Recently I have been doing some tests with C# multi-threaded performance, below are the tests and the result:
Note that all the tests run three times for an array with the size of 1000000, the first run on the main thread, the second using Parallel.For, the third uses a custom parallel function where run tasks equal to the number of Environment.ProcessorCount and the load are distributed on the tasks as evenly as possible. I run the tests on i7-10700 which comes with 8 cores and 16 threads. I'm using .NET 6 and Visual Studio 2022.
My test cases:
- Simple Write: write the index value for every element in the array
- Write Random: write a random value for every element in the array
- Date struct: create a Date object for every element in the array (where Date is struct)
- Date class: create a Date object for every element in the array (where Date is class)
- Lookup: run multiple lookups to fetch the value that will be written to the array element
- Min: find the min value in the array
- Min with wrapper: find the min value in the array, wrap the min values, so they don't share the cache line
My measured times:
| routine | single-thread | multi-thread (parallel for) | multi-thread (custom) |
|---|---|---|---|
| Simple Write | 6.4779 ms | 10.5307 ms | 2.751 ms |
| Write Random | 8.2883 ms | 20.1144 ms | 5.7795 ms |
| Date struct | 19.0835 ms | 76.5366 ms | 11.0556 ms |
| Date class | 67.922 ms | 101.3405 ms | 71.2441 ms |
| Lookup | 4.4336 ms | 2.293 ms | 1.3446 ms |
| Min | 3.4373 ms | 6.9232 ms | 1.2026 ms |
| Min with wrapper | 3.5319 ms | 5.834 ms | 1.3308 ms |
I attached the code to my test cases, my questions are:
- why is the
Parallel.Forrun so poorly? - In tests one and two, I only achieved around 2X the performance, is there a better way to achieve more?
- this is just an observation: from tests three and four we can see that object allocation on the heap doesn't work well with multi-threading.
- test number five is where multi-threading shine, we can get to around 4x the performance of the single-thread, not sure why this is the case.
- shouldn't I be reaching 16x performance, am I doing anything wrong, is there anything I should be aware of?
using System.Diagnostics;
public static class Ext
{
public static T[] Fill<T>(this T[] array, Func<T> cons)
{
for (int i = 0; i < array.Length; i )
array[i] = cons();
return array;
}
}
public static class App
{
public static void Main()
{
Simple_Write_Test();
Rand_Test();
DateStruct_Test();
DateClass_Test();
Lookup_Test();
Min_Test();
Min_Wrapper_Test();
Console.ReadKey();
}
const int size = 1000000;
public static void Simple_Write_Test()
{
RunTest("Simple Write", size, () =>
{
return (new int[size], new int[size], new int[size]);
},
(data, i) =>
{
data.Item1[i] = i;
data.Item2[i] = i;
data.Item3[i] = i;
},
(data, i) =>
{
data.Item1[i] = i;
data.Item2[i] = i;
data.Item3[i] = i;
},
(data, i, id) =>
{
data.Item1[i] = i;
data.Item2[i] = i;
data.Item3[i] = i;
});
}
public static void Rand_Test()
{
var rand_lock = new object();
RunTest("Write Random", size, () =>
{
return (new int[size],
new Random[Environment.ProcessorCount].Fill(() => new Random()));
},
(data, i) =>
{
var rand = data.Item2[0];
data.Item1[i] = rand.Next();
},
(data, i) =>
{
int value;
lock (rand_lock)
{
value = data.Item2[0].Next();
}
data.Item1[i] = value;
},
(data, i, id) =>
{
var rand = data.Item2[id];
data.Item1[i] = rand.Next();
});
}
struct DateStruct
{
public int year;
public int month;
public int day;
public DateStruct(int year, int month, int day)
{
this.year = year;
this.month = month;
this.day = day;
}
}
public static void DateStruct_Test()
{
var rand_lock = new object();
RunTest("Date struct", size, () =>
{
return (new DateStruct[size], new Random[Environment.ProcessorCount].Fill(() => new Random()));
},
(data, i) =>
{
var rand = data.Item2[0];
data.Item1[i] = new DateStruct(rand.Next(), rand.Next(), rand.Next());
},
(data, i) =>
{
DateStruct value;
lock (rand_lock)
{
var rand = data.Item2[0];
value = new DateStruct(rand.Next(), rand.Next(), rand.Next());
}
data.Item1[i] = value;
},
(data, i, id) =>
{
var rand = data.Item2[id];
data.Item1[i] = new DateStruct(rand.Next(), rand.Next(), rand.Next());
});
}
class DateClass
{
public int year;
public int month;
public int day;
public DateClass(int year, int month, int day)
{
this.year = year;
this.month = month;
this.day = day;
}
}
public static void DateClass_Test()
{
var rand_lock = new object();
RunTest("Date Class", size, () =>
{
return (new DateClass[size], new Random[Environment.ProcessorCount].Fill(() => new Random()));
},
(data, i) =>
{
var rand = data.Item2[0];
data.Item1[i] = new DateClass(rand.Next(), rand.Next(), rand.Next());
},
(data, i) =>
{
DateClass value;
lock (rand_lock)
{
var rand = data.Item2[0];
value = new DateClass(rand.Next(), rand.Next(), rand.Next());
}
data.Item1[i] = value;
},
(data, i, id) =>
{
var rand = data.Item2[id];
data.Item1[i] = new DateClass(rand.Next(), rand.Next(), rand.Next());
});
}
public static void Lookup_Test()
{
RunTest("Lookup", size, () =>
{
var rand = new Random();
return (new int[size].Fill(() => rand.Next() % 100), new int[100].Fill(() => rand.Next() % 1000), new int[1000].Fill(() => rand.Next()));
},
(data, i) =>
{
data.Item1[i] = data.Item3[data.Item2[data.Item1[i]]];
},
(data, i) =>
{
data.Item1[i] = data.Item3[data.Item2[data.Item1[i]]];
},
(data, i, id) =>
{
data.Item1[i] = data.Item3[data.Item2[data.Item1[i]]];
});
}
public static void Min_Test()
{
object min_lock = new object();
RunTest("Min", size, () =>
{
var rand = new Random();
var array = new int[size].Fill(() => rand.Next());
var minValues = new int[Environment.ProcessorCount].Fill(() => int.MaxValue);
var minIndices = new int[Environment.ProcessorCount].Fill(() => -1);
return (array, minValues, minIndices);
},
(data, i) =>
{
var array = data.Item1;
if (array[i] < data.Item2[0])
{
data.Item2[0] = array[i];
data.Item3[0] = i;
}
},
(data, i) =>
{
var array = data.Item1;
if (array[i] < data.Item2[0])
{
lock (min_lock)
{
data.Item2[0] = array[i];
data.Item3[0] = i;
}
}
},
(data, i, id) =>
{
var array = data.Item1;
if (array[i] < data.Item2[id])
{
data.Item2[id] = array[i];
data.Item3[id] = i;
}
});
}
class Wrapper<T>
{
public T value;
public Wrapper(T value)
{
this.value = value;
}
public static implicit operator Wrapper<T>(T value) => new Wrapper<T>(value);
}
public static void Min_Wrapper_Test()
{
object min_lock = new object();
//wrap the values, so they don't share the cache line
RunTest("Min with wrapper", size, () =>
{
var rand = new Random();
var array = new int[size].Fill(() => rand.Next());
var minValues = new Wrapper<int>[Environment.ProcessorCount].Fill(() => int.MaxValue);
var minIndices = new Wrapper<int>[Environment.ProcessorCount].Fill(() => -1);
return (array, minValues, minIndices);
},
(data, i) =>
{
var array = data.Item1;
if (array[i] < data.Item2[0].value)
{
data.Item2[0] = array[i];
data.Item3[0] = i;
}
},
(data, i) =>
{
var array = data.Item1;
if (array[i] < data.Item2[0].value)
{
lock (min_lock)
{
data.Item2[0] = array[i];
data.Item3[0] = i;
}
}
},
(data, i, id) =>
{
var array = data.Item1;
if (array[i] < data.Item2[id].value)
{
data.Item2[id] = array[i];
data.Item3[id] = i;
}
});
}
public static void RunTest<T>(string name, int size,
Func<T> DataIntializer,
Action<T, int> excute,
Action<T, int> parallelForExecute,
Action<T, int, int> parallelExecute)
{
Log(name);
T data = DataIntializer();
var watch = Stopwatch.StartNew();
for (int i = 0; i < size; i )
excute(data, i);
Log($"single thread time {watch.ElapsedTicks / 10000f} ms");
data = DataIntializer();
watch.Restart();
Parallel.For(0, size, (i) =>
{
excute(data, i);
});
Log($"multi thread (parallel for) time {watch.ElapsedTicks / 10000f} ms");
data = DataIntializer();
watch.Restart();
For(size, (i, id) =>
{
parallelExecute(data, i, id);
});
Log($"multi thread (custom) time {watch.ElapsedTicks / 10000f} ms\n");
}
public static void For(int size, Action<int, int> excute)
{
Task[] tasks = new Task[Environment.ProcessorCount];
int seg = size / Environment.ProcessorCount;
int r = size - seg * Environment.ProcessorCount;
int last = 0;
for (int p = 0; p < tasks.Length; p )
{
int start = last;
int end = last seg (r-- > 0 ? 1 : 0);
int id = p;
last = end;
tasks[p] = Task.Run(() =>
{
for (int i = start; i < end; i )
excute(i, id);
});
}
Task.WaitAll(tasks);
}
public static void Log(object text)
{
Console.WriteLine(text);
}
}
CodePudding user response：
why is the Parallel.For run so poorly?
because the work in each iteration is far to small, so the overhead becomes dominating. What you should be doing is splitting the work into decently sized chunks, sort of what you are doing in your custom "parallel for".
The body of your parallel loop should probably take on the order of microseconds. But there is a balance here, if you make your chunks to small you will suffer due to overheads starting to dominate, if you make your chunks to large you might not be able to use all cores efficiently. See also how to speed up small loop bodies
If you just want to limit the number of cores used you should be using the Parallel.For overload that takes a ParallelOption: new ParallelOptions(){MaxDegreeOfParallelism = Environment.ProcessorCount}.
In tests one and two, I only achieved around 2X the performance, is there a better way to achieve more?
My guess is that you are bottle necking on the delegate invocation. Writing to memory takes on the order of cycles, doing a method call is kind of expensive on these timescales. In some cases methods can be inlined to avoid this penalty, but I do not think this can be done with delegates in the current .Net runtime.
this is just an observation: from tests three and four we can see that object allocation on the heap doesn't work well with multi-threading.
That is not the observation I would make. My conclusion is that you should avoid using contested locks, and if you need something like a Random you should be using the parallel.For overload that creates a thread local objects, so each thread gets its own object.
Heap allocations on multiple threads should perform fairly well, I believe each thread has a local segment it can allocate from, but I'm really not an expert on the inner details of the memory allocator. But I can say with confidence that you should not do high frequency allocations if you are writing high performance code, regardless of how many threads you are using.
shouldn't I be reaching 16x performance, am I doing anything wrong, is there anything I should be aware of?
You only have 8 real cores, the extra "threads" can help in some very specific circumstances, but typically something like 20% in the best case. Reaching even 8 times scaling should be considered good.
Keep in mind that cores are not the only resource that matters, there is also memory bandwidth, caches, cache coherency and many other things that might limit performance.
Note that you should probably be using Benchmark.Net to avoid common pitfalls when writing benchmarks.
CodePudding user response：
Short Answer
The short answer is that the overhead of creating, managing and switching between the tasks is greater than any benefit in running your operation on multiple threads.
Simply splitting up any operation into parallelly running tasks will not make it faster. The operation needs to be one that would benefit being broken up. Generally this will be an operation that is sufficiently long or that computationally intensive.
Secondly the way that an operation is split up is important. Splitting up an operation into 100 tasks may result in an overhead to creating and processing tasks that is greater than the benefit of using them but splitting the operation into 10 tasks might provide a performance benefit.
Using an analogy to explain the situation better: getting 1000 people (together at the same time) to complete a 100 piece puzzle is likely to be worse than having just 1 person complete it. But having 4 or 10 people complete the puzzle is likely to be better than 1.
Further Information
Note that the Parallel.For implementation and the custom parallel implementation are not equivalent.
The custom parallel implementation runs a range of iterations (within the for loop) per task but the Parallel.For implementation runs a single iteration per task (in practice Parallel.For may not use one Task for every iteration but that is the extreme scenario).
This means that the Parallel.For implementation will use many more tasks which will cause a lot more overhead for the creation and processing of the tasks.
As @Wyck mentioned in a comment, range partitioning can be used with Parallel.For implementation to mitigate this issue.
Whenever using tasks for performance you need to weigh up whether the overhead of using multiple tasks will be less than any reduction in duration achieved by splitting up an operation.
Sometimes you might need to adjust the number of times you split up an operation like in this example - there is a benefit, if you split it up into 8/16 Tasks but not (up to) 1000000 tasks.