. To see why this is essential, consider the following LINQ query, which throws a DivideByZeroException on the first iteration:

try
{
  var query = from i in Enumerable.Range (0, 1000000)
              select 100 / i;
  ...
}
catch (DivideByZeroException)
{
  ...
}

If we asked PLINQ to parallelize this query and it ignored the handling of exceptions, a DivideByZeroException would probably be thrown on a separate thread, bypassing our catch block and causing the application to die.

Hence, exceptions are automatically caught and rethrown to the caller. But unfortunately, it’s not quite as simple as catching a DivideByZeroException. Because these libraries leverage many threads, it’s actually possible for two or more exceptions to be thrown simultaneously. To ensure that all exceptions are reported, exceptions are therefore wrapped in an AggregateException container, which exposes an InnerExceptions property containing each of the caught exception(s):

try
{
  var query = from i in ParallelEnumerable.Range (0, 1000000)
              select 100 / i;
  // Enumerate query
  ...
}
catch (AggregateException aex)
{
  foreach (Exception ex in aex.InnerExceptions)
    Console.WriteLine (ex.Message);
}

Both PLINQ and the Parallel class end the query or loop execution upon encountering the first exception — by not processing any further elements or loop bodies. More exceptions might be thrown, however, before the current cycle is complete. The first exception in AggregateException is visible in the InnerException property.

Flatten and Handle

The AggregateException class provides a couple of methods to simplify exception handling: Flatten and Handle.

Flatten

AggregateExceptions will quite often contain other AggregateExceptions. An example of when this might happen is if a child task throws an exception. You can eliminate any level of nesting to simplify handling by calling Flatten. This method returns a new AggregateException with a simple flat list of inner exceptions:

catch (AggregateException aex)
{
  foreach (Exception ex in aex.Flatten().InnerExceptions)
    myLogWriter.LogException (ex);
}

Handle

Sometimes it’s useful to catch only specific exception types, and have other types rethrown. The Handle method on AggregateException provides a shortcut for doing this. It accepts an exception predicate which it runs over every inner exception:

public void Handle (Func<Exception, bool> predicate)

If the predicate returns true, it considers that exception “handled.” After the delegate has run over every exception, the following happens:

For instance, the following ends up rethrowing another AggregateException that contains a single NullReferenceException:

var parent = Task.Factory.StartNew (() => 
{
  // We’ll throw 3 exceptions at once using 3 child tasks:
 
  int[] numbers = { 0 };
 
  var childFactory = new TaskFactory
   (TaskCreationOptions.AttachedToParent, TaskContinuationOptions.None);
 
  childFactory.StartNew (() => 5 / numbers[0]);   // Division by zero
  childFactory.StartNew (() => numbers [1]);      // Index out of range
  childFactory.StartNew (() => { throw null; });  // Null reference
});
 
try { parent.Wait(); }
catch (AggregateException aex)
{
  aex.Flatten().Handle (ex =>   // Note that we still need to call Flatten
  {
    if (ex is DivideByZeroException)
    {
      Console.WriteLine ("Divide by zero");
      return true;                           // This exception is "handled"
    }
    if (ex is IndexOutOfRangeException)
    {
      Console.WriteLine ("Index out of range");
      return true;                           // This exception is "handled"   
    }
    return false;    // All other exceptions will get rethrown
  });
}

Concurrent Collections

Framework 4.0 provides a set of new collections in the System.Collections.Concurrent namespace. All of these are fully thread-safe:

Concurrent collection Nonconcurrent equivalent
ConcurrentStack<T> Stack<T>
ConcurrentQueue<T> Queue<T>
ConcurrentBag<T> (none)
BlockingCollection<T> (none)
ConcurrentDictionary<TKey,TValue> Dictionary<TKey,TValue>

The concurrent collections can sometimes be useful in general multithreading when you need a thread-safe collection. However, there are some caveats:

In other words, these collections don’t merely provide shortcuts for using an ordinary collection with a lock. To demonstrate, if we execute the following code on a single thread:

var d = new ConcurrentDictionary<int,int>();
for (int i = 0; i < 1000000; i++) d[i] = 123;

it runs three times more slowly than this:

var d = new Dictionary<int,int>();
for (int i = 0; i < 1000000; i++) lock (d) d[i] = 123;

(Reading from a ConcurrentDictionary, however, is fast because reads are lock-free.)

The concurrent collections also differ from conventional collections in that they expose special methods to perform atomic test-and-act operations, such as TryPop. Most of these methods are unified via the IProducerConsumerCollection<T> interface.

IProducerConsumerCollection<T>

A producer/consumer collection is one for which the two primary use cases are:

The classic examples are stacks and queues. Producer/consumer collections are significant in parallel programming because they’re conducive to efficient lock-free implementations.

The IProducerConsumerCollection<T> interface represents a thread-safe producer/consumer collection. The following classes implement this interface:

ConcurrentStack<T>
ConcurrentQueue<T>
ConcurrentBag<T>

IProducerConsumerCollection<T> extends ICollection, adding the following methods:

void CopyTo (T[] array, int index);
T[] ToArray();
bool TryAdd (T item);
bool TryTake (out T item);

The TryAdd and TryTake methods test whether an add/remove operation can be performed, and if so, they perform the add/remove. The testing and acting are performed atomically, eliminating the need to lock as you would around a conventional collection:

int result;
lock (myStack) if (myStack.Count > 0) result = myStack.Pop();

TryTake returns false if the collection is empty. TryAdd always succeeds and returns true in the three implementations provided. If you wrote your own concurrent collection that prohibited duplicates, however, you’d make TryAdd return false if the element already existed (an example would be if you wrote a concurrent set).

The particular element that TryTake removes is defined by the subclass:

The three concrete classes mostly implement the TryTake and TryAdd methods explicitly, exposing the same functionality through more specifically named public methods such as TryDequeue and TryPop.

ConcurrentBag<T>

ConcurrentBag<T> stores an unordered collection of objects (with duplicates permitted). ConcurrentBag<T> is suitable in situations when you don’t care which element you get when calling Take or TryTake.

The benefit of ConcurrentBag<T> over a concurrent queue or stack is that a bag’s Add method suffers almost no contention when called by many threads at once. In contrast, calling Add in parallel on a queue or stack incurs some contention (although a lot less than locking around a nonconcurrent collection). Calling Take on a concurrent bag is also very efficient — as long as each thread doesn’t take more elements than it Added.

Inside a concurrent bag, each thread gets it own private linked list. Elements are added to the private list that belongs to the thread calling Add, eliminating contention. When you enumerate over the bag, the enumerator travels through each thread’s private list, yielding each of its elements in turn.

When you call Take, the bag first looks at the current thread’s private list. If there’s at least one element, it can complete the task easily and (in most cases) without contention. But if the list is empty, it must “steal” an element from another thread’s private list and incur the potential for contention.

So, to be precise, calling Take gives you the element added most recently on that thread; if there are no elements on that thread, it gives you the element added most recently on another thread, chosen at random.

Concurrent bags are ideal when the parallel operation on your collection mostly comprises Adding elements — or when the Adds and Takes are balanced on a thread. We saw an example of the former previously, when using Parallel.ForEach to implement a parallel spellchecker:

var misspellings = new ConcurrentBag<Tuple<int,string>>();
 
Parallel.ForEach (wordsToTest, (word, state, i) =>
{
  if (!wordLookup.Contains (word))
    misspellings.Add (Tuple.Create ((int) i, word));
});

A concurrent bag would be a poor choice for a producer/consumer queue, because elements are added and removed by different threads.

BlockingCollection<T>

If you call TryTake on any of the producer/consumer collections we discussed previously:

ConcurrentStack<T>
ConcurrentQueue<T>
ConcurrentBag<T>

and the collection is empty, the method returns false. Sometimes it would be more useful in this scenario to wait until an element is available.

Rather than overloading the TryTake methods with this functionality (which would have caused a blowout of members after allowing for cancellation tokens and timeouts), PFX’s designers encapsulated this functionality into a wrapper class called BlockingCollection<T>. A blocking collection wraps any collection that implements IProducerConsumerCollection<T> and lets you Take an element from the wrapped collection — blocking if no element is available.

A blocking collection also lets you limit the total size of the collection, blocking the producer if that size is exceeded. A collection limited in this manner is called a bounded blocking collection.

To use BlockingCollection<T>:

  1. Instantiate the class, optionally specifying the IProducerConsumerCollection<T> to wrap and the maximum size (bound) of the collection.
  2. Call Add or TryAdd to add elements to the underlying collection.
  3. 3.Call Take or TryTake to remove (consume) elements from the underlying collection.

If you call the constructor without passing in a collection, the class will automatically instantiate a ConcurrentQueue<T>. The producing and consuming methods let you specify cancellation tokens and timeouts. Add and TryAdd may block if the collection size is bounded; Take and TryTake block while the collection is empty.

Another way to consume elements is to call GetConsumingEnumerable. This returns a (potentially) infinite sequence that yields elements as they become available. You can force the sequence to end by calling CompleteAdding: this method also prevents further elements from being enqueued.

Previously, we wrote a producer/consumer queue using Wait and Pulse. Here’s the same class refactored to use BlockingCollection<T> (exception handling aside):

public class PCQueue : IDisposable
{
  BlockingCollection<Action> _taskQ = new BlockingCollection<Action>(); 
  public PCQueue (int workerCount)
  {
    // Create and start a separate Task for each consumer:
    for (int i = 0; i < workerCount; i++)
      Task.Factory.StartNew (Consume);
  }
 
  public void Dispose() { _taskQ.CompleteAdding(); }
 
  public void EnqueueTask (Action action) { _taskQ.Add (action); }
 
  void Consume()
  {
    // This sequence that we’re enumerating will block when no elements
    // are available and will end when CompleteAdding is called. 
    foreach (Action action in _taskQ.GetConsumingEnumerable())
      action();     // Perform task.
  }
}

Because we didn’t pass anything into BlockingCollection’s constructor, it instantiated a concurrent queue automatically. Had we passed in a ConcurrentStack, we’d have ended up with a producer/consumer stack.

BlockingCollection also provides static methods called AddToAny and TakeFromAny, which let you add or take an element while specifying several blocking collections. The action is then honored by the first collection able to service the request.

Leveraging TaskCompletionSource

The producer/consumer that we just wrote is inflexible in that we can’t track work items after they’ve been enqueued. It would be nice if we could:

An ideal solution would be to have the EnqueueTask method return some object giving us the functionality just described. The good news is that a class already exists to do exactly this — the Task class. All we need to do is to hijack control of the task via TaskCompletionSource:

public class PCQueue : IDisposable
{
  class WorkItem
  {
    public readonly TaskCompletionSource<object> TaskSource;
    public readonly Action Action;
    public readonly CancellationToken? CancelToken;
 
    public WorkItem (
      TaskCompletionSource<object> taskSource,
      Action action,
      CancellationToken? cancelToken)
    {
      TaskSource = taskSource;
      Action = action;
      CancelToken = cancelToken;
    }
  }
 
  BlockingCollection<WorkItem> _taskQ = new BlockingCollection<WorkItem>();
 
  public PCQueue (int workerCount)
  {
    // Create and start a separate Task for each consumer:
    for (int i = 0; i < workerCount; i++)
      Task.Factory.StartNew (Consume);
  }
 
  public void Dispose() { _taskQ.CompleteAdding(); }
 
  public Task EnqueueTask (Action action) 
  {
    return EnqueueTask (action, null);
  }
 
  public Task EnqueueTask (Action action, CancellationToken? cancelToken)
  {
    var tcs = new TaskCompletionSource<object>();
    _taskQ.Add (new WorkItem (tcs, action, cancelToken));
    return tcs.Task;
  }
 
  void Consume()
  {
    foreach (WorkItem workItem in _taskQ.GetConsumingEnumerable())
      if (workItem.CancelToken.HasValue && 
          workItem.CancelToken.Value.IsCancellationRequested)
      {
        workItem.TaskSource.SetCanceled();
      }
      else
        try
        {
          workItem.Action();
          workItem.TaskSource.SetResult (null);   // Indicate completion
        }
        catch (OperationCanceledException ex)
        {
          if (ex.CancellationToken == workItem.CancelToken)
            workItem.TaskSource.SetCanceled();
          else
            workItem.TaskSource.SetException (ex);
        }
        catch (Exception ex)
        {
          workItem.TaskSource.SetException (ex);
        }
  }
}

In EnqueueTask, we enqueue a work item that encapsulates the target delegate and a task completion source — which lets us later control the task that we return to the consumer.

In Consume, we first check whether a task has been canceled after dequeuing the work item. If not, we run the delegate and then call SetResult on the task completion source to indicate its completion.

Here’s how we can use this class:

var pcQ = new PCQueue (1);
Task task = pcQ.EnqueueTask (() => Console.WriteLine ("Easy!"));
...

We can now wait on task, perform continuations on it, have exceptions propagate to continuations on parent tasks, and so on. In other words, we’ve got the richness of the task model while, in effect, implementing our own scheduler.

SpinLock and SpinWait

In parallel programming, a brief episode of spinning is often preferable to blocking, as it avoids the cost of context switching and kernel transitions. SpinLock and SpinWait are designed to help in such cases. Their main use is in writing custom synchronization constructs.

SpinLock and SpinWait are structs and not classes! This design decision was an extreme optimization technique to avoid the cost of indirection and garbage collection. It means that you must be careful not to unintentionally copy instances — by passing them to another method without the ref modifier, for instance, or declaring them as readonly fields. This is particularly important in the case of SpinLock.

SpinLock

The SpinLock struct lets you lock without incurring the cost of a context switch, at the expense of keeping a thread spinning (uselessly busy). This approach is valid in high-contention scenarios when locking will be very brief (e.g., in writing a thread-safe linked list from scratch).

If you leave a spinlock contended for too long (we’re talking milliseconds at most), it will yield its time slice, causing a context switch just like an ordinary lock. When rescheduled, it will yield again — in a continual cycle of “spin yielding.” This consumes far fewer CPU resources than outright spinning — but more than blocking.

On a single-core machine, a spinlock will start “spin yielding” immediately if contended.

Using a SpinLock is like using an ordinary lock, except:

Another difference is that when you call Enter, you must follow the robust pattern of providing a lockTaken argument (which is nearly always done within a try/finally block).

Here’s an example:

var spinLock = new SpinLock (true);   // Enable owner tracking
bool lockTaken = false;
try
{
  spinLock.Enter (ref lockTaken);
  // Do stuff...
}
finally
{
  if (lockTaken) spinLock.Exit();
}

As with an ordinary lock, lockTaken will be false after calling Enter if (and only if) the Enter method throws an exception and the lock was not taken. This happens in very rare scenarios (such as Abort being called on the thread, or an OutOfMemoryException being thrown) and lets you reliably know whether to subsequently call Exit.

SpinLock also provides a TryEnter method which accepts a timeout.

Given SpinLock’s ungainly value-type semantics and lack of language support, it’s almost as if they want you to suffer every time you use it! Think carefully before dismissing an ordinary lock.

A SpinLock makes the most sense when writing your own reusable synchronization constructs. Even then, a spinlock is not as useful as it sounds. It still limits concurrency. And it wastes CPU time doing nothing useful. Often, a better choice is to spend some of that time doing something speculative — with the help of SpinWait.

SpinWait

SpinWait helps you write lock-free code that spins rather than blocks. It works by implementing safeguards to avoid the dangers of resource starvation and priority inversion that might otherwise arise with spinning.

Lock-free programming with SpinWait is as hardcore as multithreading gets and is intended for when none of the higher-level constructs will do. A prerequisite is to understand Nonblocking Synchronization.

Why we need SpinWait

Suppose we wrote a spin-based signaling system based purely on a simple flag:

bool _proceed;
void Test()
{
  // Spin until another thread sets _proceed to true:
  while (!_proceed) Thread.MemoryBarrier();
  ...
}

This would be highly efficient if Test ran when _proceed was already true — or if _proceed became true within a few cycles. But now suppose that _proceed remained false for several seconds — and that four threads called Test at once. The spinning would then fully consume a quad-core CPU! This would cause other threads to run slowly (resource starvation) — including the very thread that might eventually set _proceed to true (priority inversion). The situation is exacerbated on single-core machines, where spinning will nearly always cause priority inversion. (And although single-core machines are rare nowadays, single-core virtual machines are not.)

SpinWait addresses these problems in two ways. First, it limits CPU-intensive spinning to a set number of iterations, after which it yields its time slice on every spin (by calling Thread.Yield and Thread.Sleep), lowering its resource consumption. Second, it detects whether it’s running on a single-core machine, and if so, it yields on every cycle.

How to use SpinWait

There are two ways to use SpinWait. The first is to call its static method, SpinUntil. This method accepts a predicate (and optionally, a timeout):

bool _proceed;
void Test()
{
  SpinWait.SpinUntil (() => { Thread.MemoryBarrier(); return _proceed; });
  ...
}

The other (more flexible) way to use SpinWait is to instantiate the struct and then to call SpinOnce in a loop:

bool _proceed;
void Test()
{
  var spinWait = new SpinWait();
  while (!_proceed) { Thread.MemoryBarrier(); spinWait.SpinOnce(); }
  ...
}

The former is a shortcut for the latter.

How SpinWait works

In its current implementation, SpinWait performs CPU-intensive spinning for 10 iterations before yielding. However, it doesn’t return to the caller immediately after each of those cycles: instead, it calls Thread.SpinWait to spin via the CLR (and ultimately the operating system) for a set time period. This time period is initially a few tens of nanoseconds, but doubles with each iteration until the 10 iterations are up. This ensures some predictability in the total time spent in the CPU-intensive spinning phase, which the CLR and operating system can tune according to conditions. Typically, it’s in the few-tens-of-microseconds region — small, but more than the cost of a context switch.

On a single-core machine, SpinWait yields on every iteration. You can test whether SpinWait will yield on the next spin via the property NextSpinWillYield.

If a SpinWait remains in “spin-yielding” mode for long enough (maybe 20 cycles) it will periodically sleep for a few milliseconds to further save resources and help other threads progress.

Lock-free updates with SpinWait and Interlocked.CompareExchange

SpinWait in conjunction with Interlocked.CompareExchange can atomically update fields with a value calculated from the original (read-modify-write). For example, suppose we want to multiply field x by 10. Simply doing the following is not thread-safe:

x = x * 10;

for the same reason that incrementing a field is not thread-safe, as we saw in Nonblocking Synchronization.

The correct way to do this without locks is as follows:

  1. Take a “snapshot” of x into a local variable.
  2. Calculate the new value (in this case by multiplying the snapshot by 10).
  3. Write the calculated value back if the snapshot is still up-to-date (this step must be done atomically by calling Interlocked.CompareExchange).
  4. If the snapshot was stale, spin and return to step 1.

For example:

int x;
 
void MultiplyXBy (int factor)
{
  var spinWait = new SpinWait();
  while (true)
  {
    int snapshot1 = x;
    Thread.MemoryBarrier();
    int calc = snapshot1 * factor;
    int snapshot2 = Interlocked.CompareExchange (ref x, calc, snapshot1);
    if (snapshot1 == snapshot2) return;   // No one preempted us.
    spinWait.SpinOnce();
  }
}

We can improve performance (slightly) by doing away with the call to Thread.MemoryBarrier. We can get away with this because CompareExchange generates a memory barrier anyway — so the worst that can happen is an extra spin if snapshot1 happens to read a stale value in its first iteration.

Interlocked.CompareExchange updates a field with a specified value if the field’s current value matches the third argument. It then returns the field’s old value, so you can test whether it succeeded by comparing that against the original snapshot. If the values differ, it means that another thread preempted you, in which case you spin and try again.

CompareExchange is overloaded to work with the object type too. We can leverage this overload by writing a lock-free update method that works with all reference types:

static void LockFreeUpdate<T> (ref T field, Func <T, T> updateFunction)
  where T : class
{
  var spinWait = new SpinWait();
  while (true)
  {
    T snapshot1 = field;
    T calc = updateFunction (snapshot1);
    T snapshot2 = Interlocked.CompareExchange (ref field, calc, snapshot1);
    if (snapshot1 == snapshot2) return;
    spinWait.SpinOnce();
  }
}

Here’s how we can use this method to write a thread-safe event without locks (this is, in fact, what the C# 4.0 compiler now does by default with events):

EventHandler _someDelegate;
public event EventHandler SomeEvent
{
  add    { LockFreeUpdate (ref _someDelegate, d => d + value); }
  remove { LockFreeUpdate (ref _someDelegate, d => d - value); }
}

Finally, consider the following class:

class Test
{
  ProgressStatus _status = new ProgressStatus (0, "Starting");
 
  class ProgressStatus    // Immutable class
  {
    public readonly int PercentComplete;
    public readonly string StatusMessage;
 
    public ProgressStatus (int percentComplete, string statusMessage)
    {
      PercentComplete = percentComplete;
      StatusMessage = statusMessage;
    }
  }
}

We can use our LockFreeUpdate method to “increment” the PercentComplete field in _status as follows:

LockFreeUpdate (ref _status,
  s => new ProgressStatus (s.PercentComplete + 1, s.StatusMessage));

Notice that we’re creating a new ProgressStatus object based on existing values. Thanks to the LockFreeUpdate method, the act of reading the existing PercentComplete value, incrementing it, and writing it back can’t get unsafely preempted: any preemption is reliably detected, triggering a spin and retry.

<< Part 4


Valid XHTML 1.0 Transitional

Threading in C# is from Chapters 21 and 22 of C# 4.0 in a Nutshell.

© 2006-2014 Joseph Albahari, O'Reilly Media, Inc. All rights reserved