In this document, a *hot code path* is defined as a code path that is frequently called and where much of the execution time occurs. Hot code paths typically limit app scale-out and performance and are discussed in several parts of this document.
ASP.NET Core apps should be designed to process many requests simultaneously. Asynchronous APIs allow a small pool of threads to handle thousands of concurrent requests by not waiting on blocking calls. Rather than waiting on a long-running synchronous task to complete, the thread can work on another request.
A common performance problem in ASP.NET Core apps is blocking calls that could be asynchronous. Many synchronous blocking calls lead to [Thread Pool starvation](https://blogs.msdn.microsoft.com/vancem/2018/10/16/diagnosing-net-core-threadpool-starvation-with-perfview-why-my-service-is-not-saturating-all-cores-or-seems-to-stall/) and degraded response times.
* Block asynchronous execution by calling [Task.Wait](/dotnet/api/system.threading.tasks.task.wait) or [Task.Result](/dotnet/api/system.threading.tasks.task-1.result).
* Acquire locks in common code paths. ASP.NET Core apps are most performant when architected to run code in parallel.
* Call [Task.Run](/dotnet/api/system.threading.tasks.task.run) and immediately await it. ASP.NET Core already runs app code on normal Thread Pool threads, so calling Task.Run only results in extra unnecessary Thread Pool scheduling. Even if the scheduled code would block a thread, Task.Run does not prevent that.
* Call data access and long-running operations APIs asynchronously if an asynchronous API is available. Once again, do not use [Task.Run](/dotnet/api/system.threading.tasks.task.run) to make a synchronus API asynchronous.
* Make controller/Razor Page actions asynchronous. The entire call stack is asynchronous in order to benefit from [async/await](/dotnet/csharp/programming-guide/concepts/async/) patterns.
A profiler, such as [PerfView](https://github.com/Microsoft/perfview), can be used to find threads frequently added to the [Thread Pool](/windows/desktop/procthread/thread-pools). The `Microsoft-Windows-DotNETRuntime/ThreadPoolWorkerThread/Start` event indicates a thread added to the thread pool. <!-- For more information, see [async guidance docs](TBD-Link_To_Davifowl_Doc) -->
The [.NET Core garbage collector](/dotnet/standard/garbage-collection/) manages allocation and release of memory automatically in ASP.NET Core apps. Automatic garbage collection generally means that developers don't need to worry about how or when memory is freed. However, cleaning up unreferenced objects takes CPU time, so developers should minimize allocating objects in [hot code paths](#understand-hot-code-paths). Garbage collection is especially expensive on large objects (> 85 K bytes). Large objects are stored on the [large object heap](/dotnet/standard/garbage-collection/large-object-heap) and require a full (generation 2) garbage collection to clean up. Unlike generation 0 and generation 1 collections, a generation 2 collection requires a temporary suspension of app execution. Frequent allocation and de-allocation of large objects can cause inconsistent performance.
Memory issues, such as the preceding, can be diagnosed by reviewing garbage collection (GC) stats in [PerfView](https://github.com/Microsoft/perfview) and examining:
Interactions with a data store and other remote services are often the slowest parts of an ASP.NET Core app. Reading and writing data efficiently is critical for good performance.
* **Do not** retrieve more data than is necessary. Write queries to return just the data that's necessary for the current HTTP request.
* **Do** consider caching frequently accessed data retrieved from a database or remote service if slightly out-of-date data is acceptable. Depending on the scenario, use a [MemoryCache](xref:performance/caching/memory) or a [DistributedCache](xref:performance/caching/distributed). For more information, see <xref:performance/caching/response>.
* **Do** minimize network round trips. The goal is to retrieve the required data in a single call rather than several calls.
* **Do** use [no-tracking queries](/ef/core/querying/tracking#no-tracking-queries) in Entity Framework Core when accessing data for read-only purposes. EF Core can return the results of no-tracking queries more efficiently.
* **Do** filter and aggregate LINQ queries (with `.Where`, `.Select`, or `.Sum` statements, for example) so that the filtering is performed by the database.
* **Do** consider that EF Core resolves some query operators on the client, which may lead to inefficient query execution. For more information, see [Client evaluation performance issues](/ef/core/querying/client-eval#client-evaluation-performance-issues).
* **Do not** use projection queries on collections, which can result in executing "N + 1" SQL queries. For more information, see [Optimization of correlated subqueries](/ef/core/what-is-new/ef-core-2.1#optimization-of-correlated-subqueries).
See [EF High Performance](/ef/core/what-is-new/ef-core-2.0#explicitly-compiled-queries) for approaches that may improve performance in high-scale apps:
We recommend measuring the impact of the preceding high-performance approaches before committing the code base. The additional complexity of compiled queries may not justify the performance improvement.
Query issues can be detected by reviewing the time spent accessing data with [Application Insights](/azure/application-insights/app-insights-overview) or with profiling tools. Most databases also make statistics available concerning frequently executed queries.
Although [HttpClient](/dotnet/api/system.net.http.httpclient) implements the `IDisposable` interface, it's designed for reuse. Closed `HttpClient` instances leave sockets open in the `TIME_WAIT` state for a short period of time. If a code path that creates and disposes of `HttpClient` objects is frequently used, the app may exhaust available sockets. [HttpClientFactory](/dotnet/standard/microservices-architecture/implement-resilient-applications/use-httpclientfactory-to-implement-resilient-http-requests) was introduced in ASP.NET Core 2.1 as a solution to this problem. It handles pooling HTTP connections to optimize performance and reliability.
* **Do not** create and dispose of `HttpClient` instances directly.
* **Do** use [HttpClientFactory](/dotnet/standard/microservices-architecture/implement-resilient-applications/use-httpclientfactory-to-implement-resilient-http-requests) to retrieve `HttpClient` instances. For more information, see [Use HttpClientFactory to implement resilient HTTP requests](/dotnet/standard/microservices-architecture/implement-resilient-applications/use-httpclientfactory-to-implement-resilient-http-requests).
* Middleware components in the app's request processing pipeline, especially middleware run early in the pipeline. These components have a large impact on performance.
* Code that's executed for every request or multiple times per request. For example, custom logging, authorization handlers, or initialization of transient services.
* **Do** use performance profiling tools, such as [Visual Studio Diagnostic Tools](/visualstudio/profiling/profiling-feature-tour) or [PerfView](https://github.com/Microsoft/perfview)), to identify [hot code paths](#understand-hot-code-paths).
## Complete long-running Tasks outside of HTTP requests
Most requests to an ASP.NET Core app can be handled by a controller or page model calling necessary services and returning an HTTP response. For some requests that involve long-running tasks, it's better to make the entire request-response process asynchronous.
Recommendations:
* **Do not** wait for long-running tasks to complete as part of ordinary HTTP request processing.
* **Do** consider handling long-running requests with [background services](xref:fundamentals/host/hosted-services) or out of process with an [Azure Function](/azure/azure-functions/). Completing work out-of-process is especially beneficial for CPU-intensive tasks.
* **Do** use real-time communication options, such as [SignalR](xref:signalr/introduction), to communicate with clients asynchronously.
ASP.NET Core apps with complex front-ends frequently serve many JavaScript, CSS, or image files. Performance of initial load requests can be improved by:
* Bundling, which combines multiple files into one.
Reducing the size of the response usually increases the responsiveness of an app, often dramatically. One way to reduce payload sizes is to compress an app's responses. For more information, see [Response compression](xref:performance/response-compression).
Each new release of ASP.NET Core includes performance improvements. Optimizations in .NET Core and ASP.NET Core mean that newer versions generally outperform older versions. For example, .NET Core 2.1 added support for compiled regular expressions and benefitted from [`Span<T>`](https://msdn.microsoft.com/magazine/mt814808.aspx). ASP.NET Core 2.2 added support for HTTP/2. [ASP.NET Core 3.0 adds many improvements](xref:aspnetcore-3.0) that reduce memory usage and improve throughput. If performance is a priority, consider upgrading to the current version of ASP.NET Core.
Exceptions should be rare. Throwing and catching exceptions is slow relative to other code flow patterns. Because of this, exceptions shouldn't be used to control normal program flow.
The following sections provide performance tips and known reliability problems and solutions.
## Avoid synchronous read or write on HttpRequest/HttpResponse body
All IO in ASP.NET Core is asynchronous. Servers implement the `Stream` interface, which has both synchronous and asynchronous overloads. The asynchronous ones should be preferred to avoid blocking thread pool threads. Blocking threads can lead to thread pool starvation.
**Do not do this:** The following example uses the <xref:System.IO.StreamReader.ReadToEnd*>. It blocks the current thread to wait for the result. This is an example of [sync over async](https://github.com/davidfowl/AspNetCoreDiagnosticScenarios/blob/master/AsyncGuidance.md#warning-sync-over-async
In the preceding code, `Get` synchronously reads the entire HTTP request body into memory. If the client is slowly uploading, the app is doing sync over async. The app does sync over async because Kestrel does **NOT** support synchronous reads.
**Do this:** The following example uses <xref:System.IO.StreamReader.ReadToEndAsync*> and does not block the thread while reading.
The preceding code asynchronously reads the entire HTTP request body into memory.
> [!WARNING]
> If the request is large, reading the entire HTTP request body into memory could lead to an out of memory (OOM) condition. OOM can result in a Denial Of Service. For more information, see [Avoid reading large request bodies or response bodies into memory](#arlb) in this document.
* The cached form value is being read using `HttpContext.Request.Form`
**Do not do this:** The following example uses `HttpContext.Request.Form`. `HttpContext.Request.Form` uses [sync over async](https://github.com/davidfowl/AspNetCoreDiagnosticScenarios/blob/master/AsyncGuidance.md#warning-sync-over-async
## Avoid reading large request bodies or response bodies into memory
In .NET, every object allocation greater than 85 KB ends up in the large object heap ([LOH](https://blogs.msdn.microsoft.com/maoni/2006/04/19/large-object-heap/)). Large objects are expensive in two ways:
* The allocation cost is high because the memory for a newly allocated large object has to be cleared. The CLR guarantees that memory for all newly allocated objects is cleared.
* LOH is collected with the rest of the heap. LOH requires a full [garbage collection](/dotnet/standard/garbage-collection/fundamentals) or [Gen2 collection](/dotnet/standard/garbage-collection/fundamentals#generations).
This [blog post](https://adamsitnik.com/Array-Pool/#the-problem) describes the problem succinctly:
> When a large object is allocated, it’s marked as Gen 2 object. Not Gen 0 as for small objects. The consequences are that if you run out of memory in LOH, GC cleans up the whole managed heap, not only LOH. So it cleans up Gen 0, Gen 1 and Gen 2 including LOH. This is called full garbage collection and is the most time-consuming garbage collection. For many applications, it can be acceptable. But definitely not for high-performance web servers, where few big memory buffers are needed to handle an average web request (read from a socket, decompress, decode JSON & more).
Naively storing a large request or response body into a single `byte[]` or `string`:
* May result in quickly running out of space in the LOH.
* May cause performance issues for the app because of full GCs running.
## Working with a synchronous data processing API
When using a serializer/de-serializer that only supports synchronous reads and writes (for example, [JSON.NET](https://www.newtonsoft.com/json/help/html/Introduction.htm)):
* Buffer the data into memory asynchronously before passing it into the serializer/de-serializer.
> [!WARNING]
> If the request is large, it could lead to an out of memory (OOM) condition. OOM can result in a Denial Of Service. For more information, see [Avoid reading large request bodies or response bodies into memory](#arlb) in this document.
ASP.NET Core 3.0 uses <xref:System.Text.Json> by default for JSON serialization. <xref:System.Text.Json>:
* Reads and writes JSON asynchronously.
* Is optimized for UTF-8 text.
* Typically higher performance than `Newtonsoft.Json`.
## Do not store IHttpContextAccessor.HttpContext in a field
The [IHttpContextAccessor.HttpContext](xref:Microsoft.AspNetCore.Http.IHttpContextAccessor.HttpContext) returns the `HttpContext` of the active request when accessed from the request thread. The `IHttpContextAccessor.HttpContext` should **not** be stored in a field or variable.
**Do not do this:** The following example stores the `HttpContext` in a field, and then attempts to use it later.
## Do not access HttpContext from multiple threads
`HttpContext` is *NOT* thread-safe. Accessing `HttpContext` from multiple threads in parallel can result in undefined behavior such as hangs, crashes, and data corruption.
**Do not do this:** The following example makes three parallel requests and logs the incoming request path before and after the outgoing HTTP request. The request path is accessed from multiple threads, potentially in parallel.
## Do not use the HttpContext after the request is complete
`HttpContext` is only valid as long as there is an active HTTP request in the ASP.NET Core pipeline. The entire ASP.NET Core pipeline is an asynchronous chain of delegates that executes every request. When the `Task` returned from this chain completes, the `HttpContext` is recycled.
## Do not capture the HttpContext in background threads
**Do not do this:** The following example shows a closure is capturing the `HttpContext` from the `Controller` property. This is a bad practice because the work item could:
Background tasks should be implemented as hosted services. For more information, see [Background tasks with hosted services](xref:fundamentals/host/hosted-services).
**Do not do this:** The following example shows a closure is capturing the `DbContext` from the `Controller` action parameter. This is a bad practice. The work item could run outside of the request scope. The `ContosoDbContext` is scoped to the request, resulting in an `ObjectDisposedException`.
* Injects an <xref:Microsoft.Extensions.DependencyInjection.IServiceScopeFactory> in order to create a scope in the background work item. `IServiceScopeFactory` is a singleton.
* Creates a new dependency injection scope in the background thread.
**Do this:** The following example uses `HttpResponse.OnStarting` to set the headers before the response headers are flushed to the client.
Checking if the response has not started allows registering a callback that will be invoked just before response headers are written. Checking if the response has not started:
* Provides the ability to append or override headers just in time.
* Doesn't require knowledge of the next middleware in the pipeline.
