Networking events in .NET

Artikkeli
11/30/2023

Events give you access to:

Accurate timestamps throughout the lifetime of an operation. For example, how long it took to connect to the server and how long it took an HTTP request to receive response headers.
Debug/trace information that may not be obtainable by other means. For example, what kind of decisions the connection pool made and why.

The instrumentation is based on EventSource, allowing you to collect this information from both inside and outside the process.

Event providers

Networking information is split across the following event providers:

System.Net.Http (HttpClient and SocketsHttpHandler)
System.Net.NameResolution (Dns)
System.Net.Security (SslStream)
System.Net.Sockets
Microsoft.AspNetCore.Hosting
Microsoft-AspNetCore-Server-Kestrel

The telemetry has some performance overhead when enabled, so make sure to subscribe only to event providers you're actually interested in.

Consume events in-process

Prefer in-process collection when possible for easier event correlation and analysis.

EventListener

EventListener is an API that allows you to listen to EventSource events from within the same process that produced them.

using System.Diagnostics.Tracing;

using var listener = new MyListener();

using var client = new HttpClient();
await client.GetStringAsync("https://httpbin.org/get");

public sealed class MyListener : EventListener
{
    protected override void OnEventSourceCreated(EventSource eventSource)
    {
        if (eventSource.Name == "System.Net.Http")
        {
            EnableEvents(eventSource, EventLevel.Informational);
        }
    }

    protected override void OnEventWritten(EventWrittenEventArgs eventData)
    {
        Console.WriteLine($"{DateTime.UtcNow:ss:fff} {eventData.EventName}: " +
            string.Join(' ', eventData.PayloadNames!.Zip(eventData.Payload!).Select(pair => $"{pair.First}={pair.Second}")));
    }
}

The preceding code prints output similar to the following:

00:598 RequestStart: scheme=https host=httpbin.org port=443 pathAndQuery=/get versionMajor=1 versionMinor=1 versionPolicy=0
01:470 ConnectionEstablished: versionMajor=1 versionMinor=1
01:474 RequestLeftQueue: timeOnQueueMilliseconds=470,6214 versionMajor=1 versionMinor=1
01:476 RequestHeadersStart:
01:477 RequestHeadersStop:
01:592 ResponseHeadersStart:
01:593 ResponseHeadersStop:
01:633 ResponseContentStart:
01:635 ResponseContentStop:
01:635 RequestStop:
01:637 ConnectionClosed: versionMajor=1 versionMinor=1

Yarp.Telemetry.Consumption

While the EventListener approach outlined above is useful for quick experimentation and debugging, the APIs aren't strongly typed and force you to depend on implementation details of the instrumented library.

To address this, .NET created a library that makes it easy to consume networking events in-process: Yarp.Telemetry.Consumption. While the package is currently maintained as part of the YARP project, it can be used in any .NET application.

To use it, implement the interfaces and methods (events) that you're interested in:

public sealed class MyTelemetryConsumer : IHttpTelemetryConsumer, INetSecurityTelemetryConsumer
{
    public void OnRequestStart(DateTime timestamp, string scheme, string host, int port, string pathAndQuery, int versionMajor, int versionMinor, HttpVersionPolicy versionPolicy)
    {
        Console.WriteLine($"Request to {host} started at {timestamp}");
    }

    public void OnHandshakeStart(DateTime timestamp, bool isServer, string targetHost)
    {
        Console.WriteLine($"Starting TLS handshake with {targetHost}");
    }
}

And register the implementations with your DI container:

services.AddTelemetryConsumer<MyTelemetryConsumer>();

The library provides the following strongly typed interfaces:

These callbacks are called as part of the instrumented operation, so the general logging guidance applies. You should avoid blocking or performing any expensive calculations as part of the callback. Offload any post-processing work to different threads to avoid adding latency to the underlying operation.

Consume events from outside the process

dotnet-trace

dotnet-trace is a cross-platform tool that enables the collection of .NET Core traces of a running process without a native profiler.

dotnet tool install --global dotnet-trace

dotnet-trace collect --providers System.Net.Http,System.Net.Security,System.Threading.Tasks.TplEventSource:0x80:4 --process-id 1234

For all the available commands and parameters, see the dotnet-trace docs.

You can analyze the captured .nettrace file in Visual Studio or PerfView. For more information, see the dotnet-trace analysis docs.

PerfView

PerfView is a free, advanced performance analysis tool. It runs on Windows but can also analyze traces captured on Linux.

To configure the list of events to capture, specify them under Advanced Options / Additional Providers:

*System.Net.Sockets,*System.Net.NameResolution,*System.Net.Http,*System.Net.Security

TraceEvent

TraceEvent is a library that allows you to consume events from different processes in real time. dotnet-trace and PerfView both rely on it.

If you want to process events programmatically and in real time, see the TraceEvent docs.

Start and Stop events

Larger operations often start with a Start event and end with a Stop event. For example, you'll see RequestStart/RequestStop events from System.Net.Http or ConnectStart/ConnectStop events from System.Net.Sockets.

While large operations such as these often guarantee that a Stop event will always be present, this is not always the case. For example, seeing the RequestHeadersStart event from System.Net.Http does not guarantee that RequestHeadersStop will also be logged.

Event correlation

Now that you know how to observe these events, you often need to correlate them together, generally to the originating HTTP request.

Prefer in-process collection when possible for easier event correlation and analysis.

In-process correlation

As networking uses asynchronous I/O, you can't assume that the thread that completed a given request is also the thread that started it. This means you can't use [ThreadLocal] statics to correlate events, but you can use an AsyncLocal. Get familiar with AsyncLocal as this type is key to correlating work across different threads.

AsyncLocal allows you to access the same state deeper into the async flow of an operation. AsyncLocal values only flow down (deeper into the async call stack), and don't propagate changes up to the caller.

Consider the following example:

AsyncLocal<int> asyncLocal = new();
asyncLocal.Value = 1;

await WorkAsync();
Console.WriteLine($"Value after WorkAsync: {asyncLocal.Value}");

async Task WorkAsync()
{
    Console.WriteLine($"Value in WorkAsync: {asyncLocal.Value}");
    asyncLocal.Value = 2;
    Console.WriteLine($"Value in WorkAsync: {asyncLocal.Value}");
}

This code produces the following output:

Value in WorkAsync: 1
Value in WorkAsync: 2
Value after WorkAsync: 1

The value 1 flowed down from the caller into WorkAsync, but the modification in WorkAsync (2) did not flow back up to the caller.

As telemetry events (and their corresponding callbacks) occur as part of the underlying operation, they happen within the same async scope as the caller that initiated the request. This means that you can observe asyncLocal.Value from within the callback, but if you set the value in the callback, nothing will be able to observe it further up the stack.

The following steps show the general pattern.

Create a mutable class that can be updated from inside event callbacks.

public sealed class RequestInfo
{
    public DateTime StartTime, HeadersSent;
}

Set the AsyncLocal.Value before the main operation so that the state will flow into the operation.

private static readonly AsyncLocal<RequestInfo> _requestInfo = new();

public async Task SendRequestAsync(string url)
{
    var info = new RequestInfo();
    _requestInfo.Value = info;

    info.StartTime = DateTime.UtcNow;
    await _client.GetStringAsync(url);

Inside the event callbacks, check if the shared state is available and update it. AsyncLocal.Value will be null if the request was sent by a component that didn't set the AsyncLocal.Value in the first place.
```
public void OnRequestHeadersStop(DateTime timestamp)
{
    if (_requestInfo.Value is { } info) info.HeadersSent = timestamp;
}
```

Process the collected information after finishing the operation.

await _client.GetStringAsync(url);

Log($"Time until headers were sent {url} was {info.HeadersSent - info.StartTime}");

For more ways to do this, see the samples section.

Correlation outside the process

Every event has a piece of data attached to it called ActivityID. This ID encodes the async hierarchy at the time the event was produced.

If you look at a trace file in PerfView, you'll see events like:

System.Net.Http/Request/Start           ActivityID="/#14396/1/1/"
System.Net.Http/RequestHeaders/Start    ActivityID="/#14396/1/1/2/"
System.Net.Http/RequestHeaders/Stop     ActivityID="/#14396/1/1/2/"
System.Net.Http/ResponseHeaders/Start   ActivityID="/#14396/1/1/3/"
System.Net.Http/ResponseHeaders/Stop    ActivityID="/#14396/1/1/3/"
System.Net.Http/Request/Stop            ActivityID="/#14396/1/1/"

You'll know that these events belong to the same request because they share the /#14396/1/1/ prefix.

When doing manual investigations, a useful trick is to look for the System.Net.Http/Request/Start event of a request you're interested in, then paste its ActivityID in the Text Filter text box. If you now select all available providers, you'll see the list of all events that were produced as part of that request.

PerfView automatically collects the ActivityID, but if you use tools like dotnet-trace, you must explicitly enable the System.Threading.Tasks.TplEventSource:0x80:4 provider (see dotnet-trace example above).

HttpClient request vs. connection lifetime

Since .NET 6, an HTTP request is no longer tied to a specific connection. Instead, the request will be serviced as soon as any connection is available.

This means you may see the following order of events:

Request start
Dns start
Request stop
Dns stop

This indicates that the request did trigger a DNS resolution, but was processed by a different connection before the DNS call completed. The same goes for socket connects or TLS handshakes - the originating request may complete before they do.

You should think about such events separately. Monitor DNS resolutions or TLS handshakes on their own without tying them to the timeline of a specific request.

Internal diagnostics

Some components in .NET are instrumented with additional debug-level events that provide more insight into exactly what's happening internally. These events come with high performance overhead and their shape is constantly changing. As the name suggests, they are not part of the public API and you should therefore not rely on their behavior or existence.

Regardless, these events can offer a lot of insights when all else fails. The System.Net stack emits such events from Private.InternalDiagnostics.System.Net.* namespaces.

If you change the condition in the EventListener example above to eventSource.Name.Contains("System.Net"), you will see 100+ events from different layers in the stack. For more information, see the full example.

Samples

Measure DNS resolutions for a given endpoint
Measure time-to-headers when using HttpClient
Time to process a request in ASP.NET Core running Kestrel
Measure the latency of a .NET reverse proxy

Measure DNS resolutions for a given endpoint

services.AddTelemetryConsumer(new DnsMonitor("httpbin.org"));

public sealed class DnsMonitor : INameResolutionTelemetryConsumer
{
    private static readonly AsyncLocal<DateTime?> _startTimestamp = new();
    private readonly string _hostname;

    public DnsMonitor(string hostname) => _hostname = hostname;

    public void OnResolutionStart(DateTime timestamp, string hostNameOrAddress)
    {
        if (hostNameOrAddress.Equals(_hostname, StringComparison.OrdinalIgnoreCase))
        {
            _startTimestamp.Value = timestamp;
        }
    }

    public void OnResolutionStop(DateTime timestamp)
    {
        if (_startTimestamp.Value is { } start)
        {
            Console.WriteLine($"DNS resolution for {_hostname} took {(timestamp - start).TotalMilliseconds} ms");
        }
    }
}

Measure time-to-headers when using HttpClient

var info = RequestState.Current; // Initialize the AsyncLocal's value ahead of time

var response = await client.GetStringAsync("http://httpbin.org/get");

var requestTime = (info.RequestStop - info.RequestStart).TotalMilliseconds;
var serverLatency = (info.HeadersReceived - info.HeadersSent).TotalMilliseconds;
Console.WriteLine($"Request took {requestTime:N2} ms, server request latency was {serverLatency:N2} ms");

public sealed class RequestState
{
    private static readonly AsyncLocal<RequestState> _asyncLocal = new();
    public static RequestState Current => _asyncLocal.Value ??= new();

    public DateTime RequestStart;
    public DateTime HeadersSent;
    public DateTime HeadersReceived;
    public DateTime RequestStop;
}

public sealed class TelemetryConsumer : IHttpTelemetryConsumer
{
    public void OnRequestStart(DateTime timestamp, string scheme, string host, int port, string pathAndQuery, int versionMajor, int versionMinor, HttpVersionPolicy versionPolicy) =>
        RequestState.Current.RequestStart = timestamp;

    public void OnRequestStop(DateTime timestamp) =>
        RequestState.Current.RequestStop = timestamp;

    public void OnRequestHeadersStop(DateTime timestamp) =>
        RequestState.Current.HeadersSent = timestamp;

    public void OnResponseHeadersStop(DateTime timestamp) =>
        RequestState.Current.HeadersReceived = timestamp;
}

Time to process a request in ASP.NET Core running Kestrel

This is currently the most accurate way to measure the duration of a given request.

public sealed class KestrelTelemetryConsumer : IKestrelTelemetryConsumer
{
    private static readonly AsyncLocal<DateTime?> _startTimestamp = new();
    private readonly ILogger<KestrelTelemetryConsumer> _logger;

    public KestrelTelemetryConsumer(ILogger<KestrelTelemetryConsumer> logger) => _logger = logger;

    public void OnRequestStart(DateTime timestamp, string connectionId, string requestId, string httpVersion, string path, string method)
    {
        _startTimestamp.Value = timestamp;
    }

    public void OnRequestStop(DateTime timestamp, string connectionId, string requestId, string httpVersion, string path, string method)
    {
        if (_startTimestamp.Value is { } startTime)
        {
            var elapsed = timestamp - startTime;
            _logger.LogInformation("Request {requestId} to {path} took {elapsedMs} ms", requestId, path, elapsed.TotalMilliseconds);
        }
    }
}

Measure the latency of a .NET reverse proxy

This sample is applicable if you have a reverse proxy that receives inbound requests via Kestrel and makes outbound requests via HttpClient (for example, YARP).

This sample measures the time from receiving the request headers until they're sent out to the backend server.

public sealed class InternalLatencyMonitor : IKestrelTelemetryConsumer, IHttpTelemetryConsumer
{
    private record RequestInfo(DateTime StartTimestamp, string RequestId, string Path);

    private static readonly AsyncLocal<RequestInfo> _requestInfo = new();
    private readonly ILogger<InternalLatencyMonitor> _logger;

    public InternalLatencyMonitor(ILogger<InternalLatencyMonitor> logger) => _logger = logger;

    public void OnRequestStart(DateTime timestamp, string connectionId, string requestId, string httpVersion, string path, string method)
    {
        _requestInfo.Value = new RequestInfo(timestamp, requestId, path);
    }

    public void OnRequestHeadersStop(DateTime timestamp)
    {
        if (_requestInfo.Value is { } requestInfo)
        {
            var elapsed = (timestamp - requestInfo.StartTimestamp).TotalMilliseconds;
            _logger.LogInformation("Internal latency for {requestId} to {path} was {duration} ms", requestInfo.RequestId, requestInfo.Path, elapsed);
        }
    }
}

Need more telemetry?

If you have suggestions for other useful information that could be exposed via events or metrics, create a dotnet/runtime issue.

If you're using the Yarp.Telemetry.Consumption library and have any suggestions, create a microsoft/reverse-proxy issue.

Jaa