The Ruby 1.9.x Web Servers Booklet

The Ruby 1.9.

x Web Servers Booklet

Ruby 1.9.x Web

Servers Booklet
A SURVEY OF ARCHITECTURAL CHOICES
& ANALYSIS OF RELATIVE PERFORMANCE
MUHAMMAD A. ALI
<oldmoe>
http://oldmoe.blogspot.com
http://github.com/oldmoe
[email protected]
A Survey of Architectural Choices & Analysis of Relative Performance 1/60

The Ruby 1.9.x Web Servers Booklet
Abstract
This survey looks at different Ruby web servers and attempts to rationalize their architectural
choices. It tries to build a background of best practices in building web servers and compares
those to what is found in the Ruby servers being surveyed. Finally benchmarking results are
presented and compared to initial expectations.

About the Author
Muhammad A. Ali, A.K.A oldmoe, is a software developer from Alexandria,

Egypt. He has been writing software professionally for more than nine years
now. He authored the NeverBlock, Reactor and MySQLPlus Ruby libraries.
He loves to code in Ruby and JavaScript. While not coding he enjoys looking
at code written by the likes of _Why the Lucky Stiff and Marc-André
Cournoyer.

Disclaimer
This survey is rather limited in its scope as it focuses solely on the Ruby 1.9.x series. It does
not attempt to cover Ruby 1.8.x, Ruby EE or JRuby. The version used for the tests presented
in this document is 1.9.1.
This survey does not cover all web servers written in Ruby, due to time and resource
constraints I limited the tests to the servers I use myself (Thin & Passenger) plus two of the
defacto Ruby servers (Webrick & Mongrel). No other servers were tested.
This survey only looks at the performance and scalability (as in handling concurrent load)
aspects of the presented web servers. It doest not attempt to cover topics like security,
configurability and administrability. Thus any conclusion must be looked upon as limited in
scope.

Table of Contents
Abstract...........................................................................................................................................2
About the Author............................................................................................................................3
Disclaimer.......................................................................................................................................4
1. Introduction................................................................................................................................6
2. Web Server Primer......................................................................................................................7
3. Concurrency Models...................................................................................................................8
3.1. Introduction.....................................................................................................................................................8
3.2. Multi Processing..............................................................................................................................................8
3.3. Multi Threading.............................................................................................................................................10
3.4. Asynchronous Event Loop.............................................................................................................................14
3.5. Combinations.................................................................................................................................................16
4. Performance Bottlenecks..........................................................................................................17
4.1. Introduction....................................................................................................................................................17
4.2. Data Copying..................................................................................................................................................17
4.3. Context Switching...........................................................................................................................................17
4.4. Lock Contention.............................................................................................................................................18
4.5. Memory Management....................................................................................................................................18
4.6. Blocking Operations.......................................................................................................................................18
4.7. HTTP Parsing.................................................................................................................................................18
4.8. The TCP Stack................................................................................................................................................19
5. Ruby Web Servers.....................................................................................................................20
5.1. Introduction...................................................................................................................................................20
5.2. WEBrick..........................................................................................................................................................21
5.3. Mongrel..........................................................................................................................................................24
5.4. Thin................................................................................................................................................................27
5.5. Passenger.......................................................................................................................................................30
6. Benchmarks..............................................................................................................................33
6.1. Introduction...................................................................................................................................................33
6.2. Benchmarking Methodology.........................................................................................................................34
6.3. Test Setup.......................................................................................................................................................36
6.4. Results............................................................................................................................................................37
6.4.1. Serving Static Files.................................................................................................................................37
6.4.2. Serving Dynamic Requests (Single Rails Process)...............................................................................44
6.4.3. Serving Dynamic Requests From A Proxied Cluster (2 Processes)......................................................49
7. Benchmark Analysis..................................................................................................................54
7.1. Static File Serving Performance.....................................................................................................................54
7.1.1. Small File Sizes.......................................................................................................................................54
7.1.2.Large File Sizes........................................................................................................................................54
7.2. Dynamic Request Serving Performance........................................................................................................58
7.2.1. Quick and Small Responses...................................................................................................................58
7.2.1. Heavy Processing but Small Responses.................................................................................................58
7.2.1. Little Processing but Large Responses..................................................................................................58
8. Conclusion.................................................................................................................................59
References.....................................................................................................................................60

1. Introduction
Ruby was never famous for its capability as a language for writing high performance web
servers. The only major implementation for quite some time was Webrick. A sweet little server
written entirely in Ruby and not known to break any speed records. In 2006 though a new
contender arrived to the scene in the form of Mongrel. A server written by the now famous
Zed A. Shaw. Mongrel totally changed the web serving scene in Ruby and raised the bar
tremendously. It brought several innovations, most notably the Ragel based HTTP parser
which was small, quick and very robust. Mongrel inspired others and more servers were
written for Ruby (and in Ruby).
In this survey we will analyze the various offerings in the Ruby web server arena. We will
attempt to understand the different architectures and in then we will see how the servers
compare when faced with a set of various workloads. We will do the analysis and the
benchmarking against both Ruby 1.8.x series and Ruby 1.9.x series (whenever possible).
Although this survey focuses on Ruby web servers it should be of value to anyone researching
web server architectures in general. The survey also discusses alternative choices that are not
implemented in any of the servers being surveyed.
It should be noted (again) that this is not a comprehensive comparison between the surveyed
servers. It focuses on performance and scalability but there are many other considerations like
security, configurability and many others. It should be clear that the presented results are of
limited scope and hence any conclusion derived from them are limited to that scope as well.

2. Web Server Primer

At its core any web server is simply a never ending loop that attempts to accept connections
on a listening socket. Here is a very simple TCP server
require 'socket'
server = TCPServer.new("0.0.0.0", 8080)
loop do
connection = server.accept
inputline = connection.gets
...
connection.puts outputline
connection.close
end
Code Listing 1
The servers differ in how they construct this loop and how they process incoming connections.
The above sample is for a blocking server. Which means that it can only process one request at
a time and that other requests will be waiting for the current one to finish. A long running
request might make the server unreachable for a while. A group of those will quickly render
the server unusable. There are several strategies to overcome this shortcoming. We will
discuss those strategies and look at how they are utilized by the different servers
For a server to be called a web (HTTP) server it must speak the HTTP protocol. Hence it needs
a way to parse the incoming HTTP requests. Each of the servers presented here attempts to
solve this problem in its own way. But we will soon find that most of them rely on some clone
of Mongrel's parser. If we modify our first server to include HTTP support it could like this:
require 'socket'
loop do
request = HTTP.parse(connection.gets)# an imaginary HTTP parser
...
connection.puts status
connection.puts headers
connection.puts body
connection.close
end
Code Listing 2

3. Concurrency Models
3.1. Introduction
Since web servers need to be able to handle multiple clients at once then they must employ
some form of concurrency. Different concurrency models bring with them different concerns
and implications. We will look at various options available to Ruby web servers. The list is not
inclusive as there are other models that are more prominent outside the Ruby sphere.
3.2. Multi Processing
In its simplest form you build a server that performs all its I/O in a blocking manner. Then
you instantiate it several times. Each instance can handle one connection at a time. Hence you
can handle as many connections as there are processes.
There are usually two approaches here. One is by starting independent processes (like a
mongrel cluster, each process listens on its own port) or by dynamically forking processes
from each other. The latter enjoys the benefit of being able to adapt the number of processes
to load while the former is almost always fixed in size. In its extreme we use a process per
connection (like in CGI mode) but the common use is a pool (referred to as cluster sometimes)
of processes which are load balanced by some proxy server.
This model has its roots in the UNIX traditions of extensive use of multiprocessing for
concurrency. Many UNIX web servers where built around fork a server per connection model.
While easy, this model has scalability challenges as the cost of forking many processes
prohibits it from being able to handle a relatively large number of concurrent connections.
Here is a sample forking server written in Ruby:

require 'socket'
loop do
if fork
request = HTTP.parse(connection.gets)
...
connection.close
end
end
Code Listing 3
On the other hand using a set of pre-forked (or pre-instantiated) servers avoids the overhead
of forking servers per request but it limits the ability of the server to handle concurrent
connections to the count of the servers that were instantiated. The forked servers in that
model either inherit the server socket from the master and share the same port or the master
accepts all connections and load balances among the workers. In the first case the operating
system will be doing the load balancing among the server. Code listings 4 and 5 implement
those two models.
require 'socket'
(server_count-1).times do
break unless fork
end
# This code will run in all processes
# the kernel will synchronize calls for 'accept'
loop do
...
connection.close
end
Co d e Listin g 4

require 'socket'
workers = []
worker_count.times do
# a socket pair that the master uses to send connections to the worker
master, worker = UnixSocket.pair
workers << worker
if !fork
# this code runs in the worker
loop do
# the workers receives a connection from the master
conn = master.recv_io
request = HTTP.parse(conn.gets)
...
conn.puts status
conn.puts headers
conn.puts body
conn.close
end
end
end
# this is the master loop
loop do
w = workers.shift # pick a worker
w.send_io(connection) # send the connection to that worker
connection.close # the master closes the connection (worker has it now)
workers << w # worker is returned to the pool
end
Code Listing 5
Currently that is the only model that scales to multiple CPU cores in Ruby regardless of
whether we are talking about 1.8.x or 1.9.x.
3.3. Multi Threading
Much like Multi Process but this time we spawn multiple threads in the same process. A single
process can handle multiple concurrent requests that way. Each thread is handled in a
blocking manner but other threads are able to run while it is blocked. Usually a thread is
created per connection but pooling is also used to avoid the overhead of thread creation. Ruby
1.9 is totally different from 1.8 in this model. As it utilizes much more expensive to create
kernel threads. But on the other hand scheduling those is much faster than Ruby 1.8 threads.
Threads are not able to utilize more than one processor core because under 1.8 they are green

threads and under 1.9 there is a giant virtual machine lock that ensures that only one thread
runs at a time. That is to enable the use of non thread safe c extensions.
There are several variations of this model. One is to run several threads that implement the
full serving logic. Each of those listens on the socket, accepts connections, processes requests
and send back responses. Another model is an acceptor thread (usually the main thread) that
accepts connections and pass them over to worker threads that handle them. The models are
roughly implemented in code listings 6 and 7.
require 'socket'
require 'thread'
# function body implements the serving logic
def start
loop do
...
connection.close
end
end
(worker_count-1).times do
# create worker threads and each runs a server
Thread.new do
start
end
end
start # even the main thread can run its own server
Code Listing 6

require 'socket'
require 'thread'
loop do
# spawn a new thread to handle the connection
Thread.new do
...
connection.close
end
end
Code Listing 7
While threads are cheaper than processes they remain relatively expensive resources
(specially for Ruby 1.9.x). A server that spawns too many threads will waste considerable time
on activities not directly related to web serving like thread creation, locking and context
switching. That is why thread pooling is very common in threaded web servers. Code listing 8
presents another strategy that uses a fixed number of threads and a queue.
require 'socket'
require 'thread'
@queue = Queue.new
# spawn multiple worker threads
worker_count.times do
Thread.new do
loop do
# attempt to retrieve a connection from the queue and handle it
if conn = @queue.shift
request = HTTP.parse(conn.gets)
...
conn.puts status
conn.puts headers
conn.puts body
conn.close
end
end
end
end
loop do
# accept a new connection and put it in the queue
@queue << server.accept
end
Code Listing 8

It is important to mention that getting threads to work correctly is not a trivial problem. Since
threads run in the same process they get to share its data. Accessing this data safely from
multiple threads is a challenging problem that needs to be handled carefully. Deadlocks and
race conditions are the most common pitfalls of threaded applications. And they usually mean
that you are either locking more than you need or less than you need. It is a thin wire that you
have to walk upon when it comes to writing correct, non-trivial threaded applications.
It should be noted that while threads cannot utilize multiple CPU cores they stand out as the
easiest means of performing concurrent disk I/O along with multiple processes. This is
because disk I/O does not fit easily with event based architectures as we will see next.

3.4. Asynchronous Event Loop
In this concurrency model all I/O notifications are handled by a central tight loop calling a
method like select() or poll() on the set of file descriptors the program is dealing with.
Whenever there is activity an object is notified and it operates on the active IO object. Then it
delegates I/O notification back to the reactor core and so on.
An example would be registering interest in read activity for the server socket. Whenever
select() is called it will tell you whether the socket has data or not. If it has data then the
connection is accepted and inserted in the event loop to be processed whenever there is data
to read from it. After a connection is done with it should be removed from the event loop to
register disinterest in further events.
Such applications are single threaded and in that single thread many I/O operations can occur
concurrently. All I/O must be done in a non blocking manner though or the whole server will
block and not other operations will be done till this I/O completes.
Also it is worthy to note that this model provides I/O concurrency only. Processing
concurrency is not achieved since it is done in a single process that can only use one CPU. If a
request uses too much CPU time all other pending requests will have to wait till it finishes.
Also disk I/O will not be concurrent. This can only be done with multi processing or multi
threading (only when using Ruby1.9 for the case of multi threading). The issue is that file
system calls might block anyway even after you receive a ready notification on them. This
makes the use of file descriptors for real files in a select() call almost useless.
While select() and poll() are the most common method calls to poll they are hindered by the
fact that their performance is relative to the number of file descriptors they are watching. O(n)
to better describe it. While this is OK for a small set of file descriptors it is expensive for
servers with many concurrent clients. Modern event loops use platform specific methods like
epoll on Linux and kqueue on FreeBSD which get rid of the linear performance profile and
provide O(1) implementations that perform the same regardless of the file descriptors count.
An example asynchronous server is presented in code listing 9.

# note, this server can only handle data that can be read or written
# in a single shot. A correct implementation is more involved but details
# are removed here for clarity purposes
require 'socket'
@write, @write_actions = [], {}
@read, @read_actions = [], {}
# add the server to the read sockets and an action to fire once it is ready
@read << server
@read_actions[server] = proc{|server|
# accept a new connection and add it to the read list
@read << conn = server.accept
@read_actions[conn] = proc{|conn|
data = conn.read
...
# once a connection is readable then read the data
# and remove it from the read list
# then add it to the write list
@read.delete(conn)
@read_actions.delete(conn)
@write << conn
@write_actions[conn] = proc{|conn|
# once the connection is writable then write the response
# close the connection and remove it from the write list
conn.write(response)
conn.close
@write.delete(conn)
@write_actions.delete(conn)
}
}
}
loop do
# call select to see which sockets are ready
if res = IO.select(@read,@write,nil,0.01)
# some sockets have data to be read
res[0].each{|io|@read_actions[io].call(io)} if res[0]
# we can write to some sockets
res[1].each{|io|@write_actions[io].call(io)} if res[1]
end
end
Code Listing 9
It is clear from code listing 9 that writing servers based on an event loop is a much more
involved process that the previous models. Nevertheless, this model typically provides the best
performance.

3.5. Combinations
A server might opt to use multiple processes with multiple threads each or may be an event
loop will delegate connections to a pool of worker threads. The system on which the server will
run and the requirements of the application can help decide which is the best strategy for
handling I/O. Code listing 10 represents a hybrid multi processing and multi threading server.
require 'socket'
require 'thread'
(server_count-1).times do
break unless fork
end
# This code will run in all processes
# the kernel will synchronize calls for 'accept'
loop do
# spawn a new thread to handle the connection
Thread.new do
...
connection.close
end
end
Code Listing 10

4. Performance Bottlenecks
4.1. Introduction
Some metrics contribute heavily to web server performance. Let's go through a few of those
and see how they can affect a Ruby web server.
4.2. Data Copying
Servers usually send data back to clients. This data might come from a data fetch and a
rendering process but it also can come from a file on disk or data collected from a back end
server. In the last two cases the usual approach results in data being copied from kernel space
to the user program in the initial step and then from the user program to kernel space once
more as it is being sent to the client. There are other ways to send data directly without the
extra copying being involved. Such techniques are called zero-copy. One such technique is the
sendfile() method which sends data between two file descriptors (files, sockets or pipes)
directly via the kernel without passing it through the user program. Zero copy is particularly
useful in Ruby since not only memory copying is expensive but also is the creation of Ruby
objects and then garbage collecting them.
4.3. Context Switching
There are several forms of context switching. First any system call involves a context switch
from the user program to the kernel. Multiple threads and processes also perform context
switches to move the old one away and make way for the newly scheduled thread or process.
Traditionally Ruby interpreters did not do a great task of handling thread context switches
efficiently. This and the fact that excessive context switches are costly even for the most
optimized code should point us towards a conclusion that too much threads (or processes, or
fibers) is a bad idea. The general rule is to try to keep context switches at a minimum. We will
see how the Ruby server fare in that regard.

4.4. Lock Contention
When the server uses multiple threads and processes which share data then this data must be
properly locked to protect it from concurrent access. The problem is that locking is an
extremely hard subject and you either end up with coarse grained locking which results in too
many threads waiting for resources to be freed or even deadlocks if you are not careful or you
end up with a very complicated fine grained locking scheme which has a substantial overhead
and can easily introduce race conditions. The general rule is to avoid sharing data between
your concurrency primitives as much as possible.
4.5. Memory Management
For web servers lots of memory is being allocated and freed frequently. This poses a challenge
for the server's memory allocator/deallocator. Even worse, Ruby's garbage collector is famous
of being generally slow. Which adds a considerable overhead to the memory management
operation. Several techniques could be used to help with this problem. Like allocating the
least amount of objects and reusing as much of them as possible. We will look at how different
servers are attempting at this problem. The word on the street though is that Ruby is getting a
new GC very soon. Peeking into the fresh sources confirms that we are going to see a
generational garbage collector in Ruby's next release.
4.6. Blocking Operations
When servers perform any I/O operations they have to either wait for the operation to finish
or use some sort of nonblocking I/O. Many servers opt for simple blocking I/O in combination
with threads (and or processes) other use nonblocking I/O exclusively. We will look at how
the servers make sure that I/O operations don't keep them busy from serving other clients.
4.7. HTTP Parsing
Implementing the HTTP protocol is tricky. Due to many performance and security
considerations. We have mentioned earlier that most servers use some derivatives of
Mongrel's Ragel based C parser. We will see how this differs from rolling a hand written

parser in Ruby as is the case with Webrick.
4.8. The TCP Stack
The TCP protocol is in the heart of most server operations and hence it has a big impact on the
server's performance. The TCP protocol implementations provide a set of option for tuning
the TCP operations. These options can have a great effect on performance as they contribute
to saving bandwidth and time for certain operations. Example options for Linux are (similar
options are available for FreeBSD as well):
● TCP_NO_DELAY, eliminates wait time between transmissions, for keep alive requests
● TCP_CORK, only transmits full packets, very useful for large transfers
● TCP_DEFER_ACCEPT, delays accepting connections until clients send data
● TCP_QUICKACK, whether or not to send ACK flags in their own packets.
We will see how the servers use those options to their benefit.

5. Ruby Web Servers
5.1. Introduction
Earlier in this document we have looked at different concurrency models and explained some
of the bottlenecks facing server implementations. We are going to look at some Ruby web
servers and which concurrency models they apply and how do they handle the bottlenecks.
The list was based mostly on web server popularity among the Ruby community. It was
arranged in a chronological order to show how the Ruby community is adapting its web server
implementations as the technology evolves.
Missing from this list are some notable web servers like Ebb, Rucy, Flow, Unicorn and
Evented Mongrel. There must be other attempts at web servers that are not known to me as of
this writing. Nevertheless, the selected group represent a wide scope of architectures and
design strategies that make them a very representative sample.

5.2. WEBrick
WEBrick is a pioneering Ruby server that was created in the year 2000 by TAKAHASHI
Masayoshi and GOTOU Yuuzou. It dominated the scene as the leading standalone Ruby server
until Mongrel came out in 2006. WEBrick is a feature rich server that is written entirely in
Ruby. It has support for stuff like HTTPS, HTTP authentication and listening on several ports.
WEBrick has a very modular structure as well. Its core is concerned with network I/O, logging
and configuration while it delegates request handling to handlers that have a common
interface. WEBrick comes with several handlers out of the box that support static file serving,
CGI and other goodies. Extending WEBrick boils down to creating a new handler and adding
it to the set of existing ones.
WEBrick is implemented as a single process multi threaded server. Nothing prevents you
from starting several WEBricks, each listening on its own port and load balancing between
them via and external load balancer. But the server itself does not provide any
multiprocessing features of its own.
WEBrick's server is implemented as a never ending loop that checks the ready state of the
server socket. It does that via the select call. The main loop will spawn threads upon receiving
new connections. Those threads will wait for client connections to have data ready (using
another select call) and will them process the requests and invoke the appropriate handlers.
Since WEBrick is written entirely in Ruby, the HTTP parser it employs is hardly know for its
performance. Actually it is one of the biggest performance drains for WEBrick. This is due to
the fact that parsing requests involves a lot of string processing and generates a lot of garbage.
Thus taxing the CPU and the garbage collector heavily.
WEBrick's use of a multiple threads ensures that It can handle requests concurrently but not
in parallel since Ruby VMs cannot make use of multiple CPU cores yet. Due to the costs
involved with spawning too many threads in Ruby, it is expected that WEBrick will suffer as
the number of concurrent clients increases beyond a certain value. The benchmarking results
will give us more insight as to how WEBrick scales in the face of high loads.
Diagram 1 shows a simplified explanation of how WEBrick works.

Spawned Worker Threads

(one thread per connection)
Reads data from

connection then parses
the request, handles it
and closes the connection
when done
Accepts connections then

spawns new threads to
Server handle them
Server Class
Webrick
Diagram 1
In brief, WEBrick is a single process multi threaded application. Let's see how it attempts to
handle the various bottlenecks we discussed earlier.
1. Data Copying WEBrick is seemingly neglecting this issue. No attempts can be seen in
the code to minimize data copying.
2. Context Switching Since threading is used extensively, even when using select as
there's usually a select per connection thread. We can guess that a WEBrick server will
do a lot of context switches specially since all I/O operations are done independently
and grouping system call (again, like select) are seldom used for that purpose.
3. Lock Contention There are hardly any shared data between threads, this makes it
easy as too little locking (if any) is needed.
4. Memory Management WEBrick does not attempt to conserve memory aggressively.

There are hardly signs of reusable objects. A request will generate a good amount of
garbage objects that add a sizable overhead to the garbage collector.

5. Blocking Operations It appears that WEBrick is setting the connection sockets to

nonblocking before passing them to handlers. I am not entirely sure why it is doing that
as there are no traces of code that handles blocking exceptions (though highly unlikely
as the sockets are not operated upon except after a select call).
6. HTTP Parsing As we mentioned earlier, WEBrick parses HTTP in pure Ruby, hence
it is rather slow and heavy on the GC.
7. TCP Stack There are no traces of code attempting to optimize the TCP Stack.
Preferred Setup
Since Webrick does not utilize multiprocessing on its own and due to the fact that most Ruby
web frameworks (except a few exceptions) do not fully utilize multi-threading there must be a
way for Webrick to process multiple requests simultaneously. In this case, we start a handful
of Webrick servers (each on its own port) and pass requests to them from some proxy server.
Conclusion
WEBrick is a fine little server with loads of features but it builds atop a not very scalable I/O
model and doest not attempt to really optimize its internals and handle the known
bottlenecks. It is expected to perform poorly under heavy loads.

5.3. Mongrel
Early 2006 saw the earliest announcements of Mongrel on Ruby mailing lists by Zed A. Shaw.
Mongrel took the Ruby world by a storm. Since WEBrick was so slow users usually opted for
something like FastCGI but this was known to not be a very stable solution. It worked for
some but many had issues. Mongrel changed all that. Finally A Ruby web server that is fast
enough that you can rely on it on production. After Mongrel reached a stable state it became
the preferred deployment method for Ruby on Rails and other Ruby frameworks at the time.
Mongrel excelled by its unique implementation of the HTTP parser. It used a C parser that
was auto generated by Ragel from the HTTP specification. This made it very stable and fast
parser. And it was enough for Mongrel to surpass Webrick's performance by a wide margin.
Mongrel is composed of a single process that starts an acceptor thread. The acceptor thread
runs forever and whenever a connection is received it spawns a new thread to handle the
connection. Those spawned threads make use of the C based HTTP parser to parse requests .
Much like with Webrick, Mongrel's use of Ruby threads means that it cannot utilize multiple
CPU cores and that it will not be able to process requests in parallel, rather they will be
processed in a concurrent manner. It also means that each request is penalized by the cost of
creating a new thread which is not trivial specially for Ruby 1.9.x which uses native threads.
Aside from Mongrel's concurrency model, it also tries to optimize the TCP Stack. It enlarges
the server's socket listening queue to 1024 rather than the 5 provided by the default Ruby
implementation. It also sets TCP_CORK on the client connection sockets before serving the
response. This better aligns the response packets for optimal bandwidth usage.
One downside is that Mongrel does not support HTTP KEEP ALIVE connections. This means
that a new connection has to be established for each new request even if they originate from
the same client. Clients usually attempt to reuse the connections that were opened for earlier
requests. This prohibits their ability to do that.
Mongrel was usually used in a clustered configuration were several Mongrel processes were
started each listening to a different port and a proxy load balancer was used to pass requests
to those processes in a round robin fashion. This was the mostly adopted method of deploying

Mongrels since Rails was not able to utilize Mongrel's multiple threads. This approach was the
only way for Rails requests to run in parallel. Several proxy servers are paired with Mongrel.
The most popular are Apache and Nginx.
Diagram 2 shows a simplified explanation of how Mongrel works.
Spawned Worker Threads

(one thread per connection)
Reads data from

connection then parses
the request, handles it
and closes the
connection when done
Accepts connections
then spawns new
Acceptor threads to handle them
Thread
Mongrel
Diagram 2
To sum Mongrel, we can describe it as a single process multi threaded server. Here's how it
attempts to handle our set of bottlenecks.
1. Data Copying Mongrel does not try to avoid data copying in any way. All I/O goes
through the normal Kernel to Process to Kernel paths.
2. Context Switching Much like Webrick, Mongrel relies solely on threads to perform
concurrent I/O. This means lots of context switches each coming with its cost.
3. Lock Contention With almost no data shared between the different threads, locking
is largely a non issue for Mongrel. It can become a real issue if applications loaded
within Mongrel decide to share data. In such a case synchronization must be
introduced and this will most likely introduce some level of lock contention.

4. Memory Management Parser objects are reusable in Mongrel. Mongrel also tries to
minimize the use of string internally and replaces them with constants in an attempt to
reduce the amount of garbage generated during request handling. The efforts do yield
an improvement over Webrick's memory usage.
5. Blocking Operations Mongrel performs all I/O operations in a blocking manner.

This is basically due to its use of threads which will naturally block until I/O is finished
6. HTTP Parsing This is where Mongrel shines. Its HTTP parser is also its most touted
feature. A correct, secure and fast implementation in the form of auto generated C
code.
7. TCP Stack Mongrel optimizes the C stack using 2 methods. It increases the count of
connections that can wait in the server socket's listen queue. It also uses TCP_CORK to
optimize bandwidth usage.
Preferred Setup
Much like Webrick. Mongrel is usually setup as a cluster of processes (each with its own port)
behind a proxy server like Apache and Nginx. Those will handle the static file serving and pass
dynamic requests to the mongrel cluster.
Conclusion
Mongrel is the server that has the motto “faster is possible” and indeed it was a lot faster than
WEBrick. Even though it builds on the same concurrency model that Webrick uses it tries
hard to optimize the request path. The parser that it uses was widely adopted by other server
writers as we will see next. We expect Mongrel to provide decent performance even under the
face of moderate to high concurrent loads.

5.4. Thin
With the start of 2008 Marc-Andre Cournoyer released the first version of Thin. The new
thing about it was its departure from the thread-per-request practice of the former Ruby
servers. Rather it relied on the established EventMachine library for very fast, single threaded
based I/O. Actually what was cool about Thin was its integration of three solid components.
EventMachine as the I/O back-end, Mongrel's infamous parser and Rack as a Ruby web
application interface.
With these three components Thin offered out of the box performance and scalability that
exceeded any of the former offerings. Exceeding Webrick is understandable since it uses
Mongrel's much faster C based parser. But how can Thin outperform Mongrel when it is
sharing that bulk of code with it?.
As mentioned earlier, Thin departs from the thread-per-model and opts for an asynchronous
event loop. As we discussed earlier, those have several advantages over threaded
implementations since the require less memory and are free of context switching overheads.
Another point for Thin in the Thin vs. Mongrel issue is its support HTTP KEEP ALIVE
connections. Which means much less overhead with clients that will request multiple
resources from the server.
Internally Thin starts it event loop and waits for notifications on incoming requests. Once a
request arrives a connection is established and is then inserted in the event loop for
notifications on when it should be read from or written to. Aside from the network I/O,
requests are passed to Rack applications for processing. Due to the way Rack works and the
single threaded nature of Thin; only one request can be inside the Rack container at a time.
Recently several attempts have been made to avoid this limitation. Thin now offers an option
which will cause requests to be dispatched in a background pool of 20 threads. Another very
recent attempt is work done for asynchronous Rack processing.

Generally, Thin's request handling can be described as illustrated in Diagram 3:
Accepts connections and

registers them with the
reactor core
w
w
w w Workers objects are
Server w invoked once a
w connection has data.
They process the
requests and respond via
the reactor core
Notifies the server and

worker classes of all
network activity and
delegates the handling to
them
Reactor Core
(Event Machine)
Thin
Diagram 3
So basically Thin is a single process single threaded server that uses an asynchronous event
loop. It also has the option to dispatch requests to a thread pool. Looking at how it attempts to
solve the different bottlenecks we described earlier we can see the following:
1. Data Copying Thin does not attempt to avoid data copying in any way it even
delegates static file request handling to a very generic Rack module which fairs badly in
that regard.
2. Context Switching Here is where Thin shines. It is a single stack implementation

and hence does not context switch at all, except for the occasional error handling or if
you use the thread pool option.
3. Lock Contention There is virtually no lock contention. Except for the case when the
thread pool is used is which case threads must synchronize their use of the event loop.
4. Memory Management Thin tries hard to avoid generating many garbage objects
during request processing. It even tries to optimize all hash lookups by using constants

instead of strings. Where it fails though is in how it retrieves responses from Rack. It
does so in a way that guarantees that the data is buffered twice in memory before it is
being sent. This can easily lead to memory fragmentation specially with large response
bodies.
5. Blocking Operations Since Thin is relying in EventMachine most of the I/O

operations are done in a non-blocking manner. This helps a lot in reducing server
latencies.
6. HTTP Parsing Much like Mongrel. Thin uses the C based Ragel generated parser. It
also adds a few options like keep alive and chunked encoding support.
7. TCP Stack Thin does not try to optimize the TCP stack. Apparently it is relying on
EventMachine for this. But till this date, EventMachine does not perform any TCP
tuning to its sockets.
Preferred Setup
Like Mongrel and Webrick before it. Thin is usually setup as a cluster of processes (each with
its own port) behind a proxy server of some sort. It is recommended that the proxy handles
static file serving to offload that duty from the Thin cluster. Apache and Nginx are the most
popular options.
Conclusion
Thin deserves its name. It is a minimalistic server that acts as a thin glue for several robust
libraries. The careful selection of libraries and the use of event loops meant a server was built
that is more scalable and faster than the competition. Before we go into benchmarking we
expect that Thin will easily outperform Mongrel and Webrick, even as we scale to high
concurrent loads specially if we server quick small requests.

5.5. Passenger
The servers we described earlier are all self contained web server that can handle a request as
soon as it hits the physical server ports. The guys at Phusion thought differently. There are
many proven web servers out there and the part where Ruby is needed is running Ruby
applications rather than building the whole web stack. The idea is rather than build a
standalone server you augment an already existing one with Ruby processes that will handle
requests specific to the Ruby application. Furthermore, instead of instantiating those
processes for every request it will prefork a group of them and keep them around for faster
response time. Sounds a lot like CGI and FastCGI? Well yes, the Phusion team created their
own FastCGI look-a-like. Their implementation is much less general though and excels at
what it is intended to do.
Given a web server like Apache or Nginx, Passenger is added to it as a module. This module is
configured to handle certain types of requests. Basically the web server falls back to it when it
cannot find a static file to serve in the configured location for the Ruby application. When the
module starts it initiates a Ruby process that we can regard as a master process since it will be
responsible for all the Ruby processes handling the application. The master then forks the
worker processes. Each of those workers is connected to the master and it also initiates a
UNIX socket connection with the module.
Once the server passes a request over to the module it sends it to one of the worker processes.
The request is sent in a certain format agreed upon by the module and the workers. Along
with the request information the module also sends the client connection itself using the
UNIX socket send_io facility. At this point the web server can totally ignore this client and
even close the connection from its side. Passenger workers will be handling it from this point
forward.
The worker is basically a single threaded process that handles one request at a time. This
limits the ability of Passenger servers to handle concurrent requests beyond the number of
forked processes. Passengers ships with a special version of Ruby though that is optimized for
the forking model and reportedly saves memory by a factor of one third over vanilla Ruby.

Thus enabling the use of more worker processes per application. We will be doing our testing
using the stock Ruby 1.9.1 though which does not utilize these changes.
Diagram 4 describes the internals of Passenger:
Preforked Worker
Processes
Workers receive client

connections (via fd
passing) and the parsed
request headers. The
w w w w request is handled and
then the worker will send
the response directly to
the client
Receives parsed
requests from web server
then sends them down to
the workers along with
the client connection
Apache/NGINX itself
module
Passenger
Diagram 4
Passenger is the first genuinely multi-process server that appears in this review. Even though
it is common practice to use a cluster of Mongrel or Thin servers behind a proxy it requires
manual configuration and each process occupies its own port. This does not equate the
transparent multiprocessing capabilities provided by Passenger. Passenger also differs in how
it attempts to handle the bottlenecks as follows:
1. Data Copying Passenger does not do much about this internally but it does a great
job by passing the file descriptors to worker processes hence not needing to send data
to the clients through the parent web server. All other implementations pass data to
upstream proxies before it is sent to the clients.
2. Context Switching This is limited to process context switching. If not many worker
processes are used then the application will not face a lot of context switches while

being able to utilize multiple CPU cores at the same time.
3. Lock Contention No shared data means no lock contentions. Passenger does not rely
on sharing inter process data.
4. Memory Management This is the ares that most effort was put to make Passenger
use less memory. The Fork friendly Ruby implementation plus the ability of the master
process to recycle worker processes from time to time ensure that Passenger is the
most memory efficient Ruby application server.
5. Blocking Operations Not much is done here as the processes are all single threaded.
Multi threaded workers have appeared recently and might be used more as Rails and
other Rack libraries get more robust thread safety.
6. HTTP Parsing Passenger provides no HTTP parser at all. It relies instead on the web
server which parsed the request already. This makes lots of sense and prevents
duplicating efforts.
7. TCP Stack On its own, Passenger does not optimize the TCP stack a lot. But the web
server can set some TCP parameters on the connection before it is sent to Passenger.
Preferred Setup
Passenger is the easiest server to setup in this survey. You only need to configure Apache or
Nginx and then forget about it. Passenger will control the process loading and reloading if
needed. Using Ruby 1.9.x with Passenger does it a disservice though, as Passenger can be
coupled with Ruby Enterprise Edition and use it in conjunction with UNIX's Copy-On-Write
semantics to greatly reduce memory consumption of your servers.
Conclusion
Passenger can be described as a (finally) working FastCGI for Ruby. Along with several
memory saving optimizations and a dead easy configuration it makes much sense that it
suddenly became the preferable deployment method for Ruby web applications. The model
does not scale very well though due to the lack of a scalable concurrency model. But it fares
well for most Ruby web applications that are inherently single threaded, like most Ruby on
Rails based applications.

6. Benchmarks
6.1. Introduction
After the discussion of the several architectures we wanted to verify that the expectations will
actually map lab testing results. So a battery of benchmark tests were prepared and each of
the servers took its turn in running those tests against.

6.2. Benchmarking Methodology
The servers where tested against various workloads. The workloads were constructed to stress
various aspects of the servers. The following tests were performed:
1. Static File Serving
1. Serving a very small 200B file
2. Serving a small 1KB file
3. Serving a small 30KB file
4. Serving a medium 126KB file
5. Serving a large 1MB file
6. Serving a large 10MB file
7. Serving a very large 170MB file
2. Dynamic content serving
1. A Rails application that returns “Hello World!”
2. A Rails application that fetches a record from DB and renders an ERB template
3. A Rails application that fetches a record from DB and renders it as YAML
4. A Rails application that does heavy processing (Marshals lots of objects)
5. A Rails application that sends a large 1MB response
3. Multi-process dynamic content serving (2 processes proxied by Apache)
1. A Rails application that returns “Hello World!”
2. A Rails application that fetches a record from DB and renders an ERB template
3. A Rails application that fetches a record from DB and renders it as YAML
4. A Rails application that does heavy processing (Marshals lots of objects)

Thin was tested twice, once with worker threads enabled and once when they were disabled.
Passenger was not tested for static file serving. Apache Bench (ab) was used for
benchmarking.

6.3. Test Setup
The tests were done on a DELL Inspiron 1520 laptop. It is equipped with Core2Duo T700o
dual core CPU that is clocked at 2.0GHZ and is equipped with 2MB of cache. The system also
has 2GB of RAM and a 5400RPM hard drive.
The System had only one of the servers working on any given time to make sure all resources
were dedicated to the active test.
Some tests were repeated several times and anomalies were excluded. Long running tests
were not repeated as they ran long enough to produce stable results. Only when anomalies
where detect the long running tests were repeated to ensure a constant behavior is being
reported.
A warm up run was done to each server before an actual testing round to make sure all
required files are loaded and that the caches are filled.
Since this is a dual core processor we only used one process for all the servers. This was due to
the fact that the tests hardly had any blocking I/O operations and that the 2 cores will be split
between the server under test and Apache Bench itself as we did all the tests on the back loop
to eliminate network overhead.

6.4. Results
6.4.1. Serving Static Files
Serving a very small 200B file
Normal
Webrick Mongrel Thin Thin

(threaded)
10 users 10K requests 2931 1805 3034 2335
100 users 10K requests 295 1704 3163 2456
1K users 10K requests 288 915 3016 2398
3500
3000
Requests/Second
2500
Webrick
2000
Mongrel
1500
Thin
1000
Thin (threaded)
500
0
10 users 10K requests 100 users 10K requests 1K users 10K requests
Keep alive
Webrick Thin Thin (threaded)

10 users 10K requests 194 3780 2636
100 users 10K requests 320 3847 2814
1K users 10K requests 343 3610 3129
5000
4000
Requests/Second
3000 Webrick
Thin
2000 Thin (threaded)
1000
0
1 All results represent the number of requests per second

Serving a small 1KB file
Normal

(threaded)
10 users 10K requests 284 1715 2976 2065
100 users 10K requests 307 1579 3137 2590
3500
3000
Requests/Second
2500
Webrick
2000
Mongrel
1500
Thin
1000
Thin (threaded)
500
0
Keep alive

100 users 10K requests 307 3864 2952
5000
4000
Requests/Second
3000 Webrick
Thin
2000 Thin (threaded)
1000
0

Serving a small 30KB file
Normal

(threaded)
10 users 10K requests 297 1120 2230 1650
100 users 10K requests 293 1310 2217 1816
2500
2000
Requests/Second
1500 Webrick
Mongrel
1000 Thin
500 Thin (threaded)
0
Keep alive

100 users 10K requests 323 2487 1925
3000
2500
Requests/Second
2000 Webrick
1500 Thin
Thin (threaded)
1000
500
0

Serving a medium 126KB file
Normal

(threaded)
10 users 10K requests 275 932 1165 950
100 users 10K requests 272 922 1097 924
1400
1200
Requests/Second
1000
Webrick
800
Mongrel
600
Thin
400 Thin (threaded)
200
0
Keep alive

100 users 10K requests 303 1178 1084
1400
1200
Requests/Second
1000
Webrick
800
Thin
600
Thin (threaded)
400
200
0

Serving a large 1MB file
Normal

(threaded)
10 users 40 requests 162 156 75 155
20 users 40 requests 140 152 148 136
40 users 40 requests 66 150 139 134
200
Requests/Second
150
Webrick
100 Mongrel
Thin
50 Thin (threaded)
0
10 users 40 requests 20 users 40 requests 40 users 40 requests
Keep alive

10 users 40 requests 144 103 143
200
Requests/Second
150
Webrick
100 Thin
Thin (threaded)
50
0

Serving a large 9MB file
Normal

(threaded)
10 users 40 requests 42 30 21 20
50
40
Requests/Second
30 Webrick
Mongrel
20 Thin
Thin (threaded)
10
0
Keep alive

35
30
Requests/Second
25
Webrick
20
Thin
15
Thin (threaded)
10
5
0

Serving a huge 170MB file
Normal

(threaded)
10 users 40 requests 2 1.7 0.1 0.1
20 users 40 requests 2 1.8 FAILED FAILED
40 users 40 requests 1.8 1.6 FAILED FAILED
2.5
2
Requests/Second
1.5 Webrick
Mongrel
1 Thin
Thin (threaded)
0.5
0
Keep alive

10 users 40 requests 2 0.1 0.1
20 users 40 requests 1.8 FAILED FAILED
40 users 40 requests 1.5 FAILED FAILED
2.5
2
Requests/Second
1.5 Webrick
Thin
1 Thin (threaded)
0.5
0

6.4.2. Serving Dynamic Requests (Single Rails Process)

Serving 'Hello World!'
Normal
Webrick Mongrel Thin Thin Passenger

(threaded)
10 users 1K requests 169 438 650 589 497
100 users 1K requests 209 451 613 531 509
1K users 1K requests 250 446 615 529 324
700
600
Requests/Second
500 Webrick
400 Mongrel
300 Thin
200 Thin (threaded)
Passenger
100
0
Keep alive
Webrick Thin Thin (threaded) Passenger

10 users 1K requests 185 622 614 525
100 users 1K requests 268 611 597 517
700
600
Requests/Second
500
Webrick
400
Thin
300
Thin (threaded)
200 Passenger
100
0

Fetching a record and rendering an ERB template
Normal

(threaded)
10 users 1K requests 115 309 365 337 305
100 users 1K requests 165 299 350 335 316
400
350
Requests/Second
300
Webrick
250
Mongrel
200
Thin
150
Thin (threaded)
100
Passenger
50
0
Keep alive

10 users 1K requests 160 363 348 312
100 users 1K requests 207 339 342 303
400
350
300
Requests/Second
250 Webrick
200 Thin
150 Thin (threaded)
Passenger
100
50
0

Fetching a record and rendering it as YAML
Normal

(threaded)
10 users 1K requests 149 195 200 200 183
100 users 1K requests 145 192 199 190 186
250
200
Requests/Second
Webrick
150
Mongrel
100 Thin
Thin (threaded)
50 Passenger
0
Keep alive

10 users 1K requests 130 202 198 183
100 users 1K requests 147 194 193 184
250
200
Requests/Second
150 Webrick
Thin
100 Thin (threaded)
Passenger
50
0

Marshalling lots of objects (500 complex objects per request)

(threaded)
10 users 100 requests 41 44 43 44 45
100 users 100 requests 41 42 45 43 44
50
40
Requests/Second
Webrick
30
Mongrel
20 Thin
Thin (threaded)
10 Passenger
0
10 users 100 requests 100 users 100 requests

Sending a large 1MB response

(threaded)
10 users 100 requests 41 37 55 55 79
100 users 100 requests 41 37 54 54 79
100
80
Requests/Second
Webrick
60
Mongrel
40 Thin
Thin (threaded)
20 Passenger
0
Sending a very large 10MB response

(threaded)
10 users 100 requests 6 5 7 7 10
100 users 100 requests 6 5 2 7 10
12
10
Requests/Second
8 Webrick
Mongrel
6
Thin
4 Thin (threaded)
2 Passenger
0

6.4.3. Serving Dynamic Requests From A Proxied Cluster (2 Processes)

Serving 'Hello World!'
Normal

(threaded)
10 users 1K requests 258 512 675 767 544
100 users 1K requests 307 508 678 800 558
1000
800
Requests/Second
Webrick
600
Mongrel
400 Thin
Thin (threaded)
200 Passenger
0
Keep alive

10 users 1K requests 285 674 900 565
100 users 1K requests 330 680 884 552
1000
800
Requests/Second
600 Webrick
Thin
400 Thin (threaded)
Passenger
200
0

Fetching a record and rendering an ERB template
Normal

(threaded)
10 users 1K requests 250 375 444 477 324
100 users 1K requests 271 395 456 449 329
600
500
Requests/Second
400 Webrick
Mongrel
300
Thin
200 Thin (threaded)
100 Passenger
0
Keep alive

10 users 1K requests 241 477 479 337
100 users 1K requests 267 447 508 335
600
500
Requests/Second
400
Webrick
300 Thin
Thin (threaded)
200
Passenger
100
0

Fetching a record and rendering it as YAML
Normal

(threaded)
10 users 1K requests 201 254 278 278 214
100 users 1K requests 210 257 275 271 212
300
250
Requests/Second
200 Webrick
Mongrel
150
Thin
100 Thin (threaded)
50 Passenger
0
Keep alive

10 users 1K requests 190 272 275 224
100 users 1K requests 204 268 275 219
300
250
Requests/Second
200
Webrick
150 Thin
Thin (threaded)
100
Passenger
50
0

Marshalling lots of objects (500 complex objects per request)

(threaded)
10 users 100 requests 62 70 71 71 67
100 users 100 requests 62 69 71 70 66
80
70
Requests/Second
60
Webrick
50
Mongrel
40
Thin
30
Thin (threaded)
20
Passenger
10
0

Sending a large 1MB response

(threaded)
10 users 100 requests 54 49 71 67 71
100 users 100 requests 54 47 71 66 70
80
70
Requests/Second
60
Webrick
50
Mongrel
40
Thin
30
Thin (threaded)
20
Passenger
10
0
Sending a very large 10MB response

(threaded)
10 users 100 requests 7 7 9 9 10
100 users 100 requests 3 7 9 9 10
12
10
Requests/Second
8 Webrick
Mongrel
6
Thin
4 Thin (threaded)
2 Passenger
0

7. Benchmark Analysis
7.1. Static File Serving Performance
Generally speaking it is not recommended to utilize your Ruby server in static file serving. It is
better to offload this task to something like Apache or Nginx and free up the Ruby server to
serve dynamic requests (in Rails or any other Ruby script).
Nevertheless, these tests are interesting since they simulate (to an extent) how the servers
would fare when serving cached content. In a cached content situation there is no much data
processing and rendering, you just fetch the cache then send it through the wire. This is very
similar to static content serving.
The benchmarks can be split to two categories, small file sizes and large file sizes. Let's take a
closer look at each.
7.1.1. Small File Sizes

It is apparent that Thin leads the pack with a healthy margin, approaching a very respectable
4000 request/sec in one of the cases. While still less than half what Nginx could deliver in the
same situation it is still a very good show from an Ruby based server. Second comes Thin
again but in the threaded flavor, apparently threads and EventMachine are not happy
delivering too many quick responses. It is worth noting though that for all threaded Thin tests
the epoll back-end was disabled. Third comes Mongrel, which delivers roughly half to two-
thirds the performance of Thin. Then it the tail of theses tests Webrick establishes its
presence. Too slow to be even considered, many times it was 10x as slow as the competition.
Passenger was not tested here since it relies on the host web server (either Apache or Nginx)
for static file serving.
7.1.2.Large File Sizes

I was in for a good surprise when I performed the large file tests. I used a 1MB, a 9MB and a
170MB files. The results was so strange I kept repeating the tests but always getting the same

results. As the file size increases the performance profiles are completely reversed. Now
Webrick is leading the pack, Mongrel is second and Thin is trailing, with the threaded a
version a little faster than the pure evented one. This is the exact opposite of the small file
results. In some of the largest file tests Thin would even completely fail to deliver the files.
Let's attempt to understand what is going on here.
It is expected that Webrick and Mongrel will deliver similar performance. They are both built
using a very similar I/O strategy, Mongrel's advantage, which is the much faster HTTP parser
would not make much of a difference for long running requests which will be saturated by
disk and subsequently network I/O. Still, one would expect that Mongrel will outperform
Webrick by even a small margin due to its use to TCP-CORK which provides better TCP
packet alignment. Nevertheless, the results are not a huge departure from the expectations
even though they were tilted in the other direction. The real problem here are the results for
the Thin server. It's very strange that the fastest of the group suddenly becomes the slowest,
even failing at times.
Digging deeper into Thin's file serving code reveals the culprit. Thin is relying on the
Rack::File module for file serving. Rack::File is handled much like any other Rack middle-
ware. Thin will grab the response from the Rack::File module, buffer it and then give way to
EventMachine to send it as it sees fit. What happens is that the file (multiple instances of it if
we have concurrent clients) will be fully loaded in memory before the response is sent. Here's
how it works inside Thin:
1. When a client sends a group of concurrent requests they will be buffered in the Thin
socket's kernel buffers
2. EventMachine will notice that the socket is now readable and will loop 10 times each
time reading one request from the socket and passing it to Thin (obviously the loop will
be broken if there are less than 10 requests)
3. Thin parses the request and passes it to Rack::File. Thin then reads the response 4KB
at a time from Rack::File and sends it via the EventMachine::Connection.
4. During the whole response reading process, Thin does not give way for the
EventMachine to send any of the buffered data.

5. This process continues for (at max) 9 more requests (EventMachine attempts to read
requests from the server socket 10 at a time), all accumulating the responses in the
EventMachine buffers before a single byte is actually sent.
6. Once the batch of incoming requests are done, EventMachine starts to send responses
from those requests and accept new ones if any.
As one can see from the above, sending 10 concurrent requests for 170MB files results in
roughly 1.6GB memory usage before Thin starts sending back responses and freeing memory.
This explains why Thin fails in some benchmarks, it simply consumed all the system memory
and hitting the disk results in very low performance. To the point that responses are delayed
beyond the timeout set by Apache Bench.
So, why does Thin perform a little better with the support of threads? Let's see the processing
flow in that case and try to understand.
1. When a client sends a group of concurrent requests they will be buffered in the Thin
socket's kernel buffers
2. EventMachine will notice that the socket is now readable and will loop 10 times each
time reading one request from the socket and passing it to Thin (obviously the loop will
be broken if there are less than 10 requests)
3. Thin parses the request and passes it to Rack::File in a deferred thread. After the file is
located and opened the thread returns the response. The main EventMachine thread
will then run a call back which will start reading the response 4KB at a time from
Rack::File and sends it via the EventMachine::Connection.
4. The main difference here is that before running the callback (and between callbacks)
EventMachine is free to perform its own operations so it gets a chance (a slim one that
is) to send accumulated data before processing callbacks of further requests.
This is why the threaded flavor of Thin outperforms the pure evented one when serving large
files. There is also an interesting observation here. EventMachine attempts to accept at most
10 requests every time the server socket is readable. In practice this means that for long
running requests you can get a considerable delay with high concurrency. Consider the

following scenario:
1. Server is idle, waiting for connections
2. 20 requests come at the same time (each takes 1 second to process)
3. Socket is flagged as readable, no writable sockets (yet)
4. First connection is extracted, request is processed (1 second)
5. Resulting data from step 4 now buffered waiting for a new reactor run
6. Steps 4 & 5 are repeated 9 more times (10 seconds passed now)
7. After the 10 requests are processed the reactor is finally free for another run
8. Now the reactor checks its sockets again, we have 10 writable sockets with data on
them. But we also have a server socket with 10 pending connections.
9. The reactor will process with the server socket first and repeat the steps from 4 to 6
10. Now that 20 seconds has passed, it starts to process the writable sockets and send back
responses.
Assuming a negligible response sending time we find that the requests were answered in the
following manner, min : 20s, avg : 20s, max : 20s. On the other hand, if we were able to send
the responses immediately it could've been, min : 1s, avg : 10s, max : 20s which is a much
better distribution.

7.2. Dynamic Request Serving Performance
We have 3 categories of requests tested. The first are quick requests with little processing and
small responses. The second are requests with average response sizes but heavy processing.
The third are requests with little processing but large responses. I/O time was not a major
factor in most of the tests. Since the dataset is really small and it was surely present in the
database server's cache.
7.2.1. Quick and Small Responses

Thin leads the pack as expected. It is really optimized for those work loads. Passenger and
Mongrel fight for the second position. Passenger exhibits a strange behavior though, while it
passes Mongrel in the single process tests, it is a little bit slower in the dual process tests.
Needless to say, Webrick is the lowest performer by an embarrassingly wide margin.
7.2.1. Heavy Processing but Small Responses

Once all servers start consuming a lot of CPU the picture changes a lot. The difference have all
evaporated and we see all the servers performing almost the same. This is due to the fact that
the differences in the request handling pipeline are negligible when compared to long runs of
CPU crunching code.
7.2.1. Little Processing but Large Responses

Now Passenger is leading the pack. It is much better on the system memory even though it is
also buffering everything in memory. But since it handle one request at a time per process it
doesn't eat crazy memory amounts like Thin does. Thin follows as the second server here
which was surprising since Webrick and Mongrel were faster in large static file tests.
Apparently the file serving code of Thin is guilty. And while Thin is an overall very good server
the file serving module drags it behind on large files. Hence it is able to exceed its rivals on
large dynamic responses. The rest of the pack remain the same with Webrick exceeding
Mongrel a little.

8. Conclusion
I should not be making conclusions given the abundance of the presented data which can help
you decide on your own. But I know that some people may want to skip all the bloat and read
directly what the author thinks he found out. So, without further ado, here are my conclusions
for those deploying Ruby 1.9.x applications.
1. Offload your file serving to the likes of Nginx and Apache (in that order)
2. It won't matter much which server you go with in the case of Rails (unless it is called
Webrick, which should be avoided). For ease of use I recommend Passenger even
though Thin is a little faster for quick requests.
3. If you are building an API or a service that perform fast and quick operations then look
no further than Thin. This one will excel in those applications.

References
1. The C10K Problem, http://www.kegel.com/c10k.html
2. Benchmarking BSD and Linux http://bulk.fefe.de/scalability/
3. High-Performance Server Architecture http://pl.atyp.us/content/tech/servers.html
4. Mongrel Web Server http://mongrel.rubyforge.org/
5. Thin Web Server http://code.macournoyer.com/thin/
6. Phusion Passenger (Mod Rails) http://www.modrails.com/
7. Ruby Enterprise Edition http://www.rubyenterpriseedition.com/
8. Event Machine http://rubyeventmachine.com/
9. Ruby Event Machine – The Speed Demon http://www.igvita.com/2008/05/27/ruby-

eventmachine-the-speed-demon/
10. Concurrency is a Myth in Ruby http://www.igvita.com/2008/11/13/concurrency-is-a-

myth-in-ruby/
11. I/O models: how you move your data matters http://timetobleed.com/io-models-how-
you-move-your-data-matters/
12. Threading models: So many different ways to get stuff done

http://timetobleed.com/threading-models-so-many-different-ways-to-get-stuff-done/
13. Why Events Are a Bad Idea (for High Concurrency Servers)
http://www.usenix.org/events/hotos03/tech/vonbehren.html

The Ruby 1.9.x Web Servers Booklet

Uploaded by

Copyright:

Available Formats

The Ruby 1.9.x Web Servers Booklet

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Ruby 1.9.x Web Servers Booklet

Uploaded by

Copyright:

Available Formats

What is the purpose of this survey?

What is the purpose of this survey?

What web servers are covered in the analysis?

What web servers are covered in the analysis?

The Ruby 1.9.

x Web Servers Booklet

Ruby 1.9.x Web

A Survey of Architectural Choices & Analysis of Relative Performance 1/60

A Survey of Architectural Choices & Analysis of Relative Performance 2/60

About the Author

Muhammad A. Ali, A.K.A oldmoe, is a software developer from Alexandria,

A Survey of Architectural Choices & Analysis of Relative Performance 3/60

A Survey of Architectural Choices & Analysis of Relative Performance 4/60

A Survey of Architectural Choices & Analysis of Relative Performance 5/60

A Survey of Architectural Choices & Analysis of Relative Performance 6/60

2. Web Server Primer

A Survey of Architectural Choices & Analysis of Relative Performance 7/60

3.2. Multi Processing

Here is a sample forking server written in Ruby:

A Survey of Architectural Choices & Analysis of Relative Performance 8/60

A Survey of Architectural Choices & Analysis of Relative Performance 9/60

3.3. Multi Threading

A Survey of Architectural Choices & Analysis of Relative Performance 10/60

A Survey of Architectural Choices & Analysis of Relative Performance 11/60

A Survey of Architectural Choices & Analysis of Relative Performance 12/60

A Survey of Architectural Choices & Analysis of Relative Performance 13/60

3.4. Asynchronous Event Loop

An example asynchronous server is presented in code listing 9.

A Survey of Architectural Choices & Analysis of Relative Performance 14/60

A Survey of Architectural Choices & Analysis of Relative Performance 15/60

A Survey of Architectural Choices & Analysis of Relative Performance 16/60

4.2. Data Copying

4.3. Context Switching

A Survey of Architectural Choices & Analysis of Relative Performance 17/60

4.4. Lock Contention

4.5. Memory Management

4.6. Blocking Operations

4.7. HTTP Parsing

A Survey of Architectural Choices & Analysis of Relative Performance 18/60

parser in Ruby as is the case with Webrick.

4.8. The TCP Stack

● TCP_DEFER_ACCEPT, delays accepting connections until clients send data

● TCP_QUICKACK, whether or not to send ACK flags in their own packets.

A Survey of Architectural Choices & Analysis of Relative Performance 19/60

5. Ruby Web Servers

A Survey of Architectural Choices & Analysis of Relative Performance 20/60

Diagram 1 shows a simplified explanation of how WEBrick works.

A Survey of Architectural Choices & Analysis of Relative Performance 21/60

Spawned Worker Threads

Reads data from

Accepts connections then

4. Memory Management WEBrick does not attempt to conserve memory aggressively.

A Survey of Architectural Choices & Analysis of Relative Performance 22/60

5. Blocking Operations It appears that WEBrick is setting the connection sockets to

A Survey of Architectural Choices & Analysis of Relative Performance 23/60

A Survey of Architectural Choices & Analysis of Relative Performance 24/60

Diagram 2 shows a simplified explanation of how Mongrel works.

Spawned Worker Threads

Reads data from