Why is using AsParallel() slower than foreach in t

问题:

I am extracting data from excel that is in this format

 product1   | unnamedcol2 | product2  | unnamedcol4 | product3  | unnamedcol6 |
-------------------------------------------------------------------------------
 @1foo      |        1.10 | @1foo     |         0.3 | @1foo     |         0.3
 @2foo      |        1.00 | @2foo     |           2 | @2foo     |
 @3foo      |        1.52 | @3foo     |        2.53 | @3foo     |
 @4foo      |        1.47 |           |             | @4foo     |        1.31
 @5foo      |        1.49 |           |             | @5foo     |        1.31

The file uses all 255 fields. Using dapper-dot-net i get the data through this code

IEnumerable<IDictionary<string, object>> excelDataRaw =
                conn.Query(string.Format("select * from {0}", table)).Cast<IDictionary<string, object>>();

I pass this data to these test methods. The data is returned as an IEnumerable of IDictionaries where each key is a product and each value is an IDictionary where each key is a value from the product column and the corresponding value is a value from unnamedcol that is to the right of the product column.

var excelDataRefined = new List<IDictionary<string, IDictionary<string, decimal>>>();
excelDataRefined.Add(new Dictionary<string, IDictionary<string, decimal>>());
excelDataRefined[0].Add( "product", new Dictionary<string, decimal>());
excelDataRefined[0]["product"].Add("@1foo", 1.1m);

The methods:

private static Dictionary<string, IDictionary<string, decimal>> Benchmark_foreach(IEnumerable<IDictionary<string, object>> excelDataRaw)
{
    Console.WriteLine("1. Using foreach");
    var watch = new Stopwatch();
    watch.Start();

    List<string> headers = excelDataRaw.Select(dictionary => dictionary.Keys).First().ToList();
    bool isEven = false;
    List<string> products = headers.Where(h => isEven = !isEven).ToList();
    var dates = new List<IEnumerable<object>>();
    var prices = new List<IEnumerable<object>>();

    foreach (string field in headers)
    {
        string product1 = field;
        if (headers.IndexOf(field) % 2 == 0)
        {
            dates.Add(
                excelDataRaw.AsParallel().AsOrdered().Select(col => col[product1]).Where(row => row != null));
        }

        if (headers.IndexOf(field) % 2 == 1)
        {
            prices.Add(
                excelDataRaw.AsParallel().AsOrdered().Select(col => col[product1] ?? 0m).Take(dates.Last().Count()));
        }
    }

    watch.Stop();
    Console.WriteLine("Rearange the data in: {0}s", watch.Elapsed.TotalSeconds);
    watch.Restart();

    var excelDataRefined = new Dictionary<string, IDictionary<string, decimal>>();
    foreach (IEnumerable<object> datelist in dates)
    {
        decimal num;
        IEnumerable<object> datelist1 = datelist;
        IEnumerable<object> pricelist =
            prices[dates.IndexOf(datelist1)].Select(value => value ?? 0m).Where(
                content => decimal.TryParse(content.ToString(), out num));
        Dictionary<string, decimal> dict =
            datelist1.Zip(pricelist, (k, v) => new { k, v }).ToDictionary(
                x => (string)x.k, x => decimal.Parse(x.v.ToString()));

        if (!excelDataRefined.ContainsKey(products[dates.IndexOf(datelist1)]))
        {
            excelDataRefined.Add(products[dates.IndexOf(datelist1)], dict);
        }
    }

    watch.Stop();
    Console.WriteLine("Zipped the data in: {0}s", watch.Elapsed.TotalSeconds);

    return excelDataRefined;
}

private static Dictionary<string, IDictionary<string, decimal>> Benchmark_AsParallel(IEnumerable<IDictionary<string, object>> excelDataRaw)
{
    Console.WriteLine("2. Using AsParallel().AsOrdered().ForAll");
    var watch = new Stopwatch();
    watch.Start();

    List<string> headers = excelDataRaw.Select(dictionary => dictionary.Keys).First().ToList();
    bool isEven = false;
    List<string> products = headers.Where(h => isEven = !isEven).ToList();
    var dates = new List<IEnumerable<object>>();
    var prices = new List<IEnumerable<object>>();

    headers.AsParallel().AsOrdered().ForAll(
        field =>
        dates.Add(
            excelDataRaw.AsParallel().AsOrdered().TakeWhile(x => headers.IndexOf(field) % 2 == 0).Select(
                col => col[field]).Where(row => row != null).ToList()));
    headers.AsParallel().AsOrdered().ForAll(
        field =>
        prices.Add(
            excelDataRaw.AsParallel().AsOrdered().TakeWhile(x => headers.IndexOf(field) % 2 == 1).Select(
                col => col[field] ?? 0m).Take(256).ToList()));
    dates.RemoveAll(x => x.Count() == 0);
    prices.RemoveAll(x => x.Count() == 0);

    watch.Stop();
    Console.WriteLine("Rearange the data in: {0}s", watch.Elapsed.TotalSeconds);
    watch.Restart();

    var excelDataRefined = new Dictionary<string, IDictionary<string, decimal>>();
    foreach (IEnumerable<object> datelist in dates)
    {
        decimal num;
        IEnumerable<object> datelist1 = datelist;
        IEnumerable<object> pricelist =
            prices[dates.IndexOf(datelist1)].Select(value => value ?? 0m).Where(
                content => decimal.TryParse(content.ToString(), out num));
        Dictionary<string, decimal> dict =
            datelist1.Zip(pricelist, (k, v) => new { k, v }).ToDictionary(
                x => (string)x.k, x => decimal.Parse(x.v.ToString()));

        if (!excelDataRefined.ContainsKey(products[dates.IndexOf(datelist1)]))
        {
            excelDataRefined.Add(products[dates.IndexOf(datelist1)], dict);
        }
    }

    watch.Stop();
    Console.WriteLine("Zipped the data in: {0}s", watch.Elapsed.TotalSeconds);

    return excelDataRefined;
}

private static Dictionary<string, IDictionary<string, decimal>> Benchmark_ForEach(IEnumerable<IDictionary<string, object>> excelDataRaw)
{
    Console.WriteLine("3. Using ForEach");
    var watch = new Stopwatch();
    watch.Start();

    List<string> headers = excelDataRaw.Select(dictionary => dictionary.Keys).First().ToList();
    bool isEven = false;
    List<string> products = headers.Where(h => isEven = !isEven).ToList();
    var dates = new List<IEnumerable<object>>();
    var prices = new List<IEnumerable<object>>();

    headers.ForEach(
        field =>
        dates.Add(
            excelDataRaw.TakeWhile(x => headers.IndexOf(field) % 2 == 0).Select(col => col[field]).Where(
                row => row != null).ToList()));
    headers.ForEach(
        field =>
        prices.Add(
            excelDataRaw.TakeWhile(x => headers.IndexOf(field) % 2 == 1).Select(col => col[field] ?? 0m).
            Take(256).ToList()));
    dates.RemoveAll(x => x.Count() == 0);
    prices.RemoveAll(x => x.Count() == 0);

    watch.Stop();
    Console.WriteLine("Rearange the data in: {0}s", watch.Elapsed.TotalSeconds);
    watch.Restart();

    var excelDataRefined = new Dictionary<string, IDictionary<string, decimal>>();
    foreach (IEnumerable<object> datelist in dates)
    {
        decimal num;
        IEnumerable<object> datelist1 = datelist;
        IEnumerable<object> pricelist =
            prices[dates.IndexOf(datelist1)].Select(value => value ?? 0m).Where(
                content => decimal.TryParse(content.ToString(), out num));
        Dictionary<string, decimal> dict =
            datelist1.Zip(pricelist, (k, v) => new { k, v }).ToDictionary(
                x => (string)x.k, x => decimal.Parse(x.v.ToString()));

        if (!excelDataRefined.ContainsKey(products[dates.IndexOf(datelist1)]))
        {
            excelDataRefined.Add(products[dates.IndexOf(datelist1)], dict);
        }
    }

    watch.Stop();
    Console.WriteLine("Zipped the data in: {0}s", watch.Elapsed.TotalSeconds);

    return excelDataRefined;
}

• Benchmark_foreach needs app. 3,5s to rearrange and 3s to zip the data.
• Benchmark_AsParallel needs app. 12s to rearrange and 0,005s to zip the data.
• Benchmark_ForEach needs app. 16s to rearrange and 0,005s to zip the data.

Why does it behave like this? I expected AsParallel to be the fastest because it executes in parallel instead of sequential. Ho do i optimize this?

回答1:

In order for parallel computation to happen you have to have multiple processors or cores, otherwise you are just queueing up tasks in the threadpool waiting for the CPU. I.e. AsParallel on a single core machine is sequential plus the overhead of threadpool and thread context switch. Even on a two core machine, you may not get both cores, since lots of other things are running on the same machine.

Really .AsParallel() only becomes useful if you have long running tasks with blocking operations (I/O) where the OS can suspend the blocking thread and let another one run.

回答2:

There is an overhead to creating the additional threads and managing the work loads for each of those threads. If you have a limited amount of work, the overhead to create the extra threads, task switch between the threads, work steal and re distribute between the threads, etc. may outweigh the gains you get by parallelizing the work in the first place. You may want to profile your application to find out if you are really CPU bound when running with a single process. If not, it is going to be best to keep it single threaded and your bottleneck becomes IO which is not as easy to parallelize.

A couple additional recommendations: You are going to see a performance penalty by using AsOrdered and TakeWhile because they both need to synchronize back to the originating thread. Consider profiling without requiring ordering and see if that offers any performance improvement.

Also, consider using a ConcurrentDictionary rather than the standard generic dictionary to avoid concurrency issues when adding items.

回答3:

In Benchmark_AsParallel and Benchmark_ForEach you perform 2n operation in Benchmark_foreach n.