Benchmarks for ASP.NET Core
C# C++ Shell Lua PowerShell C Batchfile
Latest commit 2dfdb74 Jan 10, 2017 @halter73 halter73 Parse and bind to IP's configured via --server.urls
- Will revisit when aspnet/KestrelHttpServer#1290
  is addressed

README.md

Benchmarks

Benchmarks for ASP.NET Core including (but not limited to) scenarios from the TechEmpower Web Framework Benchmarks.

Running the benchmarks

The benchmark repo is set up to work against the latest sources (i.e. not packages from nuget.org) for ASP.NET Core so make sure you read through the following details to help you get started.

The ASP.NET Core benchmarks server application itself is in the ./src/Benchmarks folder. The ./experimental folder contains various experimental projects that aren't themselves part of the benchmarks.

The scenarios

Following are the details of each of the scenarios the server application contains implementations for and thus can be benchmarked:

url Name Description
/plaintext Plaintext From https://www.techempower.com/benchmarks/ This test is an exercise of the request-routing fundamentals only, designed to demonstrate the capacity of high-performance platforms in particular. Requests will be sent using HTTP pipelining.
/json JSON From https://www.techempower.com/benchmarks/ This test exercises the framework fundamentals including keep-alive support, request routing, request header parsing, object instantiation, JSON serialization, response header generation, and request count throughput.
/mvc/plaintext MVC Plaintext As for the plaintext test above, but using routing & MVC.
/mvc/json MVC JSON As for the json test above, but using routing & MVC.
/mvc/view MVC Plain View As for the plaintext test above, but using routing & MVC plus rendering the result via a Razor view.
/db/raw Single Query Raw From https://www.techempower.com/benchmarks/ This test exercises the framework's object-relational mapper (ORM), random number generator, database driver, and database connection pool.
/db/ef Single Query EF As for the single query raw test above but using Entity Framework Core as the data-access library/ORM.
/db/dapper Single Query Dapper As for the single query raw test above but using Dapper as the data-access library/ORM.
/queries/raw Multiple Queries Raw From https://www.techempower.com/benchmarks/ Multiple rows are fetched to more dramatically punish the database driver and connection pool.
/queries/ef Single Query EF As for the multiple query raw test above but using Entity Framework Core as the data-access library/ORM.
/queries/dapper Single Query Dapper As for the multiple query raw test above but using Dapper as the data-access library/ORM.
/fortunes/raw Fortunes Raw From https://www.techempower.com/benchmarks/ This test exercises the ORM, database connectivity, dynamic-size collections, sorting, server-side templates, XSS countermeasures, and character encoding.
/fortunes/ef Fortunes EF As for the Fortunes raw test above but using Entity Framework Core as the data-access library/ORM.
/fortunes/dapper Fortunes Dapper As for the Fortunes raw test above but using Dapper as the data-access library/ORM.

Setting up the web server

You can run the benchmarks application server on Windows, OSX or Linux.

  1. Follow the instructions on the ASP.NET Core docs site to get the appropriate runtime and tooling pieces for your chosen platform installed.

  2. Clone this repo to the server.

  3. Navigate to the ./src/Benchmarks directory under this repo and run the following command to restore the latest version of the ASP.NET Core runtime and dependencies for the server application:

    dotnet restore
    
  4. Finally, start the server application with the following command:

    dotnet run --configuration Release
    
  5. If you're generating load from a separate machine (which is recommended), you'll need to change the URL Kestrel binds to as it only binds to localhost by default. You can change it via a command line argument or an environment variable, "ASPNETCORE_URLS" set to "http://*:5000"

Note: You may need to open port 5000 for external traffic in your firewall for the server to successfully run

Generating Load

It's best to generate load from a completely separate machine from the server if you can, to avoid resource contention during the test.

We use the wrk load generation tool to generate the load for our benchmark runs. It's the best tool we've found for the job and supports HTTP pipelining (used by the plaintext scenario) via its scripting interface. Wrk will only run from a Linux machine however, so if you must use Windows, try using ab (Apache Bench). You can dowload ab for Windows from here.

You'll need to clone the wrk repo on your load generation machine and follow their instructions to build it.

Here's a sample wrk command to generate load for the JSON scenario. This run is using 256 connections across 32 client threads for a duration of 10 seconds.

wrk -c 256 -t 32 -d 10 http://10.0.0.100:5000/json

To generate pipelined load for the plaintext scenario, use the following command, assuming your CWD is the root of this repo and wrk is on your path. The final argument after the -- is the desired pipeline depth. We always run the plaintext scenario at a pipeline depth of 16, just like the Techempower Benchmarks.

wrk -c 256 -t 32 -d 10 -s ./scripts/pipeline.lua http://10.0.0.100:5000/plaintext -- 16

Note: You may want to tweak the number of client threads (the -t arg) being used based on the specs of your load generation machine.

Enabling scenarios

The application will prompt for which scenarios you want to enable when you run it in the command line. Follow the instructions to enable the desired scenarios. You can also add a scenarios.json configuration file to configure them that way.

Details of our perf lab

Environment

We're using the following physical machines to perform these tests:

Name OS Role CPU RAM NIC Notes
perfsvr Windows Server 2012 R2 Web Server Xeon E5-1650 32 GB Intel® Ethernet Converged Network Adapter X540-T1 10GbE
perfsvr2 Ubuntu 14.04 LTS Web Server & Load Generator Xeon E5-1650 32 GB Intel® Ethernet Converged Network Adapter X540-T1 10GbE
perf02 Windows Server 2012 R2 Load Generator Xeon W3550 24 GB Intel® Ethernet Converged Network Adapter X540-T1 10GbE
perf03 Ubuntu 14.04 LTS Load Generator Xeon W3550 12 GB Intel® Ethernet Converged Network Adapter X540-T1 10GbE

The machines are connected to an 8-port Netgear XS708E 10-Gigabit switch, that is isolated from the rest of the corporate network (the machines are all multihomed).

Load Generation

We're using wrk to generate load from one of our Linux boxes (usually perf03).

Results

For each stack, variations of the load parameters and multiple runs are tested and the highest result is recorded. Detailed results are tracked in the results spreadsheet.

Experimental Baselines

These are server experiments that are intended to measure the non-HTTP overload of different technology stacks and approaches. These generally aren't real HTTP servers but rather TCP servers that special case replying to any HTTP-looking request with a fixed HTTP response.

Stack Server Req/sec Load Params Impl Observations
Hammer (raw HTTP.SYS) perfsvr ~280,000 32 threads, 512 connections C++ directly on HTTP.SYS CPU is 100%
Hammer (raw HTTP.SYS) perfsvr ~460,000 32 threads, 256 connections, pipelining 16 deep C++ directly on HTTP.SYS CPU is 100%
libuv C# perfsvr 300,507 12 threads, 1024 connections Simple TCP server, load spread across 12 ports (port/thread/CPU) CPU is 54%, mostly in kernel mode
libuv C# perfsvr 2,379,267 36 threads, 288 connections, pipelining 16 deep Simple TCP server, load spread across 12 ports (port/thread/CPU) CPU is 100%, mostly in user mode
RIO C# perfsvr ~5,905,000 32 threads, 512 connections, pipelining 16 deep Simple TCP server using Windows Registered IO (RIO) via P/Invoke from C# CPU is 100%, 95% in user mode

Plain Text

Similar to the plain text benchmark in the TechEmpower tests. Intended to highlight the HTTP efficiency of the server & stack. Implementations are free to cache the response body aggressively and remove/disable components that aren't required in order to maximize performance.

Stack Server Req/sec Load Params Impl Observations
ASP.NET 4.6 perfsvr 57,843 32 threads, 256 connections Generic reusable handler, unused IIS modules removed CPU is 100%, almost exclusively in user mode
IIS Static File (kernel cached) perfsvr 276,727 32 threads, 512 connections hello.html containing "HelloWorld" CPU is 36%, almost exclusively in kernel mode
IIS Static File (non-kernel cached) perfsvr 231,609 32 threads, 512 connections hello.html containing "HelloWorld" CPU is 100%, almost exclusively in user mode
NodeJS perfsvr 106,479 32 threads, 256 connections The actual TechEmpower NodeJS app CPU is 100%, almost exclusively in user mode
NodeJS perfsvr2 (Linux) 127,017 32 threads, 512 connections The actual TechEmpower NodeJS app CPU is 100%, almost exclusively in user mode
ASP.NET Core on Kestrel perfsvr 313,001 32 threads, 256 connections Middleware class, multi IO thread CPU is 100%
Scala - Plain perfsvr 176,509 32 threads, 1024 connections The actual TechEmpower Scala Plain plaintext app CPU is 68%, mostly in kernel mode
Netty perfsvr 447,993 32 threads, 256 connections The actual TechEmpower Netty app CPU is 100%

Plain Text with HTTP Pipelining

Like the Plain Text scenario above but with HTTP pipelining enabled at a depth of 16. Only stacks/servers that show an improvement with pipelining are included.

Stack Server Req/sec Load Params Impl Observations
NodeJS perfsvr 147,554 32 threads, 256 connections The actual TechEmpower NodeJS app CPU is 100%, almost exclusively in user mode
NodeJS perfsvr2 (Linux) 173,641 32 threads, 512 connections The actual TechEmpower NodeJS app CPU is 100%
ASP.NET Core on Kestrel perfsvr 1,174,881 32 threads, 256 connections Middleware class, multi IO thread CPU is 100%
ASP.NET Core on Kestrel perfsvr2 (Linux) 928,023 32 threads, 256 connections Middleware class, single IO thread
Scala perfsvr 1,514,942 32 threads, 1024 connections The actual TechEmpower Scala plaintext app CPU is 100%, 70% in user mode
Netty perfsvr 2,808,515 32 threads, 256 connections The actual TechEmpower Netty app CPU is 100%

This project is part of ASP.NET Core. You can find samples, documentation and getting started instructions for ASP.NET Core at the Home repo.