Deep Dive on .NET Core String.Create() Performance

When was the last time that an insignificant detail sparked a movement in your brain? As a software engineer, I have a habit of focusing on the one tiny detail I have never seen before in a new block of code. At that point, the gears in my brain begin turning. I love these moments. The ones where the tiniest piece of trivia sends me down a rabbit-hole of discovery.

What did that look like the last time it came to you?

It recently happened to me as I was browsing Twitter. I came upon this exchange between David Fowler and Damian Edwards, discussing the .NET Span<T> API. I have used the Span<T> API before, but I found something new and different in the tweets.

The result pic.twitter.com/74tz33vfz7

— hackyderm.io/@damianedwards 🦣 (@DamianEdwards) October 20, 2020

It’s a small thing, The item that began my dive down the rabbit-hole. The String.Create method used above was the piece I had never seen before. String.Create was a hidden gem, and I was determined to uncover its mysteries. I found myself asking the question:

“Why use this method for creating strings over others?”

I began the journey, and it took me to a few interesting places that I would like to share with you. In this article, we will delve into a few topics…

• What is String.Create and how is it different from other APIs?
• What does String.Create do better and how can it make my C# code faster?
• How much more performance can it achieve?
• What pitfalls can and should be avoided?

Definitions

To make the article a little easier, I am going to refer to a few .NET Core APIs in the following ways:

• Create - Refers to using String.Create().
• Concat - Refers to using String.Concat() or the plus (+) operator.
• Format - Refers to using String.Format(), one of its many overloads, or string interpolation with the $"" syntax.
• StringBuilder - Refers to constructing a string with the fluent StringBuilder class and API.

How Does It Work?

The .NET Core codebase is open source and developed on GitHub. This provides a great opportunity to dive in and analyze Microsoft’s own practices. They provided the Create API, so seeing how they use it should provide valuable insight. Let’s start with a deep dive into the String object and its relevant APIs.

To generically build a string out of raw character data, you need to use the constructor that requires a pointer to a char array. Using this API directly would require placing individual characters into specific array locations. Below is the code that runs when you allocate a string using this constructor. There are many other ways to create strings, but this is what I consider the most comparable to the Create method.

57
58
59
60
61
62
63
64
65
66
67
68
69
70

string Ctor(char[]? value)
{
    if (value == null || value.Length == 0)
        return Empty;

    string result = FastAllocateString(value.Length);

    Buffer.Memmove(
        elementCount: (uint)result.Length, // derefing Length now allows JIT to prove 'result' not null below
        destination: ref result._firstChar,
        source: ref MemoryMarshal.GetArrayDataReference(value));

    return result;
}

To summarize the important steps:

1. Inputs are validated. An empty or null array returns String.Empty.
2. We allocate memory using FastAllocateString based on the array Length. FastAllocateString is implemented within the .NET Runtime itself and underpins nearly all string allocations.
3. We call Buffer.Memmove, which copies all bytes from the original array into the newly allocated string.
4. Return the resulting string.

To use this constructor, we need to supply it with a char array. After its job is complete, we end up with both a (now unnecessary) char array and a string, each with identical data. If we were to modify the original array, then the string would remain unmodified, because it is a separate and distinct copy of the data. In a high-performance .NET environment, saving object and array allocations can be extremely valuable because it reduces the total amount of work that the .NET Garbage Collector needs to do whenever it runs. Every extra object that is left in memory increases the frequency of collections and hurts total performance.

If you would like to learn more about .NET Garbage Collection and how it can impact performance, there are many fantastic resources online:

• ASP.NET Core Kestrel: Adventures in building a fast web server - Damian Edwards & David Fowler
• 8 Techniques to Avoid GC Pressure and Improve Performance in C# .NET
• How The Garbage Collector Can Cause Random Slowness

To contrast with the constructor and to eliminate this unnecessary array allocation, we take a look at the code that runs for the Create method:

363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379

public static string Create<TState>(int length, TState state, SpanAction<char, TState> action)
{
    if (action == null)
        throw new ArgumentNullException(nameof(action));

    if (length <= 0)
    {
        if (length == 0)
            return Empty;
        throw new ArgumentOutOfRangeException(nameof(length));
    }

    string result = FastAllocateString(length);
    action(new Span<char>(ref result.GetRawStringData(), length), state);

    return result;
}

The steps are similar but contain a critical difference:

1. Inputs are validated. Invalid action or negative Length will throw an exception and Length of 0 returns Empty;
2. We allocate memory using FastAllocateString based on the length parameter.
3. Convert the newly-allocated string to a Span<char>.
4. Invoke the provided action and pass in the new Span<char> along with the provided state.
5. Return the resulting string.

This methodology avoids the extra allocation by allowing us to pass in a SpanAction, which functions as a set of instructions for how our string should be created instead of requiring a second copy of all the bytes we need put into the string.

An analogy could be drawn from graphic design. You might want to get a professional logo created for your business, and you probably decided that you wanted it created in a vector format like SVG or EPS. You create a document that has details on the theme, colors, and general look of your logo. You now have a couple of options:

1. Draw the first draft yourself as close to the professional version as possible, send it to a graphic designer, and have them create a finalized version.
2. Send your original detailed document to the graphic designer, who creates a finalized logo from the information you provided.

In the first situation, you end up with two images, but only one that you plan to use. Your original can probably be thrown away without any ceremony, as you now have your professionally designed version. In the second case, you have not produced any extra product and therefore have produced less waste.

The process for the Create method is closer to the situation where you sent your documentation to a designer. By telling them exactly what you wanted from the beginning, you only generate one final artifact in the form that you need.

How is String.Create() Better?

At this point, you might be curious about the Create method, but you don’t necessarily know why it’s better than the methods you have used before. The Create APIs usefulness is situational, but it can be extremely powerful in the proper context.

• It pre-allocates a bucket, and then gives you an interface to fill that bucket safely. Other methods that produce similar results could require writing unsafe code or managing buffer pools.
• It avoids extra copy-operations of your data which often results in fewer allocations. This reduces pressure from your Garbage Collector which can speed up your overall program.
• It allows you to focus your high-performance code on the business requirements of your application rather than interleaving your string-building code with complex memory management.

Use Cases for String.Create()

You can only use the Create method when you already know your final string’s length. However, you can work creatively with this constraint and discover several ways to leverage Create. I performed a search through the codebases for dotnet/aspnetcore and dotnet/runtime to see where the Microsoft team decided to utilize this API. The rest of this article will dive deep into three trends that I found:

• Generating IDs
• Performance-Sensitive Concatenations
• Formatting Complex Strings

Generating IDs

Consider this class from the ASP.NET Core repository used for generating correlation IDs on each web request. It generates a correlation ID in the format of 13 characters chosen from the numbers (0-9) and most upper-case letters (A-V).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

// Copyright (c) .NET Foundation. All rights reserved.
// Licensed under the Apache License, Version 2.0. See License.txt in the project root for license information.

using System;
using System.Threading;

namespace Microsoft.AspNetCore.Connections
{
    internal static class CorrelationIdGenerator
    {
        // Base32 encoding - in ascii sort order for easy text based sorting
        private static readonly char[] s_encode32Chars = "0123456789ABCDEFGHIJKLMNOPQRSTUV".ToCharArray();

        // Seed the _lastConnectionId for this application instance with
        // the number of 100-nanosecond intervals that have elapsed since 12:00:00 midnight, January 1, 0001
        // for a roughly increasing _lastId over restarts
        private static long _lastId = DateTime.UtcNow.Ticks;

        public static string GetNextId() => GenerateId(Interlocked.Increment(ref _lastId));

        private static string GenerateId(long id)
        {
            return string.Create(13, id, (buffer, value) =>
            {
                char[] encode32Chars = s_encode32Chars;

                buffer[12] = encode32Chars[value & 31];
                buffer[11] = encode32Chars[(value >> 5) & 31];
                buffer[10] = encode32Chars[(value >> 10) & 31];
                buffer[9] = encode32Chars[(value >> 15) & 31];
                buffer[8] = encode32Chars[(value >> 20) & 31];
                buffer[7] = encode32Chars[(value >> 25) & 31];
                buffer[6] = encode32Chars[(value >> 30) & 31];
                buffer[5] = encode32Chars[(value >> 35) & 31];
                buffer[4] = encode32Chars[(value >> 40) & 31];
                buffer[3] = encode32Chars[(value >> 45) & 31];
                buffer[2] = encode32Chars[(value >> 50) & 31];
                buffer[1] = encode32Chars[(value >> 55) & 31];
                buffer[0] = encode32Chars[(value >> 60) & 31];
            });
        }
    }
}

The algorithm is brief:

1. Start your correlation ID at the latest tick count for UTC. The tick count is a 64-bit integer.
2. Increment by one on each request for a new ID.
3. For each of 13 characters:
   • Shift the value by 5 additional (character_index * 5) bits.
   • Grab the rightmost 5 bits (shifted_value & 31) and choose a character based upon a predetermined table.

For our baseline comparison, I used a naive implementation that utilizes StringBuilder. I chose this option because StringBuilder is often recommended as a good API for performance over regular string concatenation. I wrote additional implementations that attempted to use StringBuilder (with capacity), StringBuilder (without capacity), and simple concatenation.

Execution Benchmarks

Memory Benchmarks

The String.Create() method shows the best performance in performance (16.58 nanoseconds) and allocations (only 48 bytes!). Interestingly, the StringBuilder with no capacity specified also shows a small edge over the regular StringBuilder (it still loses to String.Create(), but that is interesting to note for future StringBuilder use).

Performance-Sensitive Concatenation

The C# Roslyn compiler is very intelligent about optimizing poor concatenations. The compiler will tend to convert multiple uses of the plus + operator into singular calls to Concat and likely has many additional tricks of which I am not aware. For these reasons, concatenation is generally a fast operation, but it still can be edged out by Create for simple scenarios.

The sample code to demonstrate concatenation with the Create method is straightforward.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

public static class ConcatenationStringCreate
{
    public static string Concat(string first, string second)
    {
        first ??= string.Empty;
        second ??= String.Empty;
        bool addSpace = second.Length > 0;

        int length = first.Length + (addSpace ? 1 : 0) + second.Length;
        return string.Create(length, (first, second, addSpace),
        (dst, v) =>
        {
            ReadOnlySpan<char> prefix = v.first;
            prefix.CopyTo(dst);

            if (v.addSpace)
            {
                dst[prefix.Length] = ' ';

                ReadOnlySpan<char> detail = v.second;
                detail.CopyTo(dst.Slice(prefix.Length + 1, detail.Length));
            }
        });
    }
}

I crafted this particular case after finding only one real example in the .NET Core source code. This feels like a case that could be reasonably abstracted and sprinkled throughout a codebase that is making excessive use of the plus + operator or String.Concat very frequently. The likely reason that I did not find more cases is that the benchmarks bear this out as being a marginal gain.

Note: There was nothing to report for memory benchmarks as both methods generated the same amount of allocated memory.

I find these results interesting. Even for generalized Concat usage, the Create method is marginally faster by a few percentage points. However, since it would be possible to generalize a library of fast string concatenation methods, the tradeoff of maintenance vs performance would become much cleaner. You could wrap all of your concatenation methods in a static class, add your unit tests, and never touch them again. From a maintenance perspective, Concat is still cleaner due to being very recognizable to other developers on your team. However, if you are concatenating strings at the pace of millions per second (such as a high traffic ASP.NET application), these few percentage points may be worth it. Overall, I would still advise profiling your specific case to prove that your specialized method is faster.

Formatting of Complex Strings

Complex formats can also be built using Create when you know the length of all of the smaller segments involved. Usually, this will be when you have these strings encapsulated as properties of a single class or within the parameters of a single method. It may also be advantageous to use Create when dealing with rows of fixed-width or delimited data, such as a CSV file.

I found a great example in the ASP.NET Core repository in the form of the SetCookieHeaderValue class. This class contains all of the properties necessary to write out a cookie HTTP header. Conveniently, this class also included a method that used StringBuilder to accomplish the same formatting task. This gave me a great opportunity to write a fast benchmark.

Here is the table that shows the two methods up against one another:

For a complex string like a cookie header, Create improves performance by nearly 33%! Complex formats like this show that there is a lot to gain when using the Create method.

While the SetCookieHeaderValue class provides an ideal data point, its logic is rather long and complex. I wrote a simple class to demonstrate the same formatting principles.

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108

public class Dog
{
    public string Name { get; set; }

    public int? Age { get; set; }

    public string Color { get; set; }

    public override string ToString()
    {
        return StringCreate();
    }

    /// <summary>
    /// Format the description string of the Dog class using String.Create()
    /// </summary>
    public string StringCreate()
    {
        var length = 0;
        // Constants
        const string dogPrefix = "[DOG] ";
        const string unknownName = "Unknown";
        const string leftAgeChunk = " (";
        const char rightAgeChunk = ')';
        const string leftColorChunk = " [";
        const char rightColorChunk = ']';
        static int integerLength(int val) => (int)Math.Floor(Math.Log10((double)val) + 1);

        /* Compute Lengths */
        length += dogPrefix.Length + (Name ?? unknownName).Length; // Prefix + Name

        if (Color is string)
        {
            length += 3 /* left + right chunk length */ + Color.Length;
        }

        if (Age.HasValue)
        {
            length += 3 /* left + right chunk length */ + integerLength(Age.Value); /* Digits in age */
        }
        /* Use State + Computed Length to Build String */
        return String.Create<Dog>(length, this, (buffer, dog) =>
        {
            var prefixSpan = dogPrefix.AsSpan();
            prefixSpan.CopyTo(buffer);
            var span = buffer.Slice(prefixSpan.Length);

            var nameSpan = (dog.Name ?? unknownName).AsSpan();
            nameSpan.CopyTo(span);
            span = span.Slice(nameSpan.Length);

            if(dog.Color is string)
            {
                leftColorChunk.AsSpan().CopyTo(span);
                span = span.Slice(2);
                var colorSpan = dog.Color.AsSpan();
                colorSpan.CopyTo(span);
                span = span.Slice(colorSpan.Length);
                span[0] = rightColorChunk;
                span = span.Slice(1);
            }

            if(dog.Age.HasValue)
            {
                leftAgeChunk.AsSpan().CopyTo(span);
                span = span.Slice(2);
                dog.Age.Value.TryFormat(span, out int written, provider: CultureInfo.InvariantCulture);
                span = span.Slice(written);
                span[0] = rightAgeChunk;
            }
        });
    }

    /// <summary>
    /// Format the description string of the Dog class using basic concatenation.
    /// </summary>
    public string Concatenation() 
    {
        var val = "[DOG] " + (Name ?? "Unknown");

        if(Color is string)
        {
            val += " [" + Color + "]";
        }

        if(Age.HasValue)
        {
            val += " (" + Age.Value + ")";
        }

        return val;
    }

    /// <summary>
    /// Format the description string of the Dog class using String.Format()
    /// </summary>
    public string StringFormat()
    {
        return String.Format("[DOG] {0}{5}{4}{6}{2}{1}{3}",
            Name ?? "Unknown",
            Age,
            Age.HasValue ? " (" : String.Empty,
            Age.HasValue ? ")" : String.Empty,
            Color,
            Color is string ? " [" : String.Empty,
            Color is string ? "]" : String.Empty);
    }
}

The Dog class above contains three methods that should produce the identical output: StringCreate, Concatenation, and StringFormat. Each method uses a formatting strategy that matches its name. Despite being much more complex, the actual strategy of StringCreate is relatively simple:

1. Preemptively calculate the length of each property of the Dog class. Some properties may require some creative logic to calculate length without materializing the full string (Int32 is a good example).
2. Invoke String.Create using the current class (this) as the state parameter along with the pre-calculated length.
3. Use the Span<T> APIs to fill the newly instantiated buffer variable and return the final string.

Overall, the StringFormat and Concatenation methods are far shorter and less complex. This exemplifies the difficulties you might encounter when properly formatting complex strings using the Create method and why it should only be used on performance-critical paths in your code. However, utilizing this method appropriately can pay significant dividends in overall performance.

Execution Time Benchmarks

Memory Benchmarks

For the more complex cases, formatting using the Create method is between 25% and 35% faster. The benchmark methods I chose show another reason to carefully consider this use case for the Create method. You can see that simple concatenation performs slightly better when there are fewer elements to format. In the case where the Dog has no Age or Color value, the work needed to calculate and format the string eclipses the performance gain from using Create. However, once we populate these properties, Concat quickly becomes the poorer option. This shows why it is important to understand your data before using this method as well. If 99% of your cases are simple, you simply will not need this much help. Format is consistently the slowest option, though is arguably the fastest method to write initially and likely the most maintainable in the future.

When NOT To Use String.Create()

Create shows great promise in performance-critical code, but there are many legitimate reasons to avoid it. As software engineers, we often become more tied to the metrics of our systems at the expense of the bigger picture. Generally, I think decent but maintainable code should prevail over fantastic performance. That leads me to prescribe three general cases when you should avoid using Create, even if it sacrifices performance.

1) Do Not Use When Readability is Important

Ultimately, this API is not maintenance-friendly. Your scenario should demand very high performance and your code should be well-factored and contain unit tests. When writing code that requires occasional maintenance, you can easily stick with one of these simpler formatting methods:

• Format or String Interpolation when generating simple strings with dynamic values.
• StringBuilder when creating strings that require a loop or many conditional elements.
• Concat or a simple + when you just need to combine a small number of strings.

2) Do Not Use When Culture is Important

Format, String Interpolation, and most ToString() methods respect cultural formatting options. This gives your code the crucial ability to adapt to culture-specific date, numeral, and currency formats without having to code that logic yourself. Create does not offer any support for these APIs on its own, and attempting to mimic the behavior in your code would likely require allocating additional strings, thus removing the advantages of using the Create method.

3) (Probably) Do Not Use When the Output is for Humans

This situation is a bit subjective. The reason I would not recommend using Create for formatting for humans is that humans often want things to change. Since formatting using the Create method is extremely verbose, any changes are likely to cause increased complexity over time and thereby generate an accumulating amount of technical debt. In my opinion, the best usage of Create is on machine-readable strings or more generalized string-writing APIs that are unlikely to change in the future. As is often the case in software development, the specifics of your situation are the most important, but I think this is a good general rule to avoid this for any code you anticipate modifying frequently in the future.

Is String.Create() Right for You?

The answer to this question is definitely it depends. I began this investigation by focusing on something new to me: the String.Create method. I dove down a bit of a rabbit-hole looking for new and interesting ways to use this API. During this journey, I found three primary ways of using the method: (1) generating IDs, (2) fast concatenation, and (3) complex string formatting. Our benchmarks showed that we could, with reasonable consistency, produce a faster method using String.Create at the expense of simpler, more readable code. When answering the question above, you should always consider:

• How important is this code’s readability?
• How likely is it to change?
• How much will my program use this code and how much performance will it gain?

These all depend on your code and your team’s strengths, but my hope today is that I have given you some tools to help you determine your path.

The code for this article’s benchmarks are available on GitHub.