mcs/class/referencesource/System.Core/System/Linq/Parallel/Scheduling/OrderPreservingPipeliningSpoolingTask.cs

   1 // ==++==
   2 //
   3 //   Copyright (c) Microsoft Corporation.  All rights reserved.
   4 //
   5 // ==--==
   6 // =+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
   7 //
   8 // OrderPreservingPipeliningSpoolingTask.cs
   9 //
  10 // <OWNER>[....]</OWNER>
  11 //
  12 // =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
  13
  14 using System;
  15 using System.Collections.Generic;
  16 using System.Linq;
  17 using System.Linq.Parallel;
  18 using System.Text;
  19 using System.Threading;
  20 using System.Threading.Tasks;
  21 using System.Diagnostics.Contracts;
  22
  23 namespace System.Linq.Parallel
  24 {
  25     class OrderPreservingPipeliningSpoolingTask<TOutput, TKey> : SpoolingTaskBase
  26     {
  27         private readonly QueryTaskGroupState m_taskGroupState; // State shared among tasks.
  28         private readonly TaskScheduler m_taskScheduler; // The task manager to execute the query.
  29         private readonly QueryOperatorEnumerator<TOutput, TKey> m_partition; // The source partition.
  30         private readonly bool[] m_consumerWaiting; // Whether a consumer is waiting on a particular producer
  31         private readonly bool[] m_producerWaiting; // Whether a particular producer is waiting on the consumer
  32         private readonly bool[] m_producerDone; // Whether each producer is done
  33         private readonly int m_partitionIndex; // Index of the partition owned by this task.
  34
  35         private readonly Queue<Pair<TKey, TOutput>>[] m_buffers; // The buffer for the results
  36         private readonly object m_bufferLock; // A lock for the buffer
  37
  38         /// <summary>
  39         /// Whether the producer is allowed to buffer up elements before handing a chunk to the consumer.
  40         /// If false, the producer will make each result available to the consumer immediately after it is
  41         /// produced.
  42         /// </summary>
  43         private readonly bool m_autoBuffered;
  44
  45         /// <summary>
  46         /// The number of elements to accumulate on the producer before copying the elements to the
  47         /// producer-consumer buffer. This constant is only used in the AutoBuffered mode.
  48         ///
  49         /// Experimentally, 16 appears to be sufficient buffer size to compensate for the synchronization
  50         /// cost.
  51         /// </summary>
  52         private const int PRODUCER_BUFFER_AUTO_SIZE = 16;
  53
  54         /// <summary>
  55         /// Constructor
  56         /// </summary>
  57         internal OrderPreservingPipeliningSpoolingTask(
  58             QueryOperatorEnumerator<TOutput, TKey> partition,
  59             QueryTaskGroupState taskGroupState,
  60             bool[] consumerWaiting,
  61             bool[] producerWaiting,
  62             bool[] producerDone,
  63             int partitionIndex,
  64             Queue<Pair<TKey, TOutput>>[] buffers,
  65             object bufferLock,
  66             TaskScheduler taskScheduler,
  67             bool autoBuffered)
  68             :base(partitionIndex, taskGroupState)
  69         {
  70             Contract.Requires(partition != null);
  71             Contract.Requires(taskGroupState != null);
  72             Contract.Requires(consumerWaiting != null);
  73             Contract.Requires(producerWaiting != null && producerWaiting.Length == consumerWaiting.Length);
  74             Contract.Requires(producerDone != null && producerDone.Length == consumerWaiting.Length);
  75             Contract.Requires(buffers != null && buffers.Length == consumerWaiting.Length);
  76             Contract.Requires(partitionIndex >= 0 && partitionIndex < consumerWaiting.Length);
  77
  78             m_partition = partition;
  79             m_taskGroupState = taskGroupState;
  80             m_producerDone = producerDone;
  81             m_consumerWaiting = consumerWaiting;
  82             m_producerWaiting = producerWaiting;
  83             m_partitionIndex = partitionIndex;
  84             m_buffers = buffers;
  85             m_bufferLock = bufferLock;
  86             m_taskScheduler = taskScheduler;
  87             m_autoBuffered = autoBuffered;
  88         }
  89
  90         /// <summary>
  91         /// This method is responsible for enumerating results and enqueueing them to
  92         /// the output buffer as appropriate.  Each base class implements its own.
  93         /// </summary>
  94         protected override void SpoolingWork()
  95         {
  96             TOutput element = default(TOutput);
  97             TKey key = default(TKey);
  98
  99             int chunkSize = m_autoBuffered ? PRODUCER_BUFFER_AUTO_SIZE : 1;
 100             Pair<TKey,TOutput>[] chunk = new Pair<TKey,TOutput>[chunkSize];
 101             var partition = m_partition;
 102             CancellationToken cancelToken = m_taskGroupState.CancellationState.MergedCancellationToken;
 103
 104             int lastChunkSize;
 105             do
 106             {
 107                 lastChunkSize = 0;
 108                 while (lastChunkSize < chunkSize && partition.MoveNext(ref element, ref key))
 109                 {
 110                     chunk[lastChunkSize] = new Pair<TKey, TOutput>(key, element);
 111                     lastChunkSize++;
 112                 }
 113
 114                 if (lastChunkSize == 0) break;
 115
 116                 lock (m_bufferLock)
 117                 {
 118                     // Check if the query has been cancelled.
 119                     if (cancelToken.IsCancellationRequested)
 120                     {
 121                         break;
 122                     }
 123
 124                     for (int i = 0; i < lastChunkSize; i++)
 125                     {
 126                         m_buffers[m_partitionIndex].Enqueue(chunk[i]);
 127                     }
 128
 129                     if (m_consumerWaiting[m_partitionIndex])
 130                     {
 131                         Monitor.Pulse(m_bufferLock);
 132                         m_consumerWaiting[m_partitionIndex] = false;
 133                     }
 134
 135                     // If the producer buffer is too large, wait.
 136                     // Note: we already checked for cancellation after acquiring the lock on this producer.
 137                     // That guarantees that the consumer will eventually wake up the producer.
 138                     if (m_buffers[m_partitionIndex].Count >= OrderPreservingPipeliningMergeHelper<TOutput, TKey>.MAX_BUFFER_SIZE)
 139                     {
 140                         m_producerWaiting[m_partitionIndex] = true;
 141                         Monitor.Wait(m_bufferLock);
 142                     }
 143                 }
 144             } while (lastChunkSize == chunkSize);
 145         }
 146
 147
 148         /// <summary>
 149         /// Creates and begins execution of a new set of spooling tasks.
 150         /// </summary>
 151         public static void Spool(
 152             QueryTaskGroupState groupState, PartitionedStream<TOutput, TKey> partitions,
 153             bool[] consumerWaiting, bool[] producerWaiting, bool[] producerDone,
 154             Queue<Pair<TKey,TOutput>>[] buffers, object[] bufferLocks,
 155             TaskScheduler taskScheduler, bool autoBuffered)
 156         {
 157             Contract.Requires(groupState != null);
 158             Contract.Requires(partitions != null);
 159             Contract.Requires(producerDone != null && producerDone.Length == partitions.PartitionCount);
 160             Contract.Requires(buffers != null && buffers.Length == partitions.PartitionCount);
 161             Contract.Requires(bufferLocks != null);
 162
 163             int degreeOfParallelism = partitions.PartitionCount;
 164
 165             // Initialize the buffers and buffer locks.
 166             for (int i = 0; i < degreeOfParallelism; i++)
 167             {
 168                 buffers[i] = new Queue<Pair<TKey, TOutput>>(OrderPreservingPipeliningMergeHelper<TOutput, TKey>.INITIAL_BUFFER_SIZE);
 169                 bufferLocks[i] = new object();
 170             }
 171
 172             // Ensure all tasks in this query are parented under a common root. Because this
 173             // is a pipelined query, we detach it from the parent (to avoid blocking the calling
 174             // thread), and run the query on a separate thread.
 175             Task rootTask = new Task(
 176                 () =>
 177                 {
 178                     for (int i = 0; i < degreeOfParallelism; i++)
 179                     {
 180                         QueryTask asyncTask = new OrderPreservingPipeliningSpoolingTask<TOutput, TKey>(
 181                             partitions[i], groupState, consumerWaiting, producerWaiting,
 182                             producerDone, i, buffers, bufferLocks[i], taskScheduler, autoBuffered);
 183                         asyncTask.RunAsynchronously(taskScheduler);
 184                     }
 185                 });
 186
 187             // Begin the query on the calling thread.
 188             groupState.QueryBegin(rootTask);
 189
 190             // And schedule it for execution.  This is done after beginning to ensure no thread tries to
 191             // end the query before its root task has been recorded properly.
 192             rootTask.Start(taskScheduler);
 193
 194             // We don't call QueryEnd here; when we return, the query is still executing, and the
 195             // last enumerator to be disposed of will call QueryEnd for us.
 196         }
 197
 198         /// <summary>
 199         /// Dispose the underlying enumerator and wake up the consumer if necessary.
 200         /// </summary>
 201         protected override void SpoolingFinally()
 202         {
 203             // Let the consumer know that this producer is done.
 204             lock (m_bufferLock)
 205             {
 206                 m_producerDone[m_partitionIndex] = true;
 207                 if (m_consumerWaiting[m_partitionIndex])
 208                 {
 209                     Monitor.Pulse(m_bufferLock);
 210                     m_consumerWaiting[m_partitionIndex] = false;
 211                 }
 212             }
 213
 214             // Call the base implementation.
 215             base.SpoolingFinally();
 216
 217             // Dispose of the source enumerator *after* signaling that the task is done.
 218             // We call Dispose() last to ensure that if it throws an exception, we will not cause a deadlock.
 219             m_partition.Dispose();
 220         }
 221     }
 222 }