Performance Improvements in .NET Core 3.0

Google+ Pinterest LinkedIn Tumblr +
Avatar

Back when we were getting ready to ship .NET Core 2.0, I wrote a blog post exploring some of the many performance improvements that had gone into it. I enjoyed putting it together so much and received such a positive response to the post that I did it again for .NET Core 2.1, a version for which performance was also a significant focus. With //build last week and .NET Core 3.0‘s release now on the horizon, I’m thrilled to have an opportunity to do it again.

.NET Core 3.0 has a ton to offer, from Windows Forms and WPF, to single-file executables, to async enumerables, to platform intrinsics, to HTTP/2, to fast JSON reading and writing, to assembly unloadability, to enhanced cryptography, and on and on and on… there is a wealth of new functionality to get excited about. For me, however, performance is the primary feature that makes me excited to go to work in the morning, and there’s a staggering amount of performance goodness in .NET Core 3.0.

In this post, we’ll take a tour through some of the many improvements, big and small, that have gone into the .NET Core runtime and core libraries in order to make your apps and services leaner and faster.

Setup

Benchmark.NET has become the preeminent tool for doing benchmarking of .NET libraries, and so as I did in my 2.1 post, I’ll use Benchmark.NET to demonstrate the improvements. Throughout the post, I’ll include the individual snippets of benchmarks that highlight the particular improvement being discussed. To be able to execute those benchmarks, you can use the following setup:

  1. Ensure you have .NET Core 3.0 installed, as well as .NET Core 2.1 for comparison purposes.
  2. Create a directory named BlogPostBenchmarks.
  3. In that directory, run dotnet new console.
  4. Replace the contents of BlogPostBenchmarks.csproj with the following:
  5. Replace the contents of Program.cs with the following:

To execute a particular benchmark, unless otherwise noted, copy and paste the relevant code to replace the // ...above, and execute dotnet run -c Release -f netcoreapp2.1 --runtimes netcoreapp2.1 netcoreapp3.0 --filter "*Program*". This will compile and run the tests in release builds, on both .NET Core 2.1 and .NET Core 3.0, and print out the results for comparison in a table.

Caveats

A few caveats before we get started:

  1. Any discussion involving microbenchmark results deserves a caveat that measurements can and do vary from machine to machine. I’ve tried to pick stable examples to share (and have run these tests on multiple machines in multiple configurations to help validate that), but don’t be too surprised if your numbers differ from the ones I’ve shown; hopefully, however, the magnitude of the improvements demonstrated carries through. All of the shown results are from a nightly Preview 6 build for .NET Core 3.0. Here’s my configuration, as summarized by Benchmark.NET, on my Windows configuration and on my Linux configuration:
  2. Unless otherwise mentioned, benchmarks were executed on Windows. In many cases, performance is equivalent between Windows and Unix, but in others, there can be non-trivial discrepancies between them, in particular in places where .NET relies on OS functionality, and the OS itself has different performance characteristics.
  3. I mentioned posts on .NET Core 2.0 and .NET Core 2.1, but I didn’t mention .NET Core 2.2. .NET Core 2.2 was primarily focused on ASP.NET, and while there were terrific performance improvements at the ASP.NET layer in 2.2, the release was primarily focused on servicing for the runtime and core libraries, with most improvements post-2.1 skipping 2.2 and going into 3.0.

With that out of the way, let’s have some fun.

Span and Friends

One of the more notable features introduced in .NET Core 2.1 was Span<T>, along with its friends ReadOnlySpan<T>Memory<T>, and ReadOnlyMemory<T>. The introduction of these new types came with hundreds of new methods for interacting with them, some on new types and some with overloaded functionality on existing types, as well as optimizations in the just-in-time compiler (JIT) for making working with them very efficient. The release also included some internal usage of Span<T> to make existing operations leaner and faster while still enjoying maintainable and safe code. In .NET Core 3.0, much additional work has gone into further improving all such aspects of these types: making the runtime better at generating code for them, increasing the use of them internally to help improve many other operations, and improving the various library utilities that interact with them to make consumption of these operations faster.

To work with a span, one first needs to get a span, and several PRs have made doing so faster. In particular, passing around a Memory<T> and then getting a Span<T> from it is a very common way of creating a span; this is, for example, how the various Stream.WriteAsync and ReadAsync methods work, accepting a {ReadOnly}Memory<T> (so that it can be stored on the heap) and then accessing its Span property once the actual bytes need to be read or written. PR dotnet/coreclr#20771 improved this by removing an argument validation branch (both for {ReadOnly}Memory<T>.Span and for {ReadOnly}Span<T>.Slice), and while removing a branch is a small thing, in span-heavy code (such as when doing formatting and parsing), small things done over and over and over again add up. More impactful, PR dotnet/coreclr#20386 plays tricks at the runtime level to safely eliminate some of the runtime checked casting and bit masking logic that had been used to enable {ReadOnly}Memory<T> to wrap various types, like stringT[], and MemoryManager<T>, providing a seamless veneer over all of them. The net result of these PRs is a nice speed-up when fishing a Span<T> out of a Memory<T>, which in turn improves all other operations that do so.

Method Toolchain Mean Error StdDev Ratio
GetSpan netcoreapp2.1 3.873 ns 0.0927 ns 0.0822 ns 1.00
GetSpan netcoreapp3.0 1.843 ns 0.0401 ns 0.0375 ns 0.48

 

Of course, once you get a span, you want to use it, and there a myriad of ways to use one, many of which have also been optimized further in .NET Core 3.0.

For example, just as with arrays, to pass the data from a span to native code via a P/Invoke, the data needs to be pinned (unless it’s already immovable, such as if the span were created to wrap some natively allocated memory not on the GC heap or if it were created for some data on the stack). To pin a span, the easiest way is to simply rely on the C# language’s support added in C# 7.3 that supports a pattern-based way to use any type with the fixed keyword. All a type need do is expose a GetPinnableReference method (or extension method) that returns a ref T to the data stored in that instance, and that type can be used with fixed{ReadOnly}Span<T> does exactly this. However, even though {ReadOnly}Span<T>.GetPinnableReference generally gets inlined, a call it makes internally to Unsafe.AsRef was getting blocked from inlining; PR dotnet/coreclr#18274 fixed this, enabling the whole operation to be inlined. Further, the aforementioned code was actually tweaked in PR dotnet/coreclr#20428 to eliminate one branch on the hot path. Both of these combine to result in a measurable boost when pinning a span:

Method Toolchain Mean Error StdDev Ratio RatioSD
PinSpan netcoreapp2.1 0.7930 ns 0.0177 ns 0.0189 ns 1.00 0.00
PinSpan netcoreapp3.0 0.6496 ns 0.0109 ns 0.0102 ns 0.82 0.03

 

It’s worth noting, as well, that if you’re interested in these kinds of micro-optimizations, you might also want to avoid using the default pinning at all, at least on super hot paths. The {ReadOnly}Span<T>.GetPinnableReference method was designed to behave just like pinning of arrays and strings, where null or empty inputs result in a null pointer. This behavior requires an additional check to be performed to see whether the length of the span is zero:

If in your code by construction you know that the span will not be empty, you can choose to instead use MemoryMarshal.GetReference, which performs the same operation but without the length check:

Again, while a single check adds minor overhead, when executed over and over and over, that can add up:

Method Mean Error StdDev Ratio RatioSD
PinSpan 0.6524 ns 0.0129 ns 0.0159 ns 1.00 0.00
PinSpanExplicit 0.5200 ns 0.0111 ns 0.0140 ns 0.80 0.03

 

Of course, there are many other (and generally preferred) ways to operate over a span’s data than to use fixed. For example, it’s a bit surprising that until Span<T> came along, .NET didn’t have a built-in equivalent of memcmp, but nevertheless, Span<T>‘s SequenceEqual and SequenceCompareTo methods have become go-to methods for comparing in-memory data in .NET. In .NET Core 2.1, both SequenceEqual and SequenceCompareTo were optimized to utilize System.Numerics.Vector for vectorization, but the nature of SequenceEqual made it more amenable to best take advantage. In PR dotnet/coreclr#22127, @benaadams updated SequenceCompareTo to take advantage of the new hardware instrinsics APIs available in .NET Core 3.0 to specifically target AVX2 and SSE2, resulting in significant improvements when comparing both small and large spans. (For more information on hardware intrinsics in .NET Core 3.0, see platform-intrinsics.md and using-net-hardware-intrinsics-api-to-accelerate-machine-learning-scenarios.)

Method Toolchain Length Mean Error StdDev Ratio
CompareSame netcoreapp2.1 16 16.955 ns 0.2009 ns 0.1781 ns 1.00
CompareSame netcoreapp3.0 16 4.757 ns 0.0938 ns 0.0732 ns 0.28
CompareDifferFirst netcoreapp2.1 16 11.874 ns 0.1240 ns 0.1100 ns 1.00
CompareDifferFirst netcoreapp3.0 16 5.174 ns 0.0543 ns 0.0508 ns 0.44
CompareDifferLast netcoreapp2.1 16 16.644 ns 0.2146 ns 0.2007 ns 1.00
CompareDifferLast netcoreapp3.0 16 5.373 ns 0.0479 ns 0.0448 ns 0.32
CompareSame netcoreapp2.1 256 43.740 ns 0.8226 ns 0.7292 ns 1.00
CompareSame netcoreapp3.0 256 11.055 ns 0.1625 ns 0.1441 ns 0.25
CompareDifferFirst netcoreapp2.1 256 12.144 ns 0.0849 ns 0.0752 ns 1.00
CompareDifferFirst netcoreapp3.0 256 6.663 ns 0.1044 ns 0.0977 ns 0.55
CompareDifferLast netcoreapp2.1 256 39.697 ns 0.9291 ns 2.6054 ns 1.00
CompareDifferLast netcoreapp3.0 256 11.242 ns 0.2218 ns 0.1732 ns 0.32

 

As background, “vectorization” is an approach to parallelization that performs multiple operations as part of individual instructions on a single core. Some optimizing compilers can perform automatic vectorization, whereby the compiler analyzes loops to determine whether it can generate functionally equivalent code that would utilize such instructions to run faster. The .NET JIT compiler does not currently perform auto-vectorization, but it is possible to manually vectorize loops, and the options for doing so have significantly improved in .NET Core 3.0. Just as a simple example of what vectorization can look like, imagine having an array of bytes and wanting to search it for the first non-zero byte, returning the position of that byte. The simple solution is to just iterate through all of the bytes:

That of course works functionally, and for very small arrays it’s fine. But for larger arrays, we end up doing significantly more work than is actually necessary. Consider instead in a 64-bit process re-interpreting the array of bytes as an array of longs, which Span<T> nicely supports. We then effectively compare 8 bytes at a time rather than 1 byte at a time, at the expense of added code complexity: once we find a non-zero long, we then need to look at each byte it contains to determine the position of the first non-zero one (though there are ways to improve that, too). Similarly, the array’s length may not evenly divide by 8, so we need to be able to handle the overflow.

For longer arrays, this yields really nice wins:

Method Mean Error StdDev Ratio
LoopBytes 5,462.3 ns 107.093 ns 105.180 ns 1.00
LoopLongs 568.6 ns 6.895 ns 5.758 ns 0.10

 

I’ve glossed over some details here, but it should convey the core idea. .NET includes additional mechanisms for vectorizing as well. In particular, the aforementioned System.Numerics.Vector type allows for a developer to write code using Vector and then have the JIT compiler translate that into the best instructions available on the current platform.

Method Mean Error StdDev Ratio
LoopBytes 5,462.3 ns 107.093 ns 105.180 ns 1.00
LoopLongs 568.6 ns 6.895 ns 5.758 ns 0.10
LoopVectors 306.0 ns 4.502 ns 4.211 ns 0.06

 

Further, .NET Core 3.0 includes new hardware intrinsics that allow a properly-motivated developer to eek out the best possible performance on supporting hardware, utilizing extensions like AVX or SSE that can compare well more than 8 bytes at a time. Many of the improvements in .NET Core 3.0 come from utilizing these techniques.

Back to examples, copying spans has also improved, thanks to PRs dotnet/coreclr#18006 from @benaadams and dotnet/coreclr#17889, in particular for relatively small spans…

Method Toolchain Mean Error StdDev Ratio
CopySpan netcoreapp2.1 10.913 ns 0.1960 ns 0.1737 ns 1.00
CopySpan netcoreapp3.0 3.568 ns 0.0528 ns 0.0494 ns 0.33

 

Searching is one of the most commonly performed operations in any program, and searches with spans are generally performed with IndexOf and its variants (e.g. IndexOfAny and Contains) In PR dotnet/coreclr#20738, @benaadams again utilized vectorization, this time to improve the performance of IndexOfAny when operating over bytes, a particularly common case in many networking-related scenarios (e.g. parsing bytes off the wire as part of an HTTP stack). You can see the effects of this in the following microbenchmark:

Method Toolchain Mean Error StdDev Ratio
IndexOf netcoreapp2.1 12.828 ns 0.1805 ns 0.1600 ns 1.00
IndexOf netcoreapp3.0 4.504 ns 0.0968 ns 0.0858 ns 0.35

 

I love these kinds of improvements, because they’re low-enough in the stack that they end up having multiplicative effects across so much code. The above change only affected byte, but subsequent PRs were submitted to cover char as well, and then PR dotnet/coreclr#20855 made a nice change that brought these same changes to other primitives of the same sizes. For example, we can recast the previous benchmark to use sbyte instead of byte, and as of that PR, a similar improvement applies:

Method Toolchain Mean Error StdDev Ratio
IndexOf netcoreapp2.1 24.636 ns 0.2292 ns 0.2144 ns 1.00
IndexOf netcoreapp3.0 9.795 ns 0.1419 ns 0.1258 ns 0.40

 

As another example, consider PR dotnet/coreclr#20275. That change similarly utilized vectorization to improve the performance of To{Upper/Lower}{Invariant}.

Method Toolchain Mean Error StdDev Ratio
ToUpperInvariant netcoreapp2.1 64.36 ns 0.8099 ns 0.6763 ns 1.00
ToUpperInvariant netcoreapp3.0 26.48 ns 0.2411 ns 0.2137 ns 0.41

 

PR dotnet/coreclr#19959 optimizes the Trim{Start/End} helpers on ReadOnlySpan<char>, another very commonly-applied method, with equally exciting results (it’s hard to see with the white space in the results, but the results in the table go in order of the arguments in the Params attribute):

Method Toolchain Data Mean Error StdDev Ratio
Trim netcoreapp2.1 12.999 ns 0.1913 ns 0.1789 ns 1.00
Trim netcoreapp3.0 3.078 ns 0.0349 ns 0.0326 ns 0.24
Trim netcoreapp2.1 abcdefg 17.618 ns 0.3534 ns 0.2951 ns 1.00
Trim netcoreapp3.0 abcdefg 7.927 ns 0.0934 ns 0.0828 ns 0.45
Trim netcoreapp2.1 abcdefg 15.522 ns 0.2200 ns 0.1951 ns 1.00
Trim netcoreapp3.0 abcdefg 5.227 ns 0.0750 ns 0.0665 ns 0.34

 

Sometimes optimizations are just about being smarter about code management. PR dotnet/coreclr#17890 removed an unnecessary layer of functions that were on many globalization-related code paths, and just removing those extra unnecessary method invocations results in measurable speed-ups when working with small spans, e.g.

Method Toolchain Mean Error StdDev Ratio
EndsWith netcoreapp2.1 37.80 ns 0.3290 ns 0.2917 ns 1.00
EndsWith netcoreapp3.0 12.26 ns 0.1479 ns 0.1384 ns 0.32

 

Of course, one of the great things about span is that it is a reusable building-block that enables many higher-level operations. That includes operations on both arrays and strings…

Arrays and Strings

As a theme that’s emerged within .NET Core, wherever possible, new performance-focused functionality should not only be exposed for public use but also be used internally; after all, given the depth and breadth of functionality within .NET Core, if some performance-focused feature doesn’t meet the needs of .NET Core itself, there’s a reasonable chance it also won’t meet the public need. As such, internal usage of new features is a key benchmark as to whether the design is adequate, and in the process of evaluating such criteria, many additional code paths benefit, and these improvements have a multiplicative effect.

This isn’t just about new APIs. Many of the language features introduced in C# 7.2, 7.3, and 8.0 are influenced by the needs of .NET Core itself and have been used to improve things that we couldn’t reasonably improve before (other than dropping down to unsafe code, which we try to avoid when possible). For example, PR dotnet/coreclr#17891 speeds up Array.Reverse by taking advantage of the C# 7.2 ref locals feature and the 7.3 ref local reassignment feature. Using the new feature allows for the code to be expressed in a way that lets the JIT generate better code for the inner loop, and in turn results in a measurable speed-up:

Method Toolchain Mean Error StdDev Ratio RatioSD
Reverse netcoreapp2.1 105.06 ns 2.488 ns 7.337 ns 1.00 0.00
Reverse netcoreapp3.0 74.12 ns 1.494 ns 2.536 ns 0.66 0.02

 

Another example for arrays, the Clear method improved in PR dotnet/coreclr#24302, which works around an alignment issue that results in the underlying memset used to implement the operation being up to 2x slower. The change manually clears up to a few bytes one by one, such that the pointer we then hand off to memset is properly aligned. If you got “lucky” previously and the array happened to be aligned, performance was fine, but if it wasn’t aligned, there was a non-trivial performance hit incurred. This benchmark simulates the unlucky case:

Method Toolchain Mean Error StdDev Ratio
Clear netcoreapp2.1 121.59 ns 0.8349 ns 0.6519 ns 1.00
Clear netcoreapp3.0 87.91 ns 1.7768 ns 1.6620 ns 0.73

 

That said, many of the improvements are in fact based on new APIs. Span is a great example of this. It was introduced in .NET Core 2.1, and the initial push was to get it to be usable and expose sufficient surface area to allow it to be used meaningfully. But at the same time, we started utilizing it internally in order to both vet the design and benefit from the improvements it enables. Some of this was done in .NET Core 2.1, but the effort continues in .NET Core 3.0. Arrays and strings are both prime candidates for such optimizations.

For example, many of the same vectorization optimizations applied to spans are similarly applied to arrays. PR dotnet/coreclr#21116 from @benaadams optimized Array.{Last}IndexOf for both bytes and chars, utilizing the same internal helpers that were written to enable spans, and to similar effect:

Method Toolchain Mean Error StdDev Ratio RatioSD
IndexOf netcoreapp2.1 34.976 ns 0.6352 ns 0.5631 ns 1.00 0.00
IndexOf netcoreapp3.0 9.471 ns 0.6638 ns 1.1091 ns 0.29 0.04

 

And as with spans, thanks to PR dotnet/coreclr#24293 from @dschinde, these IndexOfoptimizations also now apply to other primitives of the same size.

Method Toolchain Mean Error StdDev Ratio
IndexOf netcoreapp2.1 34.181 ns 0.6626 ns 0.6508 ns 1.00
IndexOf netcoreapp3.0 9.600 ns 0.1913 ns 0.1598 ns 0.28

 

Vectorization optimizations have been applied to strings, too. You can see the effect of PR dotnet/coreclr#21076 from @benaadams in this microbenchmark:

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
IndexOf netcoreapp2.1 75.14 ns 1.5285 ns 1.6355 ns 1.00 0.0151 32 B
IndexOf netcoreapp3.0 11.70 ns 0.2382 ns 0.2111 ns 0.16

 

Also note in the above that the .NET Core 2.1 operation allocates (due to converting the search character into a string), whereas the .NET Core 3.0 implementation does not. That’s thanks to PR dotnet/coreclr#19788 from @benaadams.

There are of course pieces of functionality that are more unique to strings (albeit also applicable to new functionality exposed on spans), such as hash code computation with various string comparison methods. For example, PR dotnet/coreclr#20309/ improved the performance of String.GetHashCode when performing OrdinalIgnoreCase operations, which along with Ordinal (the default) represent the two most common modes.

Method Toolchain Mean Error StdDev Ratio
GetHashCodeIgnoreCase netcoreapp2.1 47.70 ns 0.5751 ns 0.5380 ns 1.00
GetHashCodeIgnoreCase netcoreapp3.0 14.28 ns 0.1462 ns 0.1296 ns 0.30

 

OrdinalsIgnoreCase has been improved for other uses as well. For example, PR dotnet/coreclr#20734 improved String.Equals when using StringComparer.OrdinalIgnoreCaseby both vectorizing (checking two chars at a time instead of one) and removing branches from an inner loop:

Method Toolchain Mean Error StdDev Ratio
EqualsIC netcoreapp2.1 24.036 ns 0.3819 ns 0.3572 ns 1.00
EqualsIC netcoreapp3.0 9.165 ns 0.0589 ns 0.0551 ns 0.38

 

The previous cases are examples of functionality in String‘s implementation, but there are lots of ancillary string-related functionality that have seen improvements as well. For example, various operations on Char have been improved, such as Char.GetUnicodeCategory via PRs dotnet/coreclr#20983 and dotnet/coreclr#20864:

Method Toolchain Char Mean Error StdDev Ratio RatioSD
GetCategory netcoreapp2.1 . 1.8001 ns 0.0160 ns 0.0142 ns 1.00 0.00
GetCategory netcoreapp3.0 . 0.4925 ns 0.0141 ns 0.0132 ns 0.27 0.01
GetCategory netcoreapp2.1 a 1.7925 ns 0.0144 ns 0.0127 ns 1.00 0.00
GetCategory netcoreapp3.0 a 0.4957 ns 0.0117 ns 0.0091 ns 0.28 0.01
GetCategory netcoreapp2.1 ? 3.7836 ns 0.0493 ns 0.0461 ns 1.00 0.00
GetCategory netcoreapp3.0 ? 2.7531 ns 0.0757 ns 0.0633 ns 0.73 0.02

 

Those PRs also highlight another case of benefiting from a language improvement. As of C# 7.3, the C# compiler is able to optimize properties of the form:

Rather than emitting this exactly as written, which would allocate a new byte array on each call, the compiler takes advantage of the facts that a) the bytes backing the array are all constant and b) it’s being returned as a read-only span, which means the consumer is unable to mutate the data using safe code. As such, with PR dotnet/roslyn#24621, the C# compiler instead emits this by writing the bytes as a binary blob in metadata, and the property then simply creates a span that points directly to that data, making it very fast to access the data, more so even than if this property returned a static byte[].

Method Mean Error StdDev Median Ratio
GetArrayProp 1.3362 ns 0.0498 ns 0.0416 ns 1.3366 ns 1.000
GetSpanProp 0.0125 ns 0.0132 ns 0.0110 ns 0.0080 ns 0.009

 

Another string-related area that’s gotten some attention is StringBuilder (not necessarily improvements to StringBuilder itself, although it has received some of those, for example a new overload in PR dotnet/coreclr#20773 from @Wraith2 that helps avoid accidentally boxing and creating a string from a ReadOnlyMemory<char> appended to the builder). Rather, in many situations StringBuilders have been used for convenience but added cost, and with just a little work (and in some cases the new String.Create method introduced in .NET Core 2.1), we can eliminate that overhead, in both CPU usage and allocation. Here a few examples…

Method Toolchain Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
GetHostEntry netcoreapp2.1 532.7 us 16.59 us 46.79 us 526.8 us 1.00 0.00 1.9531 4888 B
GetHostEntry netcoreapp3.0 527.7 us 12.85 us 37.06 us 542.8 us 1.00 0.11 616 B

 

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
FormatHebrew netcoreapp2.1 626.0 ns 7.917 ns 7.405 ns 1.00 0.00 0.2890 608 B
FormatHebrew netcoreapp3.0 570.6 ns 10.504 ns 9.825 ns 0.91 0.02 0.1554 328 B

 

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
PAShort netcoreapp2.1 33.68 ns 1.0378 ns 2.9271 ns 1.00 0.00 0.0648 136 B
PAShort netcoreapp3.0 17.12 ns 0.4240 ns 0.7313 ns 0.55 0.04 0.0153 32 B
PALong netcoreapp2.1 2,761.80 ns 50.1515 ns 46.9117 ns 1.00 0.00 1.1940 2512 B
PALong netcoreapp3.0 787.78 ns 27.4673 ns 80.1234 ns 0.31 0.01 0.5007 1048 B

 

Method Toolchain Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
CertProp netcoreapp2.1 209.30 ns 4.464 ns 10.435 ns 204.35 ns 1.00 0.00 0.1256 264 B
CertProp netcoreapp3.0 95.82 ns 1.822 ns 1.704 ns 96.43 ns 0.45 0.02 0.0497 104 B

 

and so on. These PRs demonstrate that good gains can be had simply by making small tweaks that make existing code paths cheaper, and that expands well beyond StringBuilder. There are lots of places within .NET Core, for example, where String.Substring is used, and many of those cases can be replaced with use of AsSpan and Slice, for example as was done in PR dotnet/corefx#29402 by @juliushardt, or PRs dotnet/coreclr#17916 and dotnet/corefx#29539, or as was done in PRs dotnet/corefx#29227 and dotnet/corefx#29721 to remove string allocations from FileSystemWatcher, delaying the creation of such strings until only when it was known they were absolutely necessary.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
HtmlDecode netcoreapp2.1 638.2 ns 8.474 ns 7.077 ns 1.00 0.1516 320 B
HtmlDecode netcoreapp3.0 153.7 ns 2.776 ns 2.461 ns 0.24 0.0191 40 B

 

Another example of using new APIs to improve existing functionality is with String.Concat. .NET Core 3.0 has several new String.Concat overloads, ones that accept ReadOnlySpan<char> instead of string. These make it easy to avoid allocations/copies of substrings in cases where concatenating pieces of other strings: instead of using String.Concat with String.Substring, it’s used instead with String.AsSpan(...) or Slice. In fact, the PRs dotnet/coreclr#21766 and dotnet/corefx#34451 that implemented, exposed, and added tests for these new overloads also added tens of call sites to the new overloads across .NET Core. Here’s an example of the impact one of those has, improving the performance of accessing Uri.DnsSafeHost:

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
DnsSafeHost netcoreapp2.1 733.7 ns 14.448 ns 17.20 ns 1.00 0.00 0.2012 424 B
DnsSafeHost netcoreapp3.0 450.1 ns 9.013 ns 18.41 ns 0.63 0.02 0.1059 224 B

 

Another example, using Path.ChangeExtension to change from one non-null extension to another:

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
ChangeExtension netcoreapp2.1 30.57 ns 0.7124 ns 0.6664 ns 1.00 0.0495 104 B
ChangeExtension netcoreapp3.0 24.54 ns 0.3398 ns 0.2838 ns 0.80 0.0229 48 B

 

Finally, a very closely related area is that of encoding. A bunch of improvements were made in .NET Core 3.0 around Encoding, both in general and for specific encodings, such as PR dotnet/coreclr#18263 that allowed an existing corner-case optimization to be applied for Encoding.Unicode.GetString in many more cases, or dotnet/coreclr#18487 that removed a bunch of unnecessary virtual indirections from various encoding implementations, or PR dotnet/coreclr#20768 that improved the performance of Encoding.Preamble by taking advantage of the same metadata-blob span support discussed earlier, or PRs dotnet/coreclr#21948 and dotnet/coreclr#23098 that overhauled and streamlined the implementions of UTF8Encoding and AsciiEncoding.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
ASCII netcoreapp2.1 66.92 ns 0.8942 ns 0.8364 ns 1.00 0.0609 128 B
ASCII netcoreapp3.0 28.04 ns 0.6325 ns 0.9467 ns 0.42 0.0612 128 B

 

These examples all served to highlight improvements made in and around strings. That’s all well and good, but where the improvements related to strings really start to shine is when looking at improvements around formatting and parsing.

Parsing/Formatting

Parsing and formatting are the lifeblood of any modern web app or service: take data off the wire, parse it, manipulate it, format it back out. As such, in .NET Core 2.1 along with bringing up Span<T>, we invested in the formatting and parsing of primitives, from Int32 to DateTime. Many of those changes can be read about in my previous blog posts, but one of the key factors in enabling those performance improvements was in moving a lot of native code to managed. That may be counter-intuitive, in that it’s “common knowledge” that C code is faster than C# code. However, in addition to the gap between them narrowing, having (mostly) safe C# code has made the code base easier to experiment in, so whereas we may have been skittish about tweaking the native implementations, the community-at-large has dived head first into optimizing these implementations wherever possible. That effort continues in full force in .NET Core 3.0, with some very nice rewards reaped.

Let’s start with core integer primitives. PR dotnet/coreclr#18897 added a variety of special paths for the parsing of Integer-style signed values (e.g. Int32 and Int64), PR dotnet/coreclr#18930 added similar support for unsigned (e.g. UInt32 and UInt64), and PR dotnet/coreclr#18952 did a similar pass for hex. On top of those, PR dotnet/coreclr#21365 layered in additional optimizations, for example utilizing those changes for primitives like byte, skipping unnecessary layers of functions, streamlining some calls to improve inlining, and further reducing branching. The net impact here are some significant improvements to the performance of parsing integer primitive types in this release.

Method Toolchain Mean Error StdDev Ratio
ParseInt32Dec netcoreapp2.1 77.30 ns 0.8710 ns 0.8147 ns 1.00
ParseInt32Dec netcoreapp3.0 16.08 ns 0.2168 ns 0.2028 ns 0.21
ParseInt32Hex netcoreapp2.1 69.01 ns 1.0024 ns 0.9377 ns 1.00
ParseInt32Hex netcoreapp3.0 17.39 ns 0.1123 ns 0.0995 ns 0.25

 

Formatting of such types was also improved, even though it had already been improved significantly between .NET Core 2.0 and .NET Core 2.1. PR dotnet/coreclr#19551 tweaked the structure of the code to avoid needing to access the current culture number formatting data if it wouldn’t be needed (e.g. when formatting a value as hex, there’s no customization based on current culture), and PR dotnet/coreclr#18935 improved decimal formatting performance, in large part by optimizing how data is passed around (or not passed at all).

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
DecimalToString netcoreapp2.1 88.79 ns 1.4034 ns 1.3127 ns 1.00 0.0228 48 B
DecimalToString netcoreapp3.0 76.62 ns 0.5957 ns 0.5572 ns 0.86 0.0228 48 B

 

In fact, System.Decimal itself has been overhauled in .NET Core 3.0, as of PR dotnet/coreclr#18948now with an entirely managed implementation, and with additional performance work in PRs like dotnet/coreclr#20305.

Method Toolchain Mean Error StdDev Median Ratio RatioSD
Add netcoreapp2.1 12.021 ns 0.6813 ns 2.0088 ns 11.507 ns 1.00 0.00
Add netcoreapp3.0 8.300 ns 0.0553 ns 0.0518 ns 8.312 ns 0.87 0.04
Subtract netcoreapp2.1 13.026 ns 0.2599 ns 0.2431 ns 13.046 ns 1.00 0.00
Subtract netcoreapp3.0 8.613 ns 0.2024 ns 0.2770 ns 8.488 ns 0.66 0.03
Multiply netcoreapp2.1 19.215 ns 0.2813 ns 0.2631 ns 19.229 ns 1.00 0.00
Multiply netcoreapp3.0 7.182 ns 0.1795 ns 0.2457 ns 7.131 ns 0.38 0.01
Divide netcoreapp2.1 196.827 ns 4.3572 ns 4.6621 ns 194.721 ns 1.00 0.00
Divide netcoreapp3.0 75.456 ns 1.5301 ns 1.7007 ns 75.089 ns 0.38 0.01
Mod netcoreapp2.1 464.968 ns 7.0295 ns 6.5754 ns 466.825 ns 1.00 0.00
Mod netcoreapp3.0 13.756 ns 0.2476 ns 0.2316 ns 13.729 ns 0.03 0.00
Floor netcoreapp2.1 33.593 ns 0.8348 ns 2.2710 ns 32.734 ns 1.00 0.00
Floor netcoreapp3.0 12.109 ns 0.1325 ns 0.1239 ns 12.085 ns 0.33 0.02
Round netcoreapp2.1 32.181 ns 0.5660 ns 0.5294 ns 32.018 ns 1.00 0.00
Round netcoreapp3.0 12.798 ns 0.1572 ns 0.1394 ns 12.808 ns 0.40 0.01

 

Back to formatting and parsing, there are even some new formatting special-cases that might look silly at first, but that represent optimizations targeting real-world cases. In some sizeable web applications, we found that a large number of strings on the managed heap were simple integral values like “0” and “1”. And since the fastest code is code you don’t need to execute at all, why bother allocating and formatting these small numbers over and over when we can instead just cache and reuse the results (effectively our own string interning pool)? That’s what PR dotnet/coreclr#18383 does, creating a small, specialized cache of the strings for “0” through “9”, and any time we now find ourselves formatting a single-digit integer primitive, we instead just grab the relevant string from this cache.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
SingleDigitToString netcoreapp2.1 17.72 ns 0.3273 ns 0.3061 ns 1.00 0.0152 32 B
SingleDigitToString netcoreapp3.0 11.57 ns 0.1750 ns 0.1551 ns 0.65

 

Enums have also seen sizable parsing and formatting improvements in .NET Core 3.0. PR dotnet/coreclr#21214 improved the handling of Enum.Parse and Enum.TryParse, for both the generic and non-generic variants. PR dotnet/coreclr#21254 improved the performance of ToString when dealing with [Flags] enums, and PR dotnet/coreclr#21284 further improved other ToString cases. The net effect of these changes is a sizeable improvement in Enum-related performance:

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
EnumParse netcoreapp2.1 154.42 ns 1.6917 ns 1.5824 ns 1.00 0.0114 24 B
EnumParse netcoreapp3.0 62.92 ns 1.2239 ns 1.1448 ns 0.41
EnumToString netcoreapp2.1 85.81 ns 1.6458 ns 1.3743 ns 1.00 0.0305 64 B
EnumToString netcoreapp3.0 27.89 ns 0.6076 ns 0.7901 ns 0.32 0.0114 0.0001 24 B

 

In .NET Core 2.1, DateTime.TryFormat and ToString were optimized for the commonly-used “o” and “r” formats; in .NET Core 3.0, the parsing equivalents get a similar treatment. PR dotnet/coreclr#18800 significantly improves the performance of parsing DateTime{Offset}s formatted with the Roundtrip “o” format, and PR dotnet/coreclr#18771 does the same for the RFC1123 “r” format. For any serialization formats heavy in DateTimes, these improvements can make a substantial impact:

Method Toolchain Mean Error StdDev Median Ratio Gen 0 Gen 1 Gen 2 Allocated
ParseR netcoreapp2.1 2,254.6 ns 44.340 ns 45.534 ns 2,263.2 ns 1.00 0.0420 96 B
ParseR netcoreapp3.0 113.7 ns 3.440 ns 9.926 ns 112.6 ns 0.06
ParseO netcoreapp2.1 1,337.1 ns 26.542 ns 68.987 ns 1,363.8 ns 1.00 0.0744 160 B
ParseO netcoreapp3.0 354.9 ns 4.801 ns 3.748 ns 354.9 ns 0.30

 

Tying back to the StringBuilder discussion from earlier, default DateTime formatting was also improved by PR dotnet/coreclr#22111, tweaking how DateTime internally interacts with a StringBuilder that’s used to build up the resulting state.

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
DateTimeToString netcoreapp2.1 337.8 ns 6.560 ns 5.815 ns 1.00 0.00 0.0834 176 B
DateTimeToString netcoreapp3.0 269.4 ns 5.274 ns 5.416 ns 0.80 0.02 0.0300 64 B

 

TimeSpan formatting was also significantly improved, via PR dotnet/coreclr#18990:

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
TimeSpanToString netcoreapp2.1 151.11 ns 2.0037 ns 1.874 ns 1.00 0.0303 64 B
TimeSpanToString netcoreapp3.0 34.73 ns 0.7680 ns 1.304 ns 0.23 0.0305 64 B

 

Guid parsing also got in the perf-optimization game, with PR dotnet/coreclr#20183 improved parsing performance of Guid, primarily by avoiding overhead in helper routines, as well as by avoiding some searches used to determine which parsing routines to employ.

Method Toolchain Mean Error StdDev Median Ratio
ParseGuid netcoreapp2.1 287.5 ns 11.606 ns 28.688 ns 277.2 ns 1.00
ParseGuid netcoreapp3.0 111.7 ns 2.199 ns 2.057 ns 112.4 ns 0.33

 

Related, PR dotnet/coreclr#21336 again takes advantage of vectorization to improve Guid‘s construction and formatting to and from byte arrays and spans:

Method Toolchain Mean Error StdDev Ratio
GuidToFromBytes netcoreapp2.1 16.623 ns 0.2917 ns 0.2586 ns 1.00
GuidToFromBytes netcoreapp3.0 5.701 ns 0.1047 ns 0.0980 ns 0.34

 

Regular Expressions

Often related to parsing is the area of regular expressions. A bit of work was done on System.Text.RegularExpressions in .NET Core 3.0. PR dotnet/corefx#30474 replaced some usage of an internal StringBuilder cache with a ref struct-based builder that takes advantage of stack-allocated space and pooled buffers. And PR dotnet/corefx#30632 continued the effort by taking further advantage of spans. But the biggest improvement came in PR dotnet/corefx#32899 from @Alois-xx, which tweaks the code generated for a RegexOptions.Compiled Regex to avoid gratuitous thread-local accesses to look up the current culture. This is particularly impactful when also using RegexOptions.IgnoreCase. To see the impact, I found a complicated Regex that used both Compiled and IgnoreCase, and put it into a benchmark:

Method Toolchain Mean Error StdDev Ratio RatioSD
RegexCompiled netcoreapp2.1 1.946 us 0.0406 us 0.0883 us 1.00 0.00
RegexCompiled netcoreapp3.0 1.209 us 0.0432 us 0.1254 us 0.64 0.08

 

Threading

Threading is one of those things that’s ever-present and yet most apps and libraries don’t need to explicitly interact with most of the time. That makes it an area ripe for runtime performance improvements to drive down overhead as much as possible, so that user code just gets faster. Previous releases of .NET Core saw a lot of investment in this area, and .NET Core 3.0 continues the trend. This is another area where new APIs have been exposed and then also used in .NET Core itself for further gain.

For example, historically the only work item types that could be queued to the ThreadPool were ones implemented in the runtime, namely those created by ThreadPool.QueueUserWorkItem and friends, by Task, by Timer, and other such core types. But in .NET Core 3.0, the ThreadPool has an UnsafeQueueUserWorkItem overload that accepts the newly public IThreadPoolWorkItem interface. This interface is very simple, with a single method that just Executes work, and that means that any object that implements this interface can be queued directly to the thread pool. This is advanced; most code is just fine using the existing work item types. But this additional option affords a lot of flexibility, in particular in being able to implement the interface on a reusable object that can be queued over and over again to the pool. This is now used in a bunch of additional places in .NET Core 3.0.

One such place is in System.Threading.Channels. The Channels library introduced in .NET Core 2.1 already had a fairly low allocation profile, but there were still times it would allocate. For example, one of the options when creating a channel is whether continuations created by the library should run synchronously or asynchronously as part of a task completing (e.g. when a TryWrite call on a channel wakes up a corresponding ReadAsync, is the continuation from that ReadAsyncinvoked synchronously or queued by the TryWrite call). The default is that continuations are never invoked synchronously, but that also then requires allocating an object as part of queueing the continuation to the thread pool. With PR dotnet/corefx#33080, the reusable IValueTaskSource implementation that already backs the ValueTasks returned from ReadAsync calls also implements IThreadPoolWorkItem and can thus itself be queued, avoiding that allocation. This can have a measurable impact on throughput.

Method Job NuGetReferences Toolchain Mean Error StdDev Gen 0 Gen 1 Gen 2
PingPong 4.5.0 System.Threading.Channels 4.5.0 .NET Core 2.1 22.44 ms 0.3246 ms 0.4757 ms 593.7500
PingPong 4.6.0-preview5.19224.8 System.Threading.Channels 4.6.0-preview5.19224.8 .NET Core 3.0 16.81 ms 0.4246 ms 0.6356 ms 31.2500

 

IThreadPoolWorkItem is now also utilized in other places, like in ConcurrentExclusiveSchedulerPair (a little known but useful type that provides an exclusive scheduler that limits execution to only one task at a time, a concurrent scheduler that limits a user-defined number of tasks to run at a time, and that coordinate with each other so that no concurrent tasks may run while an exclusive task is running, ala a reader-writer lock), which now implements IThreadPoolWorkItem on an internally reusable work item object such that it also can avoid allocations when queueing its own processors. It’s also used in ASP.NET Core, and is one of the reasons key ASP.NET benchmarks are ammortized to 0 allocations per request. But by far the most impactful new implementer is in the async/await infrastructure.

In .NET Core 2.1, the runtime’s support for async/await was overhauled, drastically reducing the overheads involved in async methods. Previously when an async method awaited for the first time an awaitable that wasn’t yet complete, the struct-based state machine for the async method would be boxed (literally a runtime box) to the heap. With .NET Core 2.1, we changed that to instead use a generic object that stores the struct as a field on it. This has a myriad of benefits, but one of these benefits is that it now enables us to implement additional interfaces on that object, such as implementing IThreadPoolWorkItem. PR dotnet/coreclr#20159 does exactly that, and it enables another large swath of scenarios to have further reduced allocations, in particular situations where TaskCreationOptions.RunContinuationsAsynchronously was used with a TaskCompletionSource<T>. This can be seen in a benchmark like the following.

Method Job NuGetReferences Toolchain Mean Error StdDev Gen 0 Gen 1 Gen 2
AsyncAllocs 4.5.0 System.Threading.Channels 4.5.0 .NET Core 2.1 2.396 s 0.0486 s 0.0728 s 96000.0000
AsyncAllocs 4.6.0-preview5.19224.8 System.Threading.Channels 4.6.0-preview5.19224.8 .NET Core 3.0 1.512 s 0.0256 s 0.0359 s 49000.0000

 

That change allowed subsequent optimizations, such as PR dotnet/coreclr#20186 using it to make await Task.Yield(); allocation-free:

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0
Yield netcoreapp2.1 581.3 ms 11.615 ms 30.39 ms 1.00 0.00 19000.0000
Yield netcoreapp3.0 464.4 ms 9.087 ms 10.46 ms 0.81 0.06

 

It’s even utilized further in Task itself. There’s an interesting race condition that has to be handled in awaitables: what happens if the awaited operation completes after the call to IsCompleted but before the call to OnCompleted? As a reminder, the code:

compiles down to code along the lines of:

Once we go down the path of IsCompleted having returned false, we’re going to call AwaitOnCompleted and return. If the operation has completed by the time we call AwaitOnCompleted, we don’t want to synchronously invoke the continuation that re-enters this state machine, as we’ll be doing so further down the stack, and if that happened repeatedly, we’d “stack dive” and could end up overflowing the stack. Instead, we’re forced to queue the continuation. This case isn’t the common case, but it happens more often than you might expect, as it simply requires an operation that completes asynchronously very quickly (various networking operations often fall into this category). As of PR dotnet/coreclr#22373, the runtime now takes advantage of the async state machine box object implementing IThreadPoolWorkItem to avoid the allocations in this case as well!

In addition to IThreadPoolWorkItem being used with async/await to allow the async implementation to queue work items to the thread pool in a more allocation-friendly manner just as any other code can, changes were also made that give the ThreadPool 1st-hand knowledge of the state machine box in order to help it optimize additional cases. PR dotnet/coreclr#21159 from @benaadams teaches the ThreadPool to re-route some UnsafeQueueUserWorkItem(Action<object>, object, bool) calls to instead use UnsafeQueueUserWorkItem(IAsyncStateMachineBox, bool) under the covers, so that higher-level libraries can get these allocation benefits without having to be aware of the box machinery.

Another async-related area that’s seen measurable improvements are Timers. In .NET Core 2.1, some important improvements were made to System.Threading.Timers to help improve throughput and minimize contention for a common case where timers aren’t firing, but instead are quickly created and destroyed. And while those changes help a bit with the case when timers do actually fire, they didn’t help with the majority costs and sources of contention in that case, which is that potentially a lot of work (proportional to the number of timers registered) was done while holding locks. .NET Core 3.0 makes some big improvements here. PR dotnet/coreclr#20302 partitions the internal list of registered timers into two lists: one with timers that will soon fire and one with timers that won’t fire for a while. In most workloads that have a lot of registered timers, the majority of timers fall into the latter bucket at any given point in time, and this partitioning scheme enables the runtime to only consider the small bucket when firing timers most of the time. In doing so, it can significantly reduce the costs involved in firing timers, and as a result, also significantly reduce contention on the lock held while manipulating those lists. One customer who tried out these changes after having experienced issues due to tons of active timers had this to say about the impact:

“We got the change in production yesterday and the results are amazing, with 99% reduction in lock contention. We have also measured 4-5% CPU gains, and more importantly 0.15% improvement in reliability for our service (which is huge!).”

The nature of the scenario makes it a little difficult to see the impact in a Benchmark.NET benchmark, so we’ll do something a little different. Rather than measuring the thing that was actually changed, we’ll measure something else that’s indirectly impacted. In particular, these changes didn’t directly impact the performance of creating and destroying timers; in fact, one of the goals was to avoid doing so (in particular to avoid harming that important path). But by reducing the costs of firing timers, we reduce how long locks are held, which then also reduces the contention that the creating/destroying of timers faces. So, our benchmark creates a bunch of timers, ranging in when and how often they fire, and then we time how long it takes to create and destroy a bunch of additional timers.

Method Toolchain Mean Error StdDev Median Ratio RatioSD Gen 0
CreateDestroy netcoreapp2.1 289.1 us 7.131 us 20.687 us 282.8 us 1.00 0.00 80.0781
CreateDestroy netcoreapp3.0 199.5 us 3.983 us 5.584 us 199.2 us 0.71 0.04 80.3223

 

Timer improvements have also taken other forms. For example, PR dotnet/coreclr#22233 from @benaadams shrinks the allocation involved in Task.Delay when used without a CancellationToken by 24 bytes, and PR dotnet/coreclr#20509 reduces the timer-related allocations involved in creating timed CancellationTokenSources, which also has a nice effect on throughput:

Method Toolchain Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
CTSTimer netcoreapp2.1 231.3 ns 6.293 ns 16.018 ns 224.8 ns 1.00 0.00 0.0987 208 B
CTSTimer netcoreapp3.0 115.3 ns 1.769 ns 1.655 ns 115.0 ns 0.46 0.04 0.0764 160 B

 

There are other even lower-level improvements that have gone into the release. For example, PR dotnet/coreclr#21328 from @benaadams improved Thread.CurrentThread by changing the implementation to store the relevant Thread in a [ThreadStatic] field rather than forcing CurrentThread to make an InternalCall into the native portions of the runtime.

Method Toolchain Mean Error StdDev Ratio RatioSD
CurrentThread netcoreapp2.1 6.101 ns 0.2587 ns 0.7547 ns 1.00 0.00
CurrentThread netcoreapp3.0 2.822 ns 0.0439 ns 0.0389 ns 0.45 0.04

 

As other examples, PR dotnet/coreclr#23747 taught the runtime to better respect Docker –cpu limits, PRs dotnet/coreclr#21722 and dotnet/coreclr#21586 improved spinning behavior when contention was encountered across a variety of synchronization sites, PR dotnet/coreclr#22686 improved performance of SemaphoreSlim when consumers of an instance were mixing both synchronous Waits and asynchronous WaitAsyncs, and PR dotnet/coreclr#18098 from @Quogu special-cased CancellationTokenSource created with a timeout of 0 to avoid Timer-related costs.

 

Collections

Moving on from threading, let’s explore some of the performance improvements that have gone into collections. Collections are so commonly used in pretty much every program that they’ve received a lot of performance-focused attention in previous .NET Core releases. Even so, there continues to be areas for improvement. Here are some example such improvements in .NET Core 3.0.

ConcurrentDictionary<TKey, TValue> has an IsEmpty property that states whether the dictionary is empty at that moment-in-time. In previous releases, it took all of the dictionary’s locks in order to get a proper moment-in-time answer. But as it turns out, those locks only need to be held if we think the collection might be empty: if we see anything in any of the dictionary’s internals buckets, the locks aren’t needed, as we’d stop looking at additional buckets anyway the moment we found one bucket to contain anything. Thus, PR dotnet/corefx#30098 from @drewnoakes added a fast path that first checks each bucket without the locks, in order to optimize for the common case where the dictionary isn’t empty (the impact on the case where the dictionary is empty is minimal).

Method Toolchain Mean Error StdDev Ratio
IsEmpty netcoreapp2.1 73.675 ns 0.3934 ns 0.3285 ns 1.00
IsEmpty netcoreapp3.0 3.160 ns 0.0402 ns 0.0356 ns 0.04

 

ConcurrentDictionary wasn’t the only concurrent collection to get some attention. An improvement came to ConcurrentQueue<T> in dotnet/coreclr#18035, and it’s an interesting example in how performance optimization often is a trade-off between scenarios. In .NET Core 2.0, we overhauled the ConcurrentQueue implementation in a way that significantly improved throughput while also significantly reducing memory allocations, turning the ConcurrentQueue into a linked list of circular arrays. However, the change involved a concession: because of the producer/consumer nature of the arrays, if any operation needed to observe data in-place in a segment (rather than dequeueing it), the segment that was observed would be “frozen” for any further enqueues… this was to avoid problems where, for example, one thread was enumerating the contents of the segment while another thread was enqueueing and dequeueing. When there were multiple segments in the queue, accessing Count ended up being treated as an observation, but that meant that simply accessing the ConcurrentQueue‘s Count would render all of the multiple segments in the queue dead for further enqueues. The theory at the time was that such a trade-off was fine, because no one should be accessing the Count of the queue frequently enough for this to matter. That theory was wrong, and several customers reported significant slowdowns in their workloads because they were accessing the Count on every enqueue or dequeue. While the right solution is in general to avoid doing that, we wanted to fix this, and as it turns out, the fix was relatively straightforward, such that we could have our performance cake and eat it, too. The results are very obvious in the following benchmark.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
EnqueueCountDequeue netcoreapp2.1 708.48 ns 23.8638 ns 21.1546 ns 1.00 0.1669 0.0830 0.0010 704 B
EnqueueCountDequeue netcoreapp3.0 22.79 ns 0.4471 ns 0.4182 ns 0.03

 

ImmutableDictionary<TKey, TValue> also got some attention. A customer reported that they’d compared ImmutableDictionary<TKey, TValue> and Dictionary<TKey, TValue> and found the former to be measurably slower for lookups. This is to be expected, as the types use very different data structures, with ImmutableDictionary optimized for being able to inexpensively create a copy of the dictionary with a mutation, something that’s quite expensive to do with Dictionary; the trade-off is that it ends up being slower for lookups. Still, it caused us to take a look at the costs involved in ImmutableDictionary lookups, and PR dotnet/corefx#35759 includes several tweaks to improve it, changing a recursive call to be non-recursive and inlinable and avoiding some unnecessary struct wrapping. While this doesn’t make ImmutableDictionary and Dictionary lookups equivalent, it does improve ImmutableDictionarymeasurably, especially when it contains just a few elements.

Method Toolchain Mean Error StdDev Median Ratio RatioSD
Lookup netcoreapp2.1 303.9 us 7.271 us 15.016 us 297.8 us 1.00 0.00
Lookup netcoreapp3.0 174.5 us 3.360 us 2.806 us 174.5 us 0.57 0.03

 

Another collection that’s seen measurable improvements in .NET Core 3.0 is BitArray. Lots of operations, including construction, were optimized in PR dotnet/corefx#33367.

Method Toolchain Mean Error StdDev Median Ratio RatioSD
BitArrayCtor netcoreapp2.1 82.28 ns 2.601 ns 7.546 ns 77.89 ns 1.00 0.00
BitArrayCtor netcoreapp3.0 46.87 ns 2.738 ns 8.030 ns 44.63 ns 0.57 0.10

 

Core operations like Set and Get were further improved in PR dotnet/corefx#35364 from @omariom by streamlining the relevant methods and making them inlineable

Method Toolchain Mean Error StdDev Ratio
GetSet netcoreapp2.1 6.497 us 0.0854 us 0.0713 us 1.00
GetSet netcoreapp3.0 2.049 us 0.0233 us 0.0218 us 0.32

 

while other operations like OrAnd, and Xor were vectorized in PR dotnet/corefx#33781. This benchmark highlights some of the wins.

Method Toolchain Mean Error StdDev Ratio
Xor netcoreapp2.1 28.57 ns 0.4086 ns 0.3822 ns 1.00
Xor netcoreapp3.0 10.92 ns 0.0924 ns 0.0772 ns 0.38

 

Another example: SortedSet<T>. PR dotnet/corefx#30921 from @acerbusace tweaks how GetViewBetween changes how counts of the overall set and subset are managed, resulting in a nice performance boost.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
EnumerateViewBetween netcoreapp2.1 5.117 us 0.0590 us 0.0552 us 1.00 0.2518 544 B
EnumerateViewBetween netcoreapp3.0 2.510 us 0.0307 us 0.0287 us 0.49 0.1373 288 B

 

Comparers have also seen some nice improvements in .NET Core 3.0. For example, PR dotnet/coreclr#21604 overhauled how comparers for enums are implemented in the runtime, borrowing the approach used in CoreRT. It’s often the case that performance optimizations involve adding code; this is one of those fortuitous cases where the better approach is not only faster, it’s also simpler and smaller.

Method Toolchain Mean Error StdDev Ratio RatioSD
CompareEnums netcoreapp2.1 239.5 ms 10.130 ms 10.403 ms 1.00 0.00
CompareEnums netcoreapp3.0 131.7 ms 2.479 ms 2.319 ms 0.55 0.03

 

Networking

From the Kestrel web server running on System.Net.Sockets and System.Net.Security to applications accessing web services via HttpClientSystem.Net now more than ever is critical path for many applications. It received a lot of attention in .NET Core 2.1, and continues to in .NET Core 3.0.

Let’s start with HttpClient. One improvement made in PR dotnet/corefx#32820 was around how buffering is handled, and in particular better respecting larger buffer size requests made as part of copying the response data when a content length was provided by the server. On a fast connection and with a large response body (such as the 10MB in this example), this can make a sizeable difference in throughput due to reduced syscalls to transfer data.

Method Toolchain Mean Error StdDev Ratio RatioSD
HttpDownload netcoreapp2.1 8.792 ms 0.1833 ms 0.3397 ms 1.00 0.00
HttpDownload netcoreapp3.0 4.615 ms 0.0356 ms 0.0278 ms 0.52 0.02

 

Now consider SslStream. Previous releases saw work done to make reads and writes on SslStream much more efficient, but additional work was done in .NET Core 3.0 as part of PRs dotnet/corefx#35091 and dotnet/corefx#35209 (and dotnet/corefx#35367 on Unix) to make initiating the connection more efficient, in particular in terms of allocations.

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
SslConnect netcoreapp2.1 1,151.7 us 34.85 us 102.76 us 1.00 0.00 5.8594 9.82 KB
SslConnect netcoreapp3.0 915.5 us 17.73 us 26.54 us 0.80 0.08 1.9531 4.13 KB

 

In System.Net.Sockets there’s another example of taking advantage of the IThreadPoolWorkItem interface discussed earlier. On Windows for asynchronous operations, we utilize “overlapped I/O”, utilizing threads from the I/O thread pool to execute continuations from socket operations; Windows queues I/O completion packets that these I/O pool threads then process, including invoking the continuations. On Unix, however, the mechanism is very different. There’s no concept of “overlapped I/O” on Unix, and instead asynchrony in System.Net.Sockets is achieved by using epoll (or kqueues on macOS), with all of the sockets in the system registered with an epoll file descriptor, and then one thread monitoring that epoll for changes. Any time an asynchronous operation completes for a socket, the epoll is signaled and the thread blocking on it wakes up to process it. If that thread were to run the socket continuation action then and there, it would end up potentially running unbounded work that could stall every other socket’s handling indefinitely, and in the extreme case, deadlock. Instead, this thread queues a work item back to the thread pool and then immediately goes back to processing any other socket work. Prior to .NET Core 3.0, that queueing involved an allocation, which meant that every asynchronously completing socket operation on Unix involved at least one allocation. As of PR dotnet/corefx#32919, that number drops to zero, as a cached object already being used (and reused) to represent asynchronous operations was changed to also implement IThreadPoolWorkItem and be queueable directly to the thread pool.

Other areas of System.Net have benefited from the efforts already alluded to previously, as well. For example, Dns.GetHostName used to use StringBuilder in its marshaling, but as of PR dotnet/corefx#29594 it no longer does.

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
GetHostName netcoreapp2.1 85.77 us 1.656 us 1.5489 us 1.00 0.00 0.4883 1176 B
GetHostName netcoreapp3.0 81.42 us 1.016 us 0.9503 us 0.95 0.02 48 B

 

And IPAddress.HostToNetworkOrder/NetworkToHostOrder have benefiting indirectly from the intrinsics push that was mentioned previously. In .NET Core 2.1, BinaryPrimitives.ReverseEndianness was added with an optimized software implementation, and these IPAddress methods were rewritten as simple wrappers for ReverseEndianness. Now in .NET Core 3.0, PR dotnet/coreclr#18398 turned ReverseEndianness into a JIT intrinsic for which the JIT can emit a very efficient BSWAP instruction, with the resulting throughput improvements accruing to IPAddress as well.

Method Toolchain Mean Error StdDev Median Ratio RatioSD
HostToNetworkOrder netcoreapp2.1 0.4986 ns 0.0398 ns 0.0408 ns 0.4758 ns 1.000 0.00
HostToNetworkOrder netcoreapp3.0 0.0043 ns 0.0090 ns 0.0076 ns 0.0000 ns 0.009 0.02

System.IO

Often going hand in hand with networking is compression, which has also seen some improvements in .NET Core 3.0. Most notably is that a key dependency was updated. On Unix, System.IO.Compression just uses the zlib library available on the machine, as it’s a standard part of most any distro/version. On Windows, however, zlib is generally nowhere to be found, and so it’s built and shipped as part of .NET Core on Windows. Rather than shipping the standard zlib, .NET Core includes a version modified by Intel with additional performance improvements not yet merged upstream. In .NET Core 3.0, we’ve sync’d to the latest available version of ZLib-Intel, version 1.2.11. This brings some very measurable performance improvements, in particular around decompression.

There have also been compression-related improvements that take advantage of previous improvements elsewhere in .NET Core. For example, the synchronous Stream.CopyTo was originally non-virtual, but as gains were found by overriding the asynchronous CopyToAsync and specializing its implementation for particular concrete stream types, CopyTo was made virtual to enjoy similar improvements. PR dotnet/corefx#29751 capitalized on this to override CopyTo on DeflateStream, employing similar optimizations in the synchronous implementation as were employed in the asynchronous implementation, essentially entailing minimizing the interop costs with zlib.

Method Toolchain Mean Error StdDev Ratio
DeflateDecompress netcoreapp2.1 310.6 us 1.960 us 1.6367 us 1.00
DeflateDecompress netcoreapp3.0 144.9 us 1.050 us 0.9819 us 0.47

 

Improvements were also made to BrotliStream (which as of .NET Core 3.0 is also used by HttpClient to automatically decompress Brotli-encoded content). Previously every new BrotliStream would also allocate a large buffer, but as of PR dotnet/corefx#35492, that buffer is pooled, as it is with DeflateStream (additionally, BrotliStream now as of PR dotnet/corefx#30135 overrides ReadByte and WriteByte to avoid allocations in the base implementation).

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
BrotliWrite netcoreapp2.1 743.2 us 10.056 us 9.406 us 1.00 44.9219 97680 B
BrotliWrite netcoreapp3.0 575.5 us 9.181 us 8.588 us 0.77 136 B

 

Moving on from compression, it’s worth highlighting that formatting applies in more situations than just formatting individual primitives. TextWriter, for example, has multiple methods for writing with format strings, e.g. public override void Write(string format, object arg0, arg1). PR dotnet/coreclr#19235 improved on that for StreamWriter by providing specialized overrides that take a more efficient path that reduces allocation:

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
StreamWriterFormat netcoreapp2.1 207.4 ns 2.103 ns 1.864 ns 1.00 0.0455 96 B
StreamWriterFormat netcoreapp3.0 170.2 ns 1.800 ns 1.595 ns 0.82 0.0114 24 B

 

As another example, PR dotnet/coreclr#22102 from @TomerWeisberg improved the parsing performance of various primitive types on BinaryReader by special-casing the common situation where the BinaryReaderwraps a MemoryStream.

Or consider PR dotnet/corefx#30667 from @MarcoRossignoli, who added overrides to StringWriter for the Write{Line}{Async} methods that take a StringBuilder argument. StringWriter is just a wrapper around a StringBuilder, and StringBuilder knows how to append another StringBuilder to it, so these overrides on StringWriter can feed them right through.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
Write netcoreapp2.1 30.15 ns 0.6065 ns 0.5673 ns 1.00 0.0495 104 B
Write netcoreapp3.0 18.57 ns 0.1513 ns 0.1416 ns 0.62

 

System.IO.Pipelines is another IO-related library that’s received a lot of attention in .NET Core 3.0. Pipelines was introduced in .NET Core 2.1, and provides buffer-management as part of an I/O pipeline, used heavily by ASP.NET Core. A variety of PRs have gone into improving its performance. For example, PR dotnet/corefx#35171special-cases the common and default case where the Pool specified to be used by a Pipe is the default MemoryPool<byte>.Shared. Rather than go through MemoryPool<byte>.Shared in this case, the Pipe now bypasses it and goes to the underlying ArrayPool<byte>.Shared directly, which removes a layer of indirection but also the allocation of IMemoryOwner<byte> objects returned from MemoryPool<byte>.Rent. (Note that for this benchmark, since System.IO.Pipelines is part of a NuGet package rather than in the shared framework, I’ve added a Benchmark.NET config that specifies what package version to use with each run in order to show the improvements.)

Method Job NuGetReferences Toolchain Mean Error StdDev Gen 0 Gen 1 Gen 2
ReadWrite 4.5.0 System.IO.Pipelines 4.5.0 .NET Core 2.1 406.8 us 12.774 us 17.907 us 11.2305
ReadWrite 4.6.0-preview5.19224.8 System.IO.Pipelines 4.6.0-preview5.19224.8 .NET Core 3.0 324.6 us 3.208 us 4.702 us

 

PR dotnet/corefx#33658 from @benaadams allows Pipe to use the UnsafeQueueUserWorkItemboxing-related optimizations described earlier, PR dotnet/corefx#33755 avoids queueing unnecessary work items, PR dotnet/corefx#35939 tweaks the defaults used to better handle buffering in common cases, PR dotnet/corefx#35216 reduces the amount of slicing performed in various pipe operations, PR dotnet/corefx#35234 from @benaadams reduces the locking used in core operations, PR dotnet/corefx#35509 reduces argument validation (decreasing branching costs), PR dotnet/corefx#33000 focused on reducing costs associated with ReadOnlySequence<byte> that’s the main exchange type pipelines passes around, and PR dotnet/corefx#29837 further optimizes operations like GetSpan and Advance on the Pipe. The net result is to whittle away at already low CPU and allocation overheads.

Method Job NuGetReferences Toolchain Mean Error StdDev Gen 0 Gen 1 Gen 2
ReadWrite 4.5.0 System.IO.Pipelines 4.5.0 .NET Core 2.1 3.261 ms 0.0732 ms 0.1002 ms 46.8750
ReadWrite 4.6.0-preview5.19224.8 System.IO.Pipelines 4.6.0-preview5.19224.8 .NET Core 3.0 2.947 ms 0.1281 ms 0.1837 ms

 

System.Console

Console isn’t something one normally thinks of as being performance-sensitive. However, there are two changes in this release that I think are worth calling attention to here.

First, there is one area of Console about which we’ve heard numerous concerns related to performance, where the performance impact visibly impacts users. In particular, interactive console applications generally do a lot of manipulation of the cursor, which also entails asking where the cursor currently is. On Windows, both the setting and getting of the cursor are relatively fast operations, with P/Invoke calls made to functions exported from kernel32.dll. On Unix, things are more complicated. There’s no standard POSIX function for getting or setting a terminal’s cursor position. Instead, there’s a standard convention for interacting with the terminal via ANSI escape sequences. To set the cursor position, one writes a sequence of characters to stdout (e.g. “ESC [ 12 ; 34 H” to indicate 12th row, 34th column) and the terminal interprets that and reacts accordingly. Getting the cursor position is more of an ordeal. To get the current cursor position, an application writes to stdout a request (e.g. “ESC [ 6 n”), and in response the terminal writes back to the application’s stdin a response something like “ESC [ 12 ; 34 R”, to indicate the cursor is at the 12th row and 34th column. That response then needs to be read from stdin and parsed. So, in contrast to a fast interop call on Windows, on Unix we need to write, read, and parse text, and do so in a way that doesn’t cause problems with a user sitting at a keyboard using the app concurrently… not particularly cheap. When just getting the cursor position now and then, it’s not a big deal. But when getting it frequently, and when porting code originally written for Windows where the operation was so cheap the code being ported may not have been very frugal with how often it asked for the position (asking for it more than is really needed), this has resulted in visible performance problems. Thankfully, the issue has been addressed in .NET Core 3.0, by PR dotnet/corefx#36049 from @tmds. The change caches the current position and then manually handles updating that cached value based on user interactions, such as handling typing or resizing the terminal window. (Note that Benchmark.NET operates in a way that redirects standard input and output for the process running the test, and that makes Console.CursorLeft/Top return 0 immediately, so for this test, I’ve just done a simple console app with a Stopwatch, which is, as you’ll see, more than sufficient given the discrepancy between costs in versions.)

Another place where Console has been improved affects both Windows and Unix. Interestingly, this change was made for functional reasons (in particular for when running on Windows), but it has performance benefits as well for all OSes. In .NET, most of the times we specify buffer sizes it’s for performance reasons and represents a trade-off: the smaller the buffer size, the less memory is used but the more times operations may need to be performed to fill that buffer, and conversely the larger the buffer size, the more memory is used but the fewer times the buffer will need to be filled. It’s rare that the buffer size has a functional impact, but it actually can in Console. On Windows to read from the console, one calls either the ReadFile or ReadConsole functions, both of which accept a buffer to store the read data into. By default on Windows, reading from the console will not return until a newline, but Windows also needs somewhere to store the typed data, and it does so into the supplied buffer. Thus, Windows won’t let the user type more characters than can fit into the buffer, which means the line length a user can type is limited by the buffer size. For whatever historical reason, .NET has used a buffer size of 256 characters, limiting the typeable line length to that amount. PR dotnet/corefx#36212 expands that to 4096 characters, which much better matches other programming environments and allows for a much more reasonable line length. However, as is the case when increasing buffer sizes, relevant throughput involving that buffer improves as well, in particular when reading from files piped to stdin. For example, reading 8K of input data from stdin previously would have required 32 calls to ReadFile; with a 4K buffer, only 2 calls are required. The impact of that can be seen in this benchmark. (Again, this is harder to test with Benchmark.NET, so I’ve again just used a simple console app.)

System.Diagnostics.Process

There have been various functional improvements to the Process class in .NET Core 3.0, in particular on Unix, but there are a couple of performance-focused improvements I want to call out.

PR dotnet/corefx#31236 is another nice example of introducing a new performance-focused API and, at the same time, using it within .NET Core to further improve the performance of core libraries. In this case, it’s a low-level API on MemoryMarshal that enables efficiently reading structs from spans, something that’s done in spades as part of the interop in System.Diagnostics.Process. I like that example, not because it makes for a massive performance improvement, but because it highlights the general pattern I like to see: adding new APIs for others to consume and in the same breath using those APIs to better the technology itself.

A more impactful example, though, comes from @joshudson in PR dotnet/corefx#33289, which changed the native code used to fork a new process from using the fork function to instead using the vfork function. The benefit of vfork is that it avoids copying the page tables of the parent process into the child process, with the assumption that the child process is then just going to overwrite everything anyway via an almost immediate exec call. fork does copy-on-write, but if the process is modifying a lot of state concurrently (e.g. with the garbage collector running), this can get expensive quickly and unnecessarily. For this benchmark, I’ve just written a nop C program in a test.c file:

and compiled it with GCC:

to give us a target for Process.Start to invoke.

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
ProcessStartWait netcoreapp2.1 1,663.0 us 32.79 us 67.72 us 1.00 0.00 21.45 KB
ProcessStartWait netcoreapp3.0 536.0 us 10.64 us 28.40 us 0.32 0.02 1.9531 16.65 KB

 

LINQ

Previous releases have seen a ton of investment in optimizing LINQ. There’s less of that in .NET Core 3.0, as a lot of the common patterns have already been covered well. However, there are still some nice improvements to be found in the release.

It’s relatively rare that new operators are added to System.Linq itself, as the very nature of extension methods makes it easy for anyone to build up and share their own library of extension methods they consider to be useful (and several well-established such libraries exist). Even so, .NET Core 2.0 saw a new TakeLast method added. In .NET Core 3.0, PR dotnet/corefx#36051 by @Romasz updated TakeLast to integrate with the internal IPartition<T> interface that enables several operators to cooperate, helping to optimize (in some situations quite heavily) various uses of the operator.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
SumLast10 netcoreapp2.1 11,935.5 ns 102.793 ns 85.837 ns 1.00 0.1526 344 B
SumLast10 netcoreapp3.0 141.4 ns 1.310 ns 1.225 ns 0.01 0.0267 56 B

 

Just recently, PR dotnet/corefx#37410 optimized the relatively common pattern of using Enumerable.Range(...).Select(…), teaching Select about the object generated by Range and allowing for the enumeration performed by Select to skip going through IEnumerable<T> and instead just loop through the intended numerical range directly.

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
RangeSelectToArray netcoreapp2.1 953.9 ns 20.232 ns 28.363 ns 1.00 0.00 0.2460 520 B
RangeSelectToArray netcoreapp3.0 358.0 ns 7.650 ns 7.156 ns 0.37 0.02 0.2441 512 B

 

Enumerable.Empty<T>() was also changed in PR dotnet/corefx#31025 to better compose with optimizations already elsewhere in .NET Core’s System.Linq implementation. While no one should be writing code that explicitly calls additional LINQ operators directly on the result of Enumerable.Empty<T>(), it is common to return the result of Empty<T>() as one possible return value from an IEnumerable<T>-returning method, and then for the caller to tack on additional operators, such that this optimization does actually have a meaningful effect.

Method Toolchain Mean Error StdDev Ratio Gen 0 Gen 1 Gen 2 Allocated
EmptyTakeSelectToArray netcoreapp2.1 71.80 ns 1.4205 ns 1.1861 ns 1.00 0.0495 104 B
EmptyTakeSelectToArray netcoreapp3.0 30.09 ns 0.1550 ns 0.1295 ns 0.42

 

Across .NET Core, we’re also paying more attention to assembly size, in particular as it can impact ahead-of-time (AOT) compilation. PRs like dotnet/corefx#35213, which employs “ThrowHelpers” in the heavily-generic LINQ, help to reduce generated code size, which has benefits in and of itself but can also help with other areas of performance.

Interop

Interop is another one of those areas that’s critically important both to customers of .NET as well as to .NET itself, as a lot of functionality in .NET is layered on top of underlying operating system functionality that requires interop to access. As such, performance improvements in interop itself end up impacting a wide array of components.

One notable improvement is in SafeHandle, and it’s another example of where moving code from native to managed helped improve performance. SafeHandle is the recommended way for managing the lifetime of native resources, whether represented by handles on Windows or by file descriptors on Unix, and it’s used in exactly that way internally in all of our managed libraries in coreclr and corefx. One of the reasons it’s the recommended solution is that it uses appropriate synchronization to ensure that these native resources aren’t closed from managed code while they’re still being used, and that means that the interop layer needs to track every time a P/Invoke call is made with a SafeHandle, invoking DangerousAddRef prior to the call, DangerousRelease after the call, and DangerousGetHandle to extract the actual pointer value to pass to the native function. In previous releases of .NET, the core pieces of those implementations were in the runtime, which meant managed code needed to make InternalCalls to native code in the runtime for each of those operations. In .NET Core 3.0 as of PR dotnet/coreclr#22564, those operations have been ported to managed code, removing the overhead associated with each of those transitions.

Method Toolchain Mean Error StdDev Ratio
SafeHandleOps netcoreapp2.1 36.72 ns 0.7285 ns 0.6458 ns 1.00
SafeHandleOps netcoreapp3.0 16.04 ns 0.1322 ns 0.1104 ns 0.44

 

There are also examples for improvements to marshaling. Earlier in this post, I highlighted a variety of cases where StringBuilder was used as part of marshaling and interop. For the record, I personally dislike StringBuilder being used in interop, as it adds cost and complexity for relatively little benefit, and as a result did work in PRs like dotnet/corefx#33780 and dotnet/coreclr#21120 to remove almost all use of StringBuilder marshaling in coreclr and corefx. However, there is still a lot of code built around StringBuilder, and it deserves to be as fast as possible. PR dotnet/coreclr#17928 avoids a bunch of unnecessary work and allocation that happens as part of StringBuilder marshaling, and leads to improvements like this:

Method Toolchain Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
StringBuilderMarshal netcoreapp2.1 359.4 ns 7.643 ns 13.386 ns 1.00 0.00 0.2584 544 B
StringBuilderMarshal netcoreapp3.0 289.1 ns 5.773 ns 7.707 ns 0.80 0.04

 

And of course, specific uses of interop and marshaling have also improved. For example, FileSystemWatcher‘s interop on macOS had been using MarshalAs attributes, which forced the runtime to do additional marshaling work on every OS callback, including allocating arrays. PR dotnet/corefx#34715 moved FileSystemWatcher‘s interop to use a more efficient scheme that doesn’t entail additional allocations nor marshaling directives. Or consider dotnet/corefx#30099, where System.Drawing was switched to using a much more efficient scheme of marshaling and interop, with a managed array being pinned and passed directly to native code instead of allocating additional memory and copying to it.

Method Job NuGetReferences Toolchain Mean Error StdDev Gen 0 Gen 1 Gen 2 Allocated
TransformPoints 4.5.1 System.Drawing.Common 4.5.1 .NET Core 2.1 11,010.3 ns 490.050 ns 718.309 ns 0.5798 1248 B
TransformPoints 4.6.0-preview5.19224.8 System.Drawing.Common 4.6.0-preview5.19224.8 .NET Core 3.0 364.0 ns 6.704 ns 9.827 ns

 

Peanut butter

In previous sections of this post, I highlighted groups of PRs that addressed various areas of .NET in an impactful way, where some piece of mainstream functionality was significantly improved. But those aren’t the only areas or kinds of PRs that matter.

In .NET we also have what we sometimes refer to as “peanut butter”. We have a ton of code that’s generally great for most applications but that has a myriad of small opportunities for improvements. Those improvements alone don’t make anything better, but they fix a smearing of performance penalties across a large swath of code, and the more of such issues we can fix, the better performance becomes overall. An allocation removed here, some unnecessary cycles eliminated there, some unnecessary code removed there. Here are just a sampling of PRs that went in to address such “peanut butter”:

  • Lower bounds explicitly provided to Array.Copy. Calling Array.Copy(src, dst, length) requires the runtime to call GetLowerBound on each of the src and the dst arrays. When working with T[]s, the lower bound is 0, and we can just explicitly pass in 0 for both bounds and avoid the implicit GetLowerBound calls. PR dotnet/coreclr#21756 does that in a variety of places.
  • Cheaper copying to new arrays. In a variety of places, a List<T> stored some data, a new array was then allocated based on the length of the list, and the contents then copied to the array with CopyTo. PR dotnet/coreclr#22101 from @benaadams recognized the silliness of this and replaced that pattern with simply using List<T>.ToArray.
  • Nullable<T>.Value vs GetValueOrDefault. Nullable<T> has two main members to access the value: Value and GetValueOrDefault. It’s initially counter-intuitive, but GetValueOrDefault is actually cheaper: Value needs to check whether the instance has a value or not, throwing if it doesn’t, whereas GetValueOrDefault just always returns the value field, and it’ll be default if there was no value. PR dotnet/coreclr#22297 fixed up a variety of call sites where GetValueOrDefault could be used instead.
  • Array.Empty<T>(). In previous releases, lots of zero-length array allocations were changed to instead use Array.Empty<T>(), both in libraries and via compiler changes for things like params arrays. That trend continues in .NET Core 3.0, with PR dotnet/corefx#30235 doing another sweep through corefx and replacing even more zero-length allocations with the cached Array.Empty<T>().
  • Avoiding lots of little allocations all over the place. For new code being written, we’re very cost-conscious and keep an eye out for allocations that, even if small and rare, could be easily replaced by something less expensive. For existing code, the most impactful allocations show up in profiling of key scenarios and are squashed whenever possible. But there are a lot of small allocations here and there that generally don’t pop up on our radar until we have another reason to review and profile the relevant code. In every release, we end up removing a bunch of these. For example, all of these PRs contributed to reducing the allocation peanut butter across coreclr and corefx in .NET Core 3.0:
  • Avoiding explicit static cctors. Any type that has static fields initialized ends up with a static constructor (cctor) to run that initialization. But depending on how the initialization is authored can impact performance. In particular, if the developer explicitly writes a static cctor rather than initializing the fields as part of the static field declarations, the C# compiler will not mark the type as beforefieldinit. Having the type marked beforefieldinit can be beneficial for performance, because it allows the runtime more flexibility in when it performs the initialization, which in turn allows the JIT more flexibility about how it can optimize, and whether locking might be needed when accessing static methods on the type. PRs like dotnet/coreclr#21718 and dotnet/coreclr#21715 from @benaadams have removed such static cctors that can layer in small costs across a wide swath of accessing code.
  • Using a cheaper, sufficient equivalent. IndexOf on strings and spans returns the position of a found element, whereas Contains just returns whether the element was found. The latter can be slightly more efficient, because it doesn’t need to track the exact location of an element, just that it existed. Even so, lots of call sites that could have used Contains instead used IndexOf. PRs dotnet/coreclr#19874 and dotnet/corefx#32249 by @grant-d addressed that. Another example, SocketsHttpHandler(the default HttpMessageHandler behind HttpClient) was using DateTime.UtcNow when determining whether a connection could be reused for the next request or not, but Environment.TickCount is cheaper and has sufficient resolution and accuracy for this purpose, so PR dotnet/corefx#35401 switched it to use that. Another example, PR dotnet/corefx#37548 tweaks the overloads of Array.Copy used in a bunch of places to avoid unnecessary GetLowerBound() calls to lookup the lower bound for arrays we know have a lower bound of 0.
  • Simplifying interop. The interop infrastructure in .NET is quite powerful and comprehensive, with lots of knobs that allow for specifying how calls should be made and how data should be transformed. However, many come with a cost, such as needing the runtime to generate a marshaling stub to perform the various required transformations. PRs dotnet/corefx#36544 and dotnet/corefx#36071, for example, tweaked interop signatures to avoid overheads associated with such marshaling code.
  • Avoiding unnecessary globalization. Due to how various System.String APIs were designed almost two decades ago, it can be easy to accidentally employ culture-aware string comparisons when it’s not intended. Such comparisons can be functionally incorrect for a given task and also more costly, involving more expensive calls to the operating system or globalization library. In particular, String.IndexOf with a char argument uses ordinal comparison, but String.IndexOf with a string argument (even if it’s a single character) uses the current culture to perform the comparison. PRs dotnet/corefx#37499 addresses a bunch of such cases in System.Net, an area in which one almost always wants to do ordinal comparisons, generally the case when doing parsing for text-based protocols.
  • Avoiding unnecessary ExecutionContext flow. ExecutionContext is the primary vehicle for ambient state “flowing” through a program and across asynchronous calls, in particular AsyncLocal<T>. In order to achieve such flow, code that spawns an async operation (e.g. Task.RunTimer, etc.) or code that creates a continuation to run when some other operation finishes (e.g. await) needs to “capture” the current ExecutionContext, hang on to it, and then later when executing the relevant work, use that captured ExecutionContext‘s Run method to do so. If the work being performed doesn’t actually require the ExecutionContext, we can avoid flowing it to avoid the small associated overhead. PRs dotnet/corefx#37551dotnet/corefx#33235, and dotnet/corefx#33080 are examples: they switch several uses of CancellationToken.Registerover to the new CancellationToken.UnsafeRegister method, the only difference compared to Register being that it doesn’t flow ExecutionContext. As another example, PR dotnet/coreclr#18670 changed CancellationTokenSource so that when it creates a Timer, it doesn’t unnecessarily capture ExecutionContext. Or consider PR dotnet/coreclr#20294, which ensures that any such captured ExecutionContext is dropped as soon as it’s not needed from completed Tasks.
  • Centralized / optimized bit operations. PR dotnet/coreclr#22118 from @benaadams introduced a BitOperations class that serves to centralize a bunch of bit-twiddling operations (rotating, leading zero count, population count, log, etc.). This type was later augmented and enhanced in PRs from @grant-d like dotnet/coreclr#22497dotnet/coreclr#22584, and dotnet/coreclr#22630, which also serve to use these shared helpers from everywhere across System.Private.Corelib where such bit-twiddling operations are required. This ensures that all such call sites (of which there are currently ~70) get the best implementation the runtime can muster, whether that be an implementation that takes advantage of the current hardware’s instruction set or one that utilizes a software fallback.

GC

No blog post on performance would be complete without discussing the garbage collector. Many of the improvements cited thus far have involved reducing allocations, which is in part about reducing direct costs but more so about reducing the load placed on the garbage collector and minimizing the work it needs to do. But improving the GC itself is also a key focus, and one that’s gotten attention in this release, as it has in previous releases.

PR dotnet/coreclr#21523 includes a variety of performance improvements, from improvements to locking to better free list management. PR dotnet/coreclr#23251 from @mjsabby adds support to the GC for Large Pages (“Huge Pages” on Linux), which can be opted-into by very large applications that experience bottlenecks due to the translation lookaside buffer (TLB). And PR dotnet/coreclr#22003 further optimized the write barriers employed by the GC.

One notable piece of work is improving behavior on machines with a large number of processors, e.g. PR dotnet/coreclr#23824. Rather than trying to explain it here, I’ll simply refer to @Maoni0’s blog post on the subject: https://blogs.msdn.microsoft.com/maoni/2019/04/03/making-cpu-configuration-better-for-gc-on-machines-with-64-cpus/.

Similarly, a lot of work has gone into the release to improve the behavior and performance of the GC when operating in a containerized environment (and in particular in one that’s heavily constrained), such as in PR dotnet/coreclr#22180. Again, @Maoni0 can do a much better job than I can describing this work, and you can read all about it her two blog posts, running-with-server-gc-in-a-small-container-scenario-part-0 and running-with-server-gc-in-a-small-container-scenario-part-1-hard-limit-for-the-gc-heap.

JIT

A lot of goodness has gone into the just-in-time (JIT) compiler in .NET Core 3.0.

One of the most impactful changes is tiered compilation (this is split across many PRs, but for example PR dotnet/coreclr#23599). Tiered compilation is a solution for the problem that very good compilation from MSIL to native code takes time; the more analysis to be done, the more optimizations to be applied, the longer it takes. But with a JIT compiler that does that code generation at runtime, that time comes at the direct expense of application start-up, and so you’re left with a trade-off: do you spend more time generating better code but take longer to get going, or do you spend less time generating less-good code but get going faster? Tiered compilation is a scheme for accomplishing both. The idea is that methods are first compiled with a fast pass that applies few-to-no optimizations but that completes very quickly, and then as methods are seen to execute again and again, those methods are re-JIT’d, this time with more time spent on code quality.

Interestingly, though, tiered compilation isn’t just about start-up time. There are optimizations that the re-compilation can take advantage of that weren’t available the first time around. For example, tiered compilation can apply to ready-to-run (R2R) images, a form of precompilation employed by assemblies in the .NET Core shared framework. These assemblies contain precompiled native code, but in some ways the optimizations that can be applied during that native code generation are limited in order to aid in version resiliency, e.g. cross-module inlining doesn’t happen with R2R. So, the R2R code can help enable faster start-up, but then methods found to be used frequently can be re-compiled via tiered compilation, thereby taking advantage of such optimizations the original precompiled code was restricted from using.

Here’s an example of that. First, we can run the following benchmark.

Method Toolchain Mean Error StdDev Ratio
LoadXml netcoreapp2.1 9.576 us 0.1523 us 0.1425 us 1.00
LoadXml netcoreapp3.0 7.414 us 0.0980 us 0.0868 us 0.78

 

Then, we can run it again, but this time with tiered compilation disabled by setting the COMPlus_TieredCompilationenvironment variable to 0.

Method Toolchain Mean Error StdDev Ratio RatioSD
LoadXml netcoreapp2.1 9.650 us 0.1638 us 0.1279 us 1.00 0.00
LoadXml netcoreapp3.0 9.002 us 0.2018 us 0.2073 us 0.93 0.03

 

There are a variety of environment variables that configure tiered compilation and in what situations it’s enabled. For more details, see https://github.com/dotnet/coreclr/issues/24064.

Another really cool improvement in the JIT comes in PR dotnet/coreclr#20886. In previous releases of .NET, the JIT could optimize the usage of some primitive type static readonly fields as if they were constants. For example, if a static readonly int field were initialized to the value 42 by the time some code that used that field was JIT compiled, the JIT compiler would effectively treat that field instead as a const, and do constant folding and all other forms of optimizations that would otherwise apply. In .NET Core 3.0, the JIT can now utilize the type of static readonlyfields to do additional optimizations. For example, if a static readonly field is typed as a base type but is then set to a derived type, the JIT might be able to see the actual type of the object stored in the field, and then when a virtual method is called on it, devirtualize the call and even potentially inline it.

Method Toolchain Mean Error StdDev Median Ratio
AccessStatic netcoreapp2.1 0.5625 ns 0.0147 ns 0.0130 ns 0.5616 ns 1.000
AccessStatic netcoreapp3.0 0.0015 ns 0.0060 ns 0.0062 ns 0.0000 ns 0.003

 

That highlights some improvements that have gone into devirtualization, but there are others, such as in PRs dotnet/coreclr#20447dotnet/coreclr#20292, and dotnet/coreclr#20640 which, when combined with PRs like dotnet/coreclr#20637 from @benaadams, help with APIs like ArrayPool<T>.Shared.

Method Toolchain Mean Error StdDev Ratio
RentReturn netcoreapp2.1 32.92 ns 0.3357 ns 0.2803 ns 1.00
RentReturn netcoreapp3.0 25.74 ns 0.2392 ns 0.1867 ns 0.78

 

Another nice improvement is around zeroing of locals. Even when the initlocals flag isn’t set (as of PR dotnet/corefx#34406, it’s cleared for all assemblies in coreclr and corefx), the JIT still needs to zero out references in locals so that the GC doesn’t see and misinterpret garbage, and that zero’ing can take a measurable amount of time, in particular in methods that do a lot of work with spans. PRs dotnet/coreclr#23498 and dotnet/coreclr#13868 make some nice improvements in this area.

Method Toolchain Mean Error StdDev Ratio
StackZero netcoreapp2.1 8.948 ns 0.2479 ns 0.2546 ns 1.00
StackZero netcoreapp3.0 2.389 ns 0.0740 ns 0.0727 ns 0.27

 

Another example relates to structs. As more and more recognition has come to .NET performance, in particular around allocation, there’s been a significant increase in the use of value types, often wrapping one another. For example, awaiting a ValueTask<T> results in calling GetAwaiter() on that value task, and that returns a ValueTaskAwaiter<T> that wraps the ValueTask<T>. PR dotnet/coreclr#19429 improves the situation by removing unnecessary copies involved in these operations.

Method Toolchain Mean Error StdDev Median Ratio
WrapUnwrap netcoreapp2.1 1.2198 ns 0.0717 ns 0.0599 ns 1.2095 ns 1.000
WrapUnwrap netcoreapp3.0 0.0002 ns 0.0007 ns 0.0006 ns 0.0000 ns 0.000

 

What’s Next?

As I write this post, I count 29 pending performance-focused PRs in the coreclr repo and another 8  in the corefx repo. Some of those are likely to be merged in time for the .NET Core 3.0 release, as will, I’m sure, additional PRs that haven’t even been opened yet. In short, even after all of the improvements detailed in for .NET Core 2.0.NET Core 2.1, and now in this post for .NET Core 3.0, and even with all of those improvements contributing to ASP.NET Core being one of the fastest web servers on the planet, there is still incredible opportunity for performance to keep getting better and better, and for you to help achieve that. Hopefully this post has made you excited about the potential .NET Core 3.0 holds. I look forward to reviewing your PRs as we all contribute to this exciting future together!

Avatar

  Source Link

Share.

Leave A Reply