Slow performance in Kestrel compared to IIS

Question

For a demo I intend to run a benchmark where I compare .NET Core running on Kestrel, with .NET Framework 4.6.1 running on IIS. Both on local machine.

"The Internet" says that Kestrel is much faster, but in my own benchmark there's no notable difference to IIS. And most of the time IIS is even faster. How is that?

I have the "same" code running on .NET Framework (using EF) and .NET Core (using EFCore). The controller looks like this. It will write something to DB, and then fetch it and return it. I close the DBContext to make sure there's no cache. Code for .NET Core below (code for .NET Framework is similiar except no DI).

[Route("api/[controller]")]
public class ValuesController : Controller
{
    private readonly DbContextOptions<DemoContext> options;

    public ValuesController(DbContextOptionsBuilder<DemoContext> builder)
    {
        this.options = builder.Options;
    }

    // GET api/values
    [HttpGet]
    public async Task<IActionResult> Get()
    {
        var id = Guid.NewGuid();
        using (var context = new DemoContext(options))
        {

            var newItem = new DemoTable()
            {
                Id = id,
                Stamp = DateTime.Now,
                System = "Core"
            };
            context.DemoTable.Add(newItem);
            await context.SaveChangesAsync();
        }

        using (var context = new DemoContext(options))
        {
            var item = await context.DemoTable.SingleAsync(x => x.Id == id);
            return Ok(item);

        }
    }
}

My Program.cs looks like this:

public static IWebHost BuildWebHost(string[] args) =>
    WebHost.CreateDefaultBuilder(args)
        .UseKestrel(options =>
        {
            options.Limits.MaxConcurrentConnections = 1000;
            options.Listen(IPAddress.Loopback, 5050);
        })
        .ConfigureLogging((context, logging) =>
        {
            logging.ClearProviders();
        })
        .UseStartup<Startup>()
        .Build();

I run it with dotnet as dotnet demoApp.dll.

I test it with an application which makes a number of requests.

    var cts = new CancellationTokenSource();
    int processed = 0;
    int seconds = 30;

    var url = new Uri($"http://localhost:{port}");
    var tasks = new Task[20];
    for (int i = 0; i < tasks.Length; i++)
    {
        tasks[i] = Task.Factory.StartNew(() =>
        {
            while (!token.IsCancellationRequested)
            {
                var client = new RestClient(url);
                var request = new RestRequest("/api/values");
                client.ExecuteAsyncGet(request, (response, handle) =>
                {
                    if(!token.IsCancellationRequested)
                        Interlocked.Add(ref processed, 1);
                }, "GET");
            }
        });
    }
    Thread.Sleep(seconds * 1000);
    cts.Cancel();

I get more requests when I run it against .NET Framework with IIS than when I run it against Core with Kestrel.

I've tried changing options.Limits in Kestrel with not success. My assumption is that there's something wrong with my "benchmarking application" or that my local machine is the bottleneck itself.

Why is IIS handling the requests much faster than Kestrel?

UPDATE:

If I remove EF and only return OK() in the controller IIS still performs better.

I wake the servers up before benchmarking.

I build it as Release.

When I run with 10 threads for 30 seconds IIS will handle 600 request and Kestrel 300 requests.

gcServer is set to true.

Answer 1

You cant test a web hosts performance by connecting it to a db. The db or your business logic will be the bottleneck .

Start with a clean project or better yet just grab the benchmark. For high end performance many things come into play ..

The fact you are handling 300 or 600 requests ( pitiful) suggest something else in the pipeline.. Note your controller is injecting a db context on each request as well + whatever middleware you have.