Why is Parallel.ForEach much faster then AsParallel().ForAll() even though MSDN suggests otherwise?
Solution 1:
This problem is pretty debuggable, an uncommon luxury when you have problems with threads. Your basic tool here is the Debug > Windows > Threads debugger window. Shows you the active threads and gives you a peek at their stack trace. You'll easily see that, once it gets slow, that you'll have dozens of threads active that are all stuck. Their stack trace all look the same:
mscorlib.dll!System.Threading.Monitor.Wait(object obj, int millisecondsTimeout, bool exitContext) + 0x16 bytes
mscorlib.dll!System.Threading.Monitor.Wait(object obj, int millisecondsTimeout) + 0x7 bytes
mscorlib.dll!System.Threading.ManualResetEventSlim.Wait(int millisecondsTimeout, System.Threading.CancellationToken cancellationToken) + 0x182 bytes
mscorlib.dll!System.Threading.Tasks.Task.SpinThenBlockingWait(int millisecondsTimeout, System.Threading.CancellationToken cancellationToken) + 0x93 bytes
mscorlib.dll!System.Threading.Tasks.Task.InternalRunSynchronously(System.Threading.Tasks.TaskScheduler scheduler, bool waitForCompletion) + 0xba bytes
mscorlib.dll!System.Threading.Tasks.Task.RunSynchronously(System.Threading.Tasks.TaskScheduler scheduler) + 0x13 bytes
System.Core.dll!System.Linq.Parallel.SpoolingTask.SpoolForAll<ConsoleApplication1.DirWithSubDirs,int>(System.Linq.Parallel.QueryTaskGroupState groupState, System.Linq.Parallel.PartitionedStream<ConsoleApplication1.DirWithSubDirs,int> partitions, System.Threading.Tasks.TaskScheduler taskScheduler) Line 172 C#
// etc..
Whenever you see something like this, you should immediately think fire-hose problem. Probably the third-most common bug with threads, after races and deadlocks.
Which you can reason out, now that you know the cause, the problem with the code is that every thread that completes adds N more threads. Where N is the average number of sub-directories in a directory. In effect, the number of threads grows exponentially, that's always bad. It will only stay in control if N = 1, that of course never happens on an typical disk.
Do beware that, like almost any threading problem, that this misbehavior tends to repeat poorly. The SSD in your machine tends to hide it. So does the RAM in your machine, the program might well complete quickly and trouble-free the second time you run it. Since you'll now read from the file system cache instead of the disk, very fast. Tinkering with ThreadPool.SetMinThreads() hides it as well, but it cannot fix it. It never fixes any problem, it only hides them. Because no matter what happens, the exponential number will always overwhelm the set minimum number of threads. You can only hope that it completes finishing iterating the drive before that happens. Idle hope for a user with a big drive.
The difference between ParallelEnumerable.ForAll() and Parallel.ForEach() is now perhaps also easily explained. You can tell from the stack trace that ForAll() does something naughty, the RunSynchronously() method blocks until all the threads are completed. Blocking is something threadpool threads should not do, it gums up the thread pool and won't allow it to schedule the processor for another job. And has the effect you observed, the thread pool is quickly overwhelmed with threads that are waiting on the N other threads to complete. Which isn't happening, they are waiting in the pool and are not getting scheduled because there are already so many of them active.
This is a deadlock scenario, a pretty common one, but the threadpool manager has a workaround for it. It watches the active threadpool threads and steps in when they don't complete in a timely manner. It then allows an extra thread to start, one more than the minimum set by SetMinThreads(). But not more then the maximum set by SetMaxThreads(), having too many active tp threads is risky and likely to trigger OOM. This does solve the deadlock, it gets one of the ForAll() calls to complete. But this happens at a very slow rate, the threadpool only does this twice a second. You'll run out of patience before it catches up.
Parallel.ForEach() doesn't have this problem, it doesn't block so doesn't gum up the pool.
Seems to be the solution, but do keep in mind that your program is still fire-hosing the memory of your machine, adding ever more waiting tp threads to the pool. This can crash your program as well, it just isn't as likely because you have a lot of memory and the threadpool doesn't use a lot of it to keep track of a request. Some programmers however accomplish that as well.
The solution is a very simple one, just don't use threading. It is harmful, there is no concurrency when you have only one disk. And it does not like being commandeered by multiple threads. Especially bad on a spindle drive, head seeks are very, very slow. SSDs do it a lot better, it however still takes an easy 50 microseconds, overhead that you just don't want or need. The ideal number of threads to access a disk that you can't otherwise expect to be cached well is always one.
Solution 2:
The first thing to note is that you are trying to parallelise an IO-bound operation, which will distort the timings significantly.
The second thing to note is the nature of the parallelised tasks: You are recursively descending a directory tree. If you create multiple threads to do this, each thread is likely to be accessing a different part of the disk simultaneously - which will cause the disk read head to be jumping all over the place and slowing things down considerably.
Try changing your test to create an in-memory tree, and access that with multiple threads instead. Then you will be able to compare the timings properly without the results being distorted beyond all usefulness.
Additionally, you may be creating a great number of threads, and they will (by default) be threadpool threads. Having a great number of threads will actually slow things down when they exceed the number of processor cores.
Also note that when you exceed the thread pool minimum threads (defined by ThreadPool.GetMinThreads()), a delay is introduced by the thread pool manager between each new threadpool thread creation. (I think this is around 0.5s per new thread).
Also, if the number of threads exceeds the value returned by ThreadPool.GetMaxThreads(), the creating thread will block until one of the other threads has exited. I think this is likely to be happening.
You can test this hypothesis by calling ThreadPool.SetMaxThreads() and ThreadPool.SetMinThreads() to increase these values, and see if it makes any difference.
(Finally, note that if you are really trying to recursively descend from C:\, you will almost certainly get an IO exception when it reaches a protected OS folder.)
NOTE: Set the max/min threadpool threads like this:
ThreadPool.SetMinThreads(4000, 16);
ThreadPool.SetMaxThreads(4000, 16);
Follow Up
I have tried your test code with the threadpool thread counts set as described above, with the following results (not run on the whole of my C:\ drive, but on a smaller subset):
- Mode 1 took 06.5 seconds.
- Mode 2 took 15.7 seconds.
- Mode 3 took 16.4 seconds.
This is in line with my expectations; adding a load of threading to do this actually makes it slower than single-threaded, and the two parallel approaches take roughly the same time.
In case anyone else wants to investigate this, here's some determinative test code (the OP's code is not reproducible because we don't know his directory structure).
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Threading.Tasks;
namespace Demo
{
internal class Program
{
private static DirWithSubDirs RootDir;
private static void Main()
{
Console.WriteLine("Loading file system into memory...");
RootDir = new DirWithSubDirs("Root", 4, 4);
Console.WriteLine("Done");
//ThreadPool.SetMinThreads(4000, 16);
//ThreadPool.SetMaxThreads(4000, 16);
var w = Stopwatch.StartNew();
ThisIsARecursiveFunctionInMemory(RootDir);
Console.WriteLine("Elapsed seconds: " + w.Elapsed.TotalSeconds);
Console.ReadKey();
}
public static void ThisIsARecursiveFunctionInMemory(DirWithSubDirs currentDirectory)
{
var depth = currentDirectory.Path.Count(t => t == '\\');
Console.WriteLine(depth + ": " + currentDirectory.Path);
var children = currentDirectory.SubDirs;
//Edit this mode to switch what way of parallelization it should use
int mode = 3;
switch (mode)
{
case 1:
foreach (var child in children)
{
ThisIsARecursiveFunctionInMemory(child);
}
break;
case 2:
children.AsParallel().ForAll(t =>
{
ThisIsARecursiveFunctionInMemory(t);
});
break;
case 3:
Parallel.ForEach(children, t =>
{
ThisIsARecursiveFunctionInMemory(t);
});
break;
default:
break;
}
}
}
internal class DirWithSubDirs
{
public List<DirWithSubDirs> SubDirs = new List<DirWithSubDirs>();
public String Path { get; private set; }
public DirWithSubDirs(String path, int width, int depth)
{
this.Path = path;
if (depth > 0)
for (int i = 0; i < width; ++i)
SubDirs.Add(new DirWithSubDirs(path + "\\" + i, width, depth - 1));
}
}
}
Solution 3:
The Parallel.For and .ForEach methods are implemented internally as equivalent to running iterations in Tasks, e.g. that a loop like:
Parallel.For(0, N, i =>
{
DoWork(i);
});
is equivalent to:
var tasks = new List<Task>(N);
for(int i=0; i<N; i++)
{
tasks.Add(Task.Factory.StartNew(state => DoWork((int)state), i));
}
Task.WaitAll(tasks.ToArray());
And from the perspective of every iteration potentially running in parallel with every other iteration, this is an ok mental model, but does not happen in reality. Parallel, in fact, does not necessarily use one Task per iteration, as that is significantly more overhead than is necessary. Parallel.ForEach tries to use the minimum number of tasks necessary to complete the loop as fast as possible. It spins up tasks as threads become available to process those tasks, and each of those tasks participates in a management scheme (I think its called chunking): A task asks for multiple iterations to be done, gets them, and then processes that work, and then goes back for more. The chunk sizes vary based the number of tasks participating, the load on the machine, etc.
PLINQ’s .AsParallel() has a different implementation, but it ‘can’ still similarly fetch multiple iterations into a temporary store, do the calculations in a thread (but not as a task), and put the query results into a small buffer. (You get something based on ParallelQuery, and then further .Whatever() functions bind to an alternative set of extension methods that provide parallel implementations).
So now that we have a small idea of how these two mechanisms work, I will try to provide an answer to your original question:
So why is .AsParallel() slower than Parallel.ForEach? The reason stems from the following. Tasks (or their equivalent implementation here) do NOT block on I/O-like calls. They ‘await’ and free up the CPU to do something else. But (quoting C# nutshell book): “PLINQ cannot perform I/O-bound work without blocking threads”. The calls are synchronous. They were written with the intention that you increase the degree of parallelism if (and ONLY if) you are doing such things as downloading web pages per task that do not hog CPU time.
And the reason why your function calls are exactly analogous to I/O bound calls is this: One of your threads (call it T) blocks and does nothing until all of its child threads have finished, which can be a slow process here. T itself is not CPU-intensive while it waits for the children to unblock, it is doing nothing but waiting. Hence it is identical to a typical I/O bound function call.