前段时间看到园子里面有同学在用Parallel进行批量插入数据库。后面也有很多同学针对这一事件给出了自己的看法和见解。我在这里不评论内容的好坏,至少能将自己东西总结分享这个是要靠勇气和毅力。
闲话少说,我在最近看崔鹏飞的github的时候,发现他对这块也做了一定的总结,那么我就他这块进行板书与展示。案例是怎么回事呢?话说我有一个公司,里面需要统计一下总收入,另外有一个公司被我收购了,我一起计算总收入。当一天我收购了N个公司,计算总收入的时候,我们采用并行计算。
1 internal class Company 2 { 3 public decimal TotalIncome; 4 5 public Company Merge(Company that) 6 { 7 Calc(); 8 TotalIncome += that.TotalIncome; 9 return this; 10 } 11 12 /// <summary> 13 /// 复杂运算 14 /// </summary> 15 private void Calc() 16 { 17 //TODO:省略500字 18 } 19 }
首先我们想到的是采用直接累加就行了吧,这是所谓的线性预算。
/// <summary>
/// 线性运行
/// </summary>
/// <param name="bigCompany"></param>
/// <param name="smallCompanies"></param>
/// <returns></returns>
private static Company LinearMerge(Company bigCompany, IEnumerable<Company> smallCompanies)
{
foreach (Company smallCompany in smallCompanies)
{
bigCompany.Merge(smallCompany);
}
return bigCompany;
}
采用线性运算,毫无疑问结果是正确的。但是,如果的N大一点,例如30000000个,可能就要花一点时间了。
那么是否我们可以采用并行处理呢?OK,直接上代码。
1 /// <summary> 2 /// 并行处理 3 /// </summary> 4 /// <param name="bigCompany"></param> 5 /// <param name="smallCompanies"></param> 6 /// <returns></returns> 7 private static Company ParallelMerge(Company bigCompany, IEnumerable<Company> smallCompanies) 8 { 9 Parallel.ForEach(smallCompanies, smallCompany => bigCompany.Merge(smallCompany)); 10 return bigCompany; 11 }
时间很快,但是结果呢?结果和上面线性的一致么?
那么我如果在并行的基础上面加一把锁呢,保证每次独占资源。
1 /// <summary> 2 /// 并行加锁 3 /// </summary> 4 /// <param name="bigCompany"></param> 5 /// <param name="smallCompanies"></param> 6 /// <returns></returns> 7 private static Company ParallelMergeLock(Company bigCompany, IEnumerable<Company> smallCompanies) 8 { 9 var obj = new object(); 10 Parallel.ForEach(smallCompanies, smallCompany => 11 { 12 lock (obj) 13 { 14 bigCompany.Merge(smallCompany); 15 } 16 }); 17 return bigCompany; 18 }
毫无疑问,结果也是正确的,那么耗时可能我们就要关心了。那么耗时究竟怎么样呢?
我们可以采用函数式处理嘛。
1 /// <summary> 2 /// 函数式合并 3 /// </summary> 4 /// <param name="bigCompany"></param> 5 /// <param name="smallCompanies"></param> 6 /// <returns></returns> 7 private static Company FunctionalMerger(Company bigCompany, IEnumerable<Company> smallCompanies) 8 { 9 return smallCompanies.Aggregate(bigCompany, (buyer, seller) => buyer.Merge(seller)); 10 }
那么我们在在函数式的基础上面进行并行化处理呢?
1 /// <summary> 2 /// 函数式的并行化 3 /// </summary> 4 /// <param name="bigCompany"></param> 5 /// <param name="smallCompanies"></param> 6 /// <returns></returns> 7 private static Company FunctionParallelMerge(Company bigCompany, IEnumerable<Company> smallCompanies) 8 { 9 return smallCompanies.AsParallel().Aggregate(() => new Company(), (shell, smallCompany) => shell.Merge(smallCompany), (shell1, shell2) => shell1.Merge(shell2), bigCompany.Merge); 10 }
上面提出了一些问题,这里我们用实际的测试数据查看。
测试代码
1 private static IEnumerable<Company> GenerateSmallCompanies() 2 { 3 return Enumerable.Range(0, 30000000).Select(number => new Company { TotalIncome = number }).ToArray(); 4 } 5 6 private static void PrintMergeResult(Func<Company, IEnumerable<Company>, Company> mergeMethod, string funcApproach) 7 { 8 var stopWatch = new Stopwatch(); 9 stopWatch.Start(); 10 var mergeResult = mergeMethod(new Company { TotalIncome = 1000000 }, m_SmallCompanies); 11 stopWatch.Stop(); 12 Console.WriteLine("{0}:{1} Time:{2}", funcApproach, mergeResult.TotalIncome, stopWatch.ElapsedMilliseconds); 13 } 14 15 private static void TryAll() 16 { 17 Console.WriteLine("============================"); 18 PrintMergeResult(LinearMerge, "简单直接 "); 19 PrintMergeResult(ParallelMerge, "错误并行 "); 20 PrintMergeResult(ParallelMergeLock, "加锁并行 "); 21 Console.WriteLine("***********"); 22 PrintMergeResult(FunctionalMerge,"函数式合并 "); 23 PrintMergeResult(FunctionParallelMerge, "函数式并行合并 "); 24 } 25 26 27 private static readonly IEnumerable<Company> m_SmallCompanies = GenerateSmallCompanies(); 28 static void Main() 29 { 30 Console.WriteLine("测试数据30000000个"); 31 for (int i = 0; i < 5; i++) 32 { 33 TryAll(); 34 } 35 Console.ReadKey(); 36 }
测试结果如下:
按照理论情况,错误并行应该比直接更快,但是不知道我机器(CPU AMD)上面出现这样的情况,其他情况还算正常。在另一台计算机(CPU Intel)上面运行测试,数据如下: