痛苦地减慢Azure表插入和删除批处理操作

Question

使用Azure表存储时,我遇到了巨大的性能瓶颈.我的愿望是使用表作为一种缓存,因此一个漫长的过程可能会导致数百到数千行的数据.然后可以通过分区和行键快速查询数据.

查询工作速度非常快(仅使用分区和行键时速度极快,速度稍慢,但在搜索特定匹配项的属性时仍然可以接受).

但是,插入和删除行都非常慢.

澄清

我想澄清一下,即使插入一批100件物品也需要几秒钟.这不仅仅是数千行总吞吐量的问题.当我只插入100时,它正在影响我.

以下是我的代码批量插入到我的表中的示例:

static async Task BatchInsert( CloudTable table, List<ITableEntity> entities )
    {
        int rowOffset = 0;

        while ( rowOffset < entities.Count )
        {
            Stopwatch sw = Stopwatch.StartNew();

            var batch = new TableBatchOperation();

            // next batch
            var rows = entities.Skip( rowOffset ).Take( 100 ).ToList();

            foreach ( var row in rows )
                batch.Insert( row );

            // submit
            await table.ExecuteBatchAsync( batch );

            rowOffset += rows.Count;

            Trace.TraceInformation( "Elapsed time to batch insert " + rows.Count + " rows: " + sw.Elapsed.ToString( "g" ) );
        }
    }

我正在使用批处理操作,这是调试输出的一个示例:

Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Starting asynchronous request to http://127.0.0.1:10002/devstoreaccount1.
Microsoft.WindowsAzure.Storage Verbose: 4 : b08a07da-fceb-4bec-af34-3beaa340239b: StringToSign = POST..multipart/mixed; boundary=batch_6d86d34c-5e0e-4c0c-8135-f9788ae41748.Tue, 30 Jul 2013 18:48:38 GMT./devstoreaccount1/devstoreaccount1/$batch.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Preparing to write request data.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Writing request data.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Waiting for response.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Response received. Status code = 202, Request ID = , Content-MD5 = , ETag = .
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Response headers were processed successfully, proceeding with the rest of the operation.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Processing response body.
Microsoft.WindowsAzure.Storage Information: 3 : b08a07da-fceb-4bec-af34-3beaa340239b: Operation completed successfully.
iisexpress.exe Information: 0 : Elapsed time to batch insert 100 rows: 0:00:00.9351871

如您所见,此示例需要大约1秒钟才能插入100行.我的开发机器(3.4 Ghz四核)的平均值似乎约为0.8秒.

这看起来很荒谬.

以下是批量删除操作的示例:

Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Starting asynchronous request to http://127.0.0.1:10002/devstoreaccount1.
Microsoft.WindowsAzure.Storage Verbose: 4 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: StringToSign = POST..multipart/mixed; boundary=batch_7e3d229f-f8ac-4aa0-8ce9-ed00cb0ba321.Tue, 30 Jul 2013 18:47:41 GMT./devstoreaccount1/devstoreaccount1/$batch.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Preparing to write request data.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Writing request data.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Waiting for response.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Response received. Status code = 202, Request ID = , Content-MD5 = , ETag = .
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Response headers were processed successfully, proceeding with the rest of the operation.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Processing response body.
Microsoft.WindowsAzure.Storage Information: 3 : 4c271cb5-7463-44b1-b2e5-848b8fb10a93: Operation completed successfully.
iisexpress.exe Information: 0 : Elapsed time to batch delete 100 rows: 0:00:00.6524402

持续超过0.5秒.

我也将此部署到Azure(小实例),并记录了20分钟的时间来插入28000行.

我目前正在使用2.1 RC版本的Storage Client Library:MSDN Blog

我一定是做错了.有什么想法吗？

UPDATE

我尝试了并行性与整体速度改进的净效果(以及8个最大逻辑处理器),但在我的开发机器上仍然每秒只有150行插入.

总而言之,我可以告诉他们,当部署到Azure(小实例)时可能更糟糕.

我已经增加了线程池,并通过遵循此建议增加了我的WebRole的最大HTTP连接数.

我仍然觉得我遗漏了一些限制我的插入/删除到150 ROPS的基本内容.

更新2

在分析部署到Azure的小实例(使用内置于2.1 RC Storage Client的新日志记录)中的一些诊断日志后,我有更多信息.

批量插入的第一个存储客户端日志位于635109046781264034刻度:

caf06fca-1857-4875-9923-98979d850df3: Starting synchronous request to https://?.table.core.windows.net/.; TraceSource 'Microsoft.WindowsAzure.Storage' event

然后差不多3秒钟后,我看到这个日志在635109046810104314滴答声中:

caf06fca-1857-4875-9923-98979d850df3: Preparing to write request data.; TraceSource 'Microsoft.WindowsAzure.Storage' event

然后再添加几个日志,这些日志占用0.15秒,以这个结尾的方式结束635109046811645418插入:

caf06fca-1857-4875-9923-98979d850df3: Operation completed successfully.; TraceSource 'Microsoft.WindowsAzure.Storage' event

我不知道该怎么做,但是在我检查的批量插入日志中它非常一致.

更新3

以下是用于并行批量插入的代码.在此代码中,仅用于测试,我确保将每批100个插入到一个唯一的分区中.

static async Task BatchInsert( CloudTable table, List<ITableEntity> entities )
    {
        int rowOffset = 0;

        var tasks = new List<Task>();

        while ( rowOffset < entities.Count )
        {
            // next batch
            var rows = entities.Skip( rowOffset ).Take( 100 ).ToList();

            rowOffset += rows.Count;

            string partition = "$" + rowOffset.ToString();

            var task = Task.Factory.StartNew( () =>
                {
                    Stopwatch sw = Stopwatch.StartNew();

                    var batch = new TableBatchOperation();

                    foreach ( var row in rows )
                    {
                        row.PartitionKey = row.PartitionKey + partition;
                        batch.InsertOrReplace( row );
                    }

                    // submit
                    table.ExecuteBatch( batch );

                    Trace.TraceInformation( "Elapsed time to batch insert " + rows.Count + " rows: " + sw.Elapsed.ToString( "F2" ) );
                } );

            tasks.Add( task );
        }

        await Task.WhenAll( tasks );
    }

如上所述,这确实有助于缩短插入数千行的总时间,但每批100仍然需要几秒钟.

更新4

因此,我使用VS2012.2创建了一个全新的Azure云服务项目,将Web角色作为单页模板(带有TODO示例的新模板).

这是开箱即用的,没有新的NuGet包或任何东西.它默认使用Storage Client库v2,以及EDM和相关库v5.2.

我只是将HomeController代码修改为以下内容(使用一些随机数据来模拟我想要存储在真实应用程序中的列):

public ActionResult Index( string returnUrl )
    {
        ViewBag.ReturnUrl = returnUrl;

        Task.Factory.StartNew( () =>
            {
                TableTest();
            } );

        return View();
    }

    static Random random = new Random();
    static double RandomDouble( double maxValue )
    {
        // the Random class is not thread safe!
        lock ( random ) return random.NextDouble() * maxValue;
    }

    void TableTest()
    {
        // Retrieve storage account from connection-string
        CloudStorageAccount storageAccount = CloudStorageAccount.Parse(
            CloudConfigurationManager.GetSetting( "CloudStorageConnectionString" ) );

        // create the table client
        CloudTableClient tableClient = storageAccount.CreateCloudTableClient();

        // retrieve the table
        CloudTable table = tableClient.GetTableReference( "test" );

        // create it if it doesn't already exist
        if ( table.CreateIfNotExists() )
        {
            // the container is new and was just created
            Trace.TraceInformation( "Created table named " + "test" );
        }


        Stopwatch sw = Stopwatch.StartNew();

        // create a bunch of objects
        int count = 28000;
        List<DynamicTableEntity> entities = new List<DynamicTableEntity>( count );

        for ( int i = 0; i < count; i++ )
        {
            var row = new DynamicTableEntity()
            {
                PartitionKey = "filename.txt",
                RowKey = string.Format( "$item{0:D10}", i ),
            };

            row.Properties.Add( "Name", EntityProperty.GeneratePropertyForString( i.ToString() ) );
            row.Properties.Add( "Data", EntityProperty.GeneratePropertyForString( string.Format( "data{0}", i ) ) );
            row.Properties.Add( "Value1", EntityProperty.GeneratePropertyForDouble( RandomDouble( 10000 ) ) );
            row.Properties.Add( "Value2", EntityProperty.GeneratePropertyForDouble( RandomDouble( 10000 ) ) );
            row.Properties.Add( "Value3", EntityProperty.GeneratePropertyForDouble( RandomDouble( 1000 ) ) );
            row.Properties.Add( "Value4", EntityProperty.GeneratePropertyForDouble( RandomDouble( 90 ) ) );
            row.Properties.Add( "Value5", EntityProperty.GeneratePropertyForDouble( RandomDouble( 180 ) ) );
            row.Properties.Add( "Value6", EntityProperty.GeneratePropertyForDouble( RandomDouble( 1000 ) ) );

            entities.Add( row );
        }

        Trace.TraceInformation( "Elapsed time to create record rows: " + sw.Elapsed.ToString() );

        sw = Stopwatch.StartNew();

        Trace.TraceInformation( "Inserting rows" );

        // batch our inserts (100 max)
        BatchInsert( table, entities ).Wait();

        Trace.TraceInformation( "Successfully inserted " + entities.Count + " rows into table " + table.Name );
        Trace.TraceInformation( "Elapsed time: " + sw.Elapsed.ToString() );

        Trace.TraceInformation( "Done" );
    }


            static async Task BatchInsert( CloudTable table, List<DynamicTableEntity> entities )
    {
        int rowOffset = 0;

        var tasks = new List<Task>();

        while ( rowOffset < entities.Count )
        {
            // next batch
            var rows = entities.Skip( rowOffset ).Take( 100 ).ToList();

            rowOffset += rows.Count;

            string partition = "$" + rowOffset.ToString();

            var task = Task.Factory.StartNew( () =>
            {
                var batch = new TableBatchOperation();

                foreach ( var row in rows )
                {
                    row.PartitionKey = row.PartitionKey + partition;
                    batch.InsertOrReplace( row );
                }

                // submit
                table.ExecuteBatch( batch );

                Trace.TraceInformation( "Inserted batch for partition " + partition );
            } );

            tasks.Add( task );
        }

        await Task.WhenAll( tasks );
    }

这是我得到的输出:

iisexpress.exe Information: 0 : Elapsed time to create record rows: 00:00:00.0719448
iisexpress.exe Information: 0 : Inserting rows
iisexpress.exe Information: 0 : Inserted batch for partition $100
...
iisexpress.exe Information: 0 : Successfully inserted 28000 rows into table test
iisexpress.exe Information: 0 : Elapsed time: 00:01:07.1398928

这比我的其他应用程序快一点,超过460 ROPS.这仍然是不可接受的.再次在此测试中,我的CPU(8个逻辑处理器)几乎被淘汰,磁盘访问几乎空闲.

我不知道出了什么问题.

更新5

一轮又一轮的调整和调整已经取得了一些进展,但我只是不能比500-700(ish)ROPS更快地进行批量InsertOrReplace操作(批量为100).

此测试在Azure云中使用一个小实例(或两个)完成.根据下面的评论,我已经接受了这样一个事实,即本地测试充其量只会很慢.

这里有几个例子.每个例子都是它自己的PartitionKey:

Successfully inserted 904 rows into table org1; TraceSource 'w3wp.exe' event
Elapsed time: 00:00:01.3401031; TraceSource 'w3wp.exe' event

Successfully inserted 4130 rows into table org1; TraceSource 'w3wp.exe' event
Elapsed time: 00:00:07.3522871; TraceSource 'w3wp.exe' event

Successfully inserted 28020 rows into table org1; TraceSource 'w3wp.exe' event
Elapsed time: 00:00:51.9319217; TraceSource 'w3wp.exe' event

也许这是我的MSDN Azure帐户有一些性能上限？我不知道.

在这一点上,我想我已经完成了这件事.也许它足够快,可以用于我的目的,或者我可能会遵循不同的道路.

结论

以下所有答案都很好!

对于我的具体问题,我已经能够在小型Azure实例上看到高达2k ROPS的速度,更典型的是大约1k.由于我需要降低成本(因此降低实例大小),这定义了我将能够使用表格的内容.

感谢大家的帮助.

Answer 1

基本概念 - 使用并行来加速这一点.

第1步 - 给你的线程池足够的线程来解决这个问题 - ThreadPool.SetMinThreads(1024,256);

第2步 - 使用分区.我使用guids作为Ids,我使用最后一个字符分成256个独特的分区(实际上我将它们分组为N个子集,在我的情况下为48个分区)

第3步 - 使用任务插入,我使用表refs的对象池

public List<T> InsertOrUpdate(List<T> items)
        {
            var subLists = SplitIntoPartitionedSublists(items);

            var tasks = new List<Task>();

            foreach (var subList in subLists)
            {
                List<T> list = subList;
                var task = Task.Factory.StartNew(() =>
                    {
                        var batchOp = new TableBatchOperation();
                        var tableRef = GetTableRef();

                        foreach (var item in list)
                        {
                            batchOp.Add(TableOperation.InsertOrReplace(item));
                        }

                        tableRef.ExecuteBatch(batchOp);
                        ReleaseTableRef(tableRef);
                    });
                tasks.Add(task);
            }

            Task.WaitAll(tasks.ToArray());

            return items;
        }

private IEnumerable<List<T>> SplitIntoPartitionedSublists(IEnumerable<T> items)
        {
            var itemsByPartion = new Dictionary<string, List<T>>();

            //split items into partitions
            foreach (var item in items)
            {
                var partition = GetPartition(item);
                if (itemsByPartion.ContainsKey(partition) == false)
                {
                    itemsByPartion[partition] = new List<T>();
                }
                item.PartitionKey = partition;
                item.ETag = "*";
                itemsByPartion[partition].Add(item);
            }

            //split into subsets
            var subLists = new List<List<T>>();
            foreach (var partition in itemsByPartion.Keys)
            {
                var partitionItems = itemsByPartion[partition];
                for (int i = 0; i < partitionItems.Count; i += MaxBatch)
                {
                    subLists.Add(partitionItems.Skip(i).Take(MaxBatch).ToList());
                }
            }

            return subLists;
        }

        private void BuildPartitionIndentifiers(int partitonCount)
        {
            var chars = new char[] { '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'a', 'b', 'c', 'd', 'e', 'f' }.ToList();
            var keys = new List<string>();

            for (int i = 0; i < chars.Count; i++)
            {
                var keyA = chars[i];
                for (int j = 0; j < chars.Count; j++)
                {
                    var keyB = chars[j];
                    keys.Add(string.Concat(keyA, keyB));
                }
            }


            var keySetMaxSize = Math.Max(1, (int)Math.Floor((double)keys.Count / ((double)partitonCount)));
            var keySets = new List<List<string>>();

            if (partitonCount > keys.Count)
            {
                partitonCount = keys.Count;
            }

            //Build the key sets
            var index = 0;
            while (index < keys.Count)
            {
                var keysSet = keys.Skip(index).Take(keySetMaxSize).ToList();
                keySets.Add(keysSet);
                index += keySetMaxSize;
            }

            //build the lookups and datatable for each key set
            _partitions = new List<string>();
            for (int i = 0; i < keySets.Count; i++)
            {
                var partitionName = String.Concat("subSet_", i);
                foreach (var key in keySets[i])
                {
                    _partitionByKey[key] = partitionName;
                }
                _partitions.Add(partitionName);
            }

        }

        private string GetPartition(T item)
        {
            var partKey = item.Id.ToString().Substring(34,2);
            return _partitionByKey[partKey];
        }

        private string GetPartition(Guid id)
        {
            var partKey = id.ToString().Substring(34, 2);
            return _partitionByKey[partKey];
        }

        private CloudTable GetTableRef()
        {
            CloudTable tableRef = null;
            //try to pop a table ref out of the stack
            var foundTableRefInStack = _tableRefs.TryPop(out tableRef);
            if (foundTableRefInStack == false)
            {
                //no table ref available must create a new one
                var client = _account.CreateCloudTableClient();
                client.RetryPolicy = new ExponentialRetry(TimeSpan.FromSeconds(1), 4);
                tableRef = client.GetTableReference(_sTableName);
            }

            //ensure table is created
            if (_bTableCreated != true)
            {
                tableRef.CreateIfNotExists();
                _bTableCreated = true;
            }

            return tableRef;
        }

结果 - 最大存储帐户为19-22kops

如果你对完整的资源感兴趣,请打我

需要什么？使用多个存储帐户!

这是几个月的试验和错误,测试,在桌子上敲打我的头.我真的希望它有所帮助.

Answer 2

好的,3号回答了一个魅力？

http://blogs.msdn.com/b/windowsazurestorage/archive/2010/11/06/how-to-get-most-out-of-windows-azure-tables.aspx

一些事情 - 存储模拟器 - 来自一位认真研究它的朋友.

"一切都在单个数据库中击中一个表(更多分区不会影响任何东西).每个表插入操作至少有3个sql操作.每个批处理都在一个事务中.根据事务隔离级别,这些批次将受到限制能够并行执行.

由于sql server行为,串行批处理应该比单个插入更快.(单个插入本质上是很少的事务,每个事务都刷新到磁盘,而真正的事务作为一个组刷新到磁盘)."

使用多个分区的IE不会影响仿真器的性能,而它会影响实际的天蓝色存储.

还可以启用日志记录并稍微检查日志 - c:\ users\username\appdata\local\developmentstorage

批量大小100似乎提供最好的真实性能,关闭naggle,关闭期望100,加强连接限制.

还要确保你不会意外地插入重复项,这将导致错误并减慢一切方式.

并测试真实存储.有一个相当不错的库可以为你处理大部分内容 - http://www.nuget.org/packages/WindowsAzure.StorageExtensions/,只是确保你实际上在添加中调用ToList,因为它不会真的执行直到枚举.此外,该库使用dynamictableentity,因此序列化有一个小的性能,但它确实允许您使用没有TableEntity内容的纯POCO对象.

~JT

Answer 3

在经历了大量的痛苦,实验之后,最终能够获得单表分区的最佳吞吐量(每秒2,000多次批量写入操作),以及使用Azure表存储获得更高的存储帐户吞吐量(每秒3,500多次批量写入操作).我尝试了所有不同的方法,但是以编程方式设置.net连接限制(我尝试了配置示例,但对我不起作用)解决了问题(基于Microsoft提供的白皮书),如下所示:

ServicePoint tableServicePoint = ServicePointManager
    .FindServicePoint(_StorageAccount.TableEndpoint);

//This is a notorious issue that has affected many developers. By default, the value 
//for the number of .NET HTTP connections is 2.
//This implies that only 2 concurrent connections can be maintained. This manifests itself
//as "underlying connection was closed..." when the number of concurrent requests is
//greater than 2.

tableServicePoint.ConnectionLimit = 1000;

对于每个存储帐户获得20K +批量写入操作的任何其他人,请分享您的经验.

Answer 4

为了更有趣,这是一个新的答案 - 隔离的独立测试,它为生产中的写入性能提供了一些惊人的数字,并且在避免IO阻塞和连接管理方面做得更好.我很有兴趣看到这对你有用,因为我们的写入速度很可笑(> 7kps).

webconfig

 <system.net>
    <connectionManagement>
      <add address="*" maxconnection="48"/>
    </connectionManagement>
  </system.net>

对于测试,我使用的是基于卷的参数,因此像25000个项目,24个分区,批量大小100似乎总是最好的,并且ref计数为20.这是使用TPL数据流(http://www.nuget.org) /packages/Microsoft.Tpl.Dataflow/)为BufflerBlock提供了一个很好的等待线程安全表引用拉.

public class DyanmicBulkInsertTestPooledRefsAndAsynch : WebTest, IDynamicWebTest
{
    private int _itemCount;
    private int _partitionCount;
    private int _batchSize;
    private List<TestTableEntity> _items;
    private GuidIdPartitionSplitter<TestTableEntity> _partitionSplitter;
    private string _tableName;
    private CloudStorageAccount _account;
    private CloudTableClient _tableClient;
    private Dictionary<string, List<TestTableEntity>> _itemsByParition;
    private int _maxRefCount;
    private BufferBlock<CloudTable> _tableRefs;


    public DyanmicBulkInsertTestPooledRefsAndAsynch()
    {
        Properties = new List<ItemProp>();    
        Properties.Add(new ItemProp("ItemCount", typeof(int)));
        Properties.Add(new ItemProp("PartitionCount", typeof(int)));
        Properties.Add(new ItemProp("BatchSize", typeof(int)));
        Properties.Add(new ItemProp("MaxRefs", typeof(int)));


    }

    public List<ItemProp> Properties { get; set; }

    public void SetProps(Dictionary<string, object> propValuesByPropName)
    {
        _itemCount = (int)propValuesByPropName["ItemCount"];
        _partitionCount = (int)propValuesByPropName["PartitionCount"];
        _batchSize = (int)propValuesByPropName["BatchSize"];
        _maxRefCount = (int)propValuesByPropName["MaxRefs"];
    }

    protected override void SetupTest()
    {
        base.SetupTest();

        ThreadPool.SetMinThreads(1024, 256);
        ServicePointManager.DefaultConnectionLimit = 256;
        ServicePointManager.UseNagleAlgorithm = false;
        ServicePointManager.Expect100Continue = false;


        _account = CloudStorageAccount.Parse(CloudConfigurationManager.GetSetting("DataConnectionString"));
        _tableClient = _account.CreateCloudTableClient();
        _tableName = "testtable" + new Random().Next(100000);

        //create the refs
        _tableRefs = new BufferBlock<CloudTable>();
        for (int i = 0; i < _maxRefCount; i++)
        {
            _tableRefs.Post(_tableClient.GetTableReference(_tableName));
        }

        var tableRefTask = GetTableRef();
        tableRefTask.Wait();
        var tableRef = tableRefTask.Result;

        tableRef.CreateIfNotExists();
        ReleaseRef(tableRef);

        _items = TestUtils.GenerateTableItems(_itemCount);
        _partitionSplitter = new GuidIdPartitionSplitter<TestTableEntity>();
        _partitionSplitter.BuildPartitions(_partitionCount);

        _items.ForEach(o =>
            {
                o.ETag = "*";
                o.Timestamp = DateTime.Now;
                o.PartitionKey = _partitionSplitter.GetPartition(o);
            });

        _itemsByParition = _partitionSplitter.SplitIntoPartitionedSublists(_items);
    }

    private async Task<CloudTable> GetTableRef()
    {
        return await _tableRefs.ReceiveAsync();            
    }

    private void ReleaseRef(CloudTable tableRef)
    {
        _tableRefs.Post(tableRef);
    }

    protected override void ExecuteTest()
    {
        Task.WaitAll(_itemsByParition.Keys.Select(parition => Task.Factory.StartNew(() => InsertParitionItems(_itemsByParition[parition]))).ToArray());
    }

    private void InsertParitionItems(List<TestTableEntity> items)
    {

        var tasks = new List<Task>();

        for (int i = 0; i < items.Count; i += _batchSize)
        {
            int i1 = i;

            var task = Task.Factory.StartNew(async () =>
            {
                var batchItems = items.Skip(i1).Take(_batchSize).ToList();

                if (batchItems.Select(o => o.PartitionKey).Distinct().Count() > 1)
                {
                    throw new Exception("Multiple partitions batch");
                }

                var batchOp = new TableBatchOperation();
                batchItems.ForEach(batchOp.InsertOrReplace);   

                var tableRef = GetTableRef.Result();
                tableRef.ExecuteBatch(batchOp);
                ReleaseRef(tableRef);
            });

            tasks.Add(task);

        }

        Task.WaitAll(tasks.ToArray());


    }

    protected override void CleanupTest()
    {
        var tableRefTask = GetTableRef();
        tableRefTask.Wait();
        var tableRef = tableRefTask.Result;
        tableRef.DeleteIfExists();
        ReleaseRef(tableRef);
    }

我们目前正在开发一个可以处理多个存储帐户的版本,希望能够获得一些疯狂的速度.此外,我们在8个核心虚拟机上运行这些大型数据集,但是对于新的非阻塞IO,它应该在有限的vm上运行良好.祝好运!

 public class SimpleGuidIdPartitionSplitter<T> where T : IUniqueId
{
    private ConcurrentDictionary<string, string> _partitionByKey = new ConcurrentDictionary<string, string>();
    private List<string> _partitions;
    private bool _bPartitionsBuilt;

    public SimpleGuidIdPartitionSplitter()
    {

    }

    public void BuildPartitions(int iPartCount)
    {
        BuildPartitionIndentifiers(iPartCount);
    }

    public string GetPartition(T item)
    {
        if (_bPartitionsBuilt == false)
        {
            throw new Exception("Partitions Not Built");
        }

        var partKey = item.Id.ToString().Substring(34, 2);
        return _partitionByKey[partKey];
    }

    public string GetPartition(Guid id)
    {
        if (_bPartitionsBuilt == false)
        {
            throw new Exception("Partitions Not Built");
        }

        var partKey = id.ToString().Substring(34, 2);
        return _partitionByKey[partKey];
    }

    #region Helpers
    private void BuildPartitionIndentifiers(int partitonCount)
    {
        var chars = new char[] { '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'a', 'b', 'c', 'd', 'e', 'f' }.ToList();
        var keys = new List<string>();

        for (int i = 0; i < chars.Count; i++)
        {
            var keyA = chars[i];
            for (int j = 0; j < chars.Count; j++)
            {
                var keyB = chars[j];
                keys.Add(string.Concat(keyA, keyB));
            }
        }


        var keySetMaxSize = Math.Max(1, (int)Math.Floor((double)keys.Count / ((double)partitonCount)));
        var keySets = new List<List<string>>();

        if (partitonCount > keys.Count)
        {
            partitonCount = keys.Count;
        }

        //Build the key sets
        var index = 0;
        while (index < keys.Count)
        {
            var keysSet = keys.Skip(index).Take(keySetMaxSize).ToList();
            keySets.Add(keysSet);
            index += keySetMaxSize;
        }

        //build the lookups and datatable for each key set
        _partitions = new List<string>();
        for (int i = 0; i < keySets.Count; i++)
        {
            var partitionName = String.Concat("subSet_", i);
            foreach (var key in keySets[i])
            {
                _partitionByKey[key] = partitionName;
            }
            _partitions.Add(partitionName);
        }

        _bPartitionsBuilt = true;
    }
    #endregion
}



internal static List<TestTableEntity> GenerateTableItems(int count)
        {
            var items = new List<TestTableEntity>();
            var random = new Random();

            for (int i = 0; i < count; i++)
            {
                var itemId = Guid.NewGuid();

                items.Add(new TestTableEntity()
                {
                    Id = itemId,
                    TestGuid = Guid.NewGuid(),
                    RowKey = itemId.ToString(),
                    TestBool = true,
                    TestDateTime = DateTime.Now,
                    TestDouble = random.Next() * 1000000,
                    TestInt = random.Next(10000),
                    TestString = Guid.NewGuid().ToString(),
                });
            }

            var dupRowKeys = items.GroupBy(o => o.RowKey).Where(o => o.Count() > 1).Select(o => o.Key).ToList();
            if (dupRowKeys.Count > 0)
            {
                throw  new Exception("Dupicate Row Keys");
            }

            return items;
        }

还有一件事 - 你的时机以及框架如何受到影响指向这个http://blogs.msdn.com/b/windowsazurestorage/archive/2013/08/08/net-clients-encountering-port-exhaustion-after-安装-kb2750149-OR-kb2805227.aspx