NFX Wave Benchmarks

The purpose of this project is to realistically assess the performance of our web stack and compare it to the top tech offerings on the market. This work was sparked by a post about Kestrel achieving +1MM requests a second on .Net Core. Indeed, the MS team has implemented some tight optimizations and achieved very good results, however the question stands: how is this applicable to the real web apps laden with business logic? Is this even needed? Does it justify the complexity?. When you need to carry a 1000lb rock, does it matter if your bag weights 1 lb or 10 lb?

NFX Wave is a web-centric part of NFX UNISTACK library. WAVE provides: web server, request handling pipeline (filters/handlers), and MVCHandler that operate on top.

At the present moment we have based the WaveServer implementation on the classic HttpListener class. We really do not depend on HttpListener and plan to get rid of it altogether in favor of libuv or plain sockets.

Testing Scenarios

We have tested two kinds of request handling: direct handler that writes to the stream, and MVC Controller which is of course slower than plain handler as it is an extra layer that dispatches request/verbs into a class method invocation.

url	Type	MVC Action	Description
/plaintext	NFX.Wave.Benchmarks.Handlers.PlainTextHandler	n/a	Plain Text with direct response writing
/json	NFX.Wave.Benchmarks.Handlers.JsonHandler	n/a	Simple JSON with direct response writing
/mvc/plaintext	NFX.Wave.Handlers.MVCHandler	Index.PlainText	Return Plain Text via MVC controller action
/mvc/json	NFX.Wave.Handlers.MVCHandler	Index.Json	Return simple JSON via MVC controller action

Setting Up

1. Clone this repo

2. Build VS solution in Release. This is a regular .NET project, not a Core project

3. Either run as Admin or grant HttpListener rights like so: netsh http add urlacl url=http://+:8080/ user=Everyone

4. Run

Output\Release\nfx-wave-bench

Note: You may need to open port 8080 for external traffic to access the server rig

Details

Our Test Environment

We are testing using the following physical machines:

Name	OS	Role	CPU	RAM	NIC	Notes
CNONIM	Windows 10	Desktop	Core i7-6700K = 4 core @ 4.2 ghz	64 GB	1 Gbit
SEXTOD	Windows 7	Desktop	Core i7-3930K = 6 core @ 3.2 ghz	64 GB	1 Gbit

Load Generation

We generate load with Apache Bench (AB) for now, using -k for keep alives, -nXXXX number of requests, -cXXXXX concurrency level

ab -n200000 -c8 -k ...

Plaintext

Stack	Server	Client	Req/sec
NFX.Wave	CNONIM	AB local	102,320
ASP.Net Core on Kestrel	CNONIM	AB local	138,789

Json

Stack	Server	Client	Req/sec
NFX.Wave	CNONIM	AB local	93,647
ASP.Net Core on Kestrel	CNONIM	AB local	135,224

MVC PlainText

Stack	Server	Client	Req/sec
NFX.Wave	CNONIM	AB local	83,480
ASP.Net Core on Kestrel	CNONIM	AB local	106,297

MVC Json

Stack	Server	Client	Req/sec
NFX.Wave	CNONIM	AB local	75,608
ASP.Net Core on Kestrel	CNONIM	AB local	6,349