Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

288 lines (208 sloc) 10.42 kb

Overview

This directory contains examples of running WebApplications with MagLev using various configurations of HTTPd and protocols to talk between MagLev and the front-end HTTP server.

Note on C-Extensions

Currently, MagLev does not support C-Extensions, hence does not support many popular web components like Thin, Mongrel, Unicorn, eventmachine based servers etc. We are working on an implementation of C-extensions, and when that implementation is done, we will update this investigation.

Until then, you can at least play with getting a MagLev solution working using the currently supported options.

Quick Start

To run the demo “magtag” application on one instance of MagLev with WEBrick, first commit the code to the MagLev repository, and then run the app:

$ rake magtag:commit
$ rake magtag:maglev

Then hit http://localhost:3333/app/

Performance Overview

We ran several combinations of front-end HTTP server against MagLev using HTTP, FastCGI and SCGI as the connection ptorocol. The tests were run with “rake client:ab”, which runs the following apache bench command:

ab -n 5000 -c 10 http://127.0.0.1:3333/app/magtag.css

This just requests a static file 5,000 times. The HTTP servers were configured to pass all requests (even static file requests) to the back end Ruby VM. So this measures round trips with minimal data and no dynamic calculation.

I got the following performance on my “laptop class” “server”.

|-----------------------------+-------|
| Configuration               | Req/s |
|-----------------------------+-------|
| 1 MagLev + WEBrick          |   441 |
| 1 MRI 1.8.7 + WEBrick       |   420 |
| 1 MRI 1.9.2 + WEBrick       |   359 |
|-----------------------------+-------|
|-----------------------------+-------|
| Lighttpd + 1 SCGI           |   843 |
| Lighttpd + 4 SCGI           |  1334 |
|-----------------------------+-------|
| Lighttpd + 1 FastCGI        |   792 |
| Lighttpd + 4 FastCGI        |  1282 |
|-----------------------------+-------|
| Lighttpd + 1 HTTP (WEBrick) |   369 |
| Lighttpd + 4 HTTP (WEBrick) |   661 |
|-----------------------------+-------|
|-----------------------------+-------|
| nginx + 1 HTTP (WEBrick)    |   387 |
| nginx + 4 HTTP (WEBrick)    |   676 |
|-----------------------------+-------|

These performance results shouldn't be taken too seriously. The real test will be to run a web app for many hours or days at continuous high load to see how garbage collection, DB transactions, memory leaks (if any) etc. all play into things. Then, we will also have to look for cyclic patterns in the output (e.g., perhaps we average X req/sec, but there is a clear cycle going from 0 upto 10X and back to 0 and std dev is really large).

I also didn't fight configuring nginx for FastCGI/SCGI, as they are somewhat deprecated, and I expect Unix domain sockets + Mongrel HTTP front end to be much faster when C-Extensions are done.

It's curious to note that 1.9.2 was the slowest running WEBrick…

Rakefile overview

The examples in this directory (except those based on WEBrick) require starting an external HTTPd, and then connecting one or more appropriately configured Ruby VMs to the httpd. So, in general, you'll need to have two terminals open and issue two rake commands. E.g.,

# Terminal 1: start 1 httpd server, configured to listen to 4 SCGI
# servers.
$ rake lighttpd:scgi[4]

# Terminal 2: start 4 MagLev SCGI servers
$ rake maglev:scgi[4]

And then you can hit the base url: localhost:3333/app/

The recipe is:

  1. Figure out which httpd you want (e.g., lighttpd)

  2. Figure out which httpd <-> ruby protocol you want (e.g., scgi)

  3. Figure out how many ruby VMs you want running (e.g., 4)

Then issue the following rake tasks:

  1. Start httpd using protocol and number: rake lighttpd:scgi[4]

  2. Start number of Ruby VMs on protocol: rake maglev:scgi[4]

By default, all HTTP servers are started on port 3333, and all Ruby VMs are started on ports 3000 - 3003. These facts are used in the rake tasks that start and kill processes.

Some targets:

  1. rake kill:all To kill an entire experiment (HTTPd and VMs)

  2. rake kill:vms To kill just the Ruby VMs on ports 3000-3003

  3. rake kill:httpd To kill just the HTTPd on port 3333

Discussion

Summary:

| HTTP Server | HTTP  | FastCGI | SCGI  |
|-------------+-------+---------+-------|
| nginx       | Works |   ?     |   ?   |
| lighttpd    | Works | Works   | works |
| WEBrick     | Works | N/A     | N/A   |

This investigation considers the HTTP server, and the connection protocol between the HTTP server and the MagLev VM(s) running Ruby. We assume that all of the Web App frameworks of interest integrate with Rack, so we limit our survey to Rack adapters. The web stack looks like:

  • HTTP servers (handle static files, and routing to WebApps)

  • optional load balancer (haproxy)
  • Protocol to Web App (CGI, FastCGI, SCGI, HTTP, etc.)

  • Rack (handles other end of protocol)

  • MagLev Web App and Framework

HTTP Server Options

We looked at the following HTTP servers:

  • WEBrick (works HTTP)

  • nginx (works HTTP)

  • lighttpd (works HTTP, SCGI, FastCGI)

We have not yet investigated the following HTTP options (many of them depend on C-Extensions):

  • Mongrel's HTTP parser

  • Thin

  • Unicorn

  • Apache

Connection Protocols

We first look at protocols that MagLev can support, and then we look at how to configure and hook-in various front-end HTTP servers.

  • HTTP

  • SCGI

  • FastCGI

  • CGI (not tested)

  • LSAPI (not tested)

We have solutions for HTTP, SCGI and FastCGI. CGI should also work, but wasn't tested.

HTTP

Most, if not all, HTTP servers that are in use as front-ends to a web site will talk HTTP to back-end application servers.

WEBrick is the standard, pure ruby HTTP solution, and the only HTTP solution currently supported by MagLev. WEBrick is often used in development, but rarely in production. The quick survey of performance reported above shows that WEBrick just can't keep up with the faster, simpler protocols.

Thin, Unicorn and Mongrel all use the Ragel HTTP parser (written in C). Ditching WEBrick for the Ragel front-end should give MagLev a good performance boost. We will expand on this section once C-Extensions are implemented.

SCGI

SCGI is an alternative to FastCGI or CGI. See the SCGI WikiPedia article for details.

The scgi gem is a pure ruby implementation of an SCGI Rack Handler, and works with MagLev.

Install and Configuration
  1. Install the scgi gem (one time):

    $ maglev-gem install scgi
  2. You'll need to configure your HTTP server to connect to the MagLev VM(s) running SCGI. Each HTTP server has its own way of configuring. For an example of a lighttpd configuration file, see config/lighttpd-scgi-4.conf.

  3. Start your HTTP server using the config to listen to four servers:

    $ rake lighttpd:scgi[4]
  4. Start your MagLev VMs: you'll need to do this in a different terminal than the one your HTTP server is running in:

    $ rake maglev:scgi[4]
  5. Connect to the web app: localhost:3333/app/login

  6. Run apache bench against it:

    $ rake client:ab
Discussion

On the plus side, SCGI is easy to setup and is reasonably snappy.

Some of the down-sides of SCGI:

  • The SCGI gem does not support Unix domain sockets.

  • TIME_WAIT problem: the SCGI protocol indicates the end of a response by closing the connection (Rack::Handler::SCGI closes the connection from the HTTP Server). The connection is TCP, so this can leave a lot of sockets in TIME_WAIT, and you may run out of sockets for a while if you have a lot of requests.

  • Concurrent Connection limits: The SCGI gem creates a TCPServer socket on a particular port (ports 300-3004, in the example config files). The HTTP server connects to that port when it has an HTTP request for Ruby to handle. The default configuration for the listen backlog on the TCPServer socket is rather low, and the system does not degrade nicely if that limit is exceeded.

    TODO: Need to increase the listen backlog and try again.

FastCGI

FastCGI is an old protocol that isn't much used for serious web sites. The performance summary suggests that SCGI is slightly faster.

Install and Configuration

The ruby-fcgi gem supports both a pure ruby and a C-extension based FastCGI engine. Unfortunately, a simple “gem install ruby-fcgi” doesn't work (can't compile the C-extension) and there appears to be no way to only install the pure ruby version. I've taken the one file we need (plus the accompanying license file) and placed those in the lib directory. The FastCGI examples just need to “require 'fcgi'” to work.

Discussion

FastCGI is faster than WEBrick, but is a bit of a pain to configure, and even then, you need to fiddle with the HTTP headers before Sinatra sees the request. Lighttpd was far easier to configure than nginx, and as I don't believe FastCGI will be in the running once MagLev supports C-Extensions, I didn't bother getting FastCGI + nginx working.

Not (yet) Considered

Mongrel (evented, swiftiplied)

We did not consider Mongrel, as it is based on a Ruby C-extension to implement the HTTP parser.

The idea behind Mongrel is to use the standard HTTP protocol between the Ruby process and the front end web server. This makes it easy to plug ruby VMs into many different web serving environments. To make things fast, Mongrel implements the HTTP parser as a Ruby C-extension, and then calls into the Ruby VM once the request is “cooked”. Mongrel creates a thread per request, so there is the possibility for multi-threaded web processing, depending on whether the web framework grabs a lock or not.

Thin: Mongrel HTTP Parser + EventMachine + Rack

Thin

Thin uses the Mongrel front-end HTTP parser, and then uses EventMachine to process the requests. Both the Mongrel HTTP parser and EventMachine are based on Ruby C-extensions, so we do not consider them.

Passenger (mod_rack for Apache/nginx)

Passenger

Passenger uses a C-extension.

Unicorn

Unicorn

Jump to Line
Something went wrong with that request. Please try again.