Bruno Fernandez-Ruiz

I spent this weekend testing and installing nginx and I have said Adios to Apache! Let me tell you why I have done that.

I am running this blog on a slicehost.com 256Mb Xen VPS, using WordPress. I have been using Apache 2.2 for almost a year now without any problems. But recently I deployed a little Ruby on Rails application using mongrel for my wife’s business, and then suddenly I ran out of memory. Each Apache worker was easily taking 15 to 20Mb, even with a really cut-down configuration and built on Gentoo. With three Mongrel backends and the MySQL server I was starting to swap to disk.

I tried FastCGI instead of Mongrel, but it only cut down from about 50Mb to 40Mb for each process, not enough really to make a difference. I had to trim the fat out of the web server.

So about dropping Apache. First I considered Pound as a load balancer, but given that I would also still need a web server for static content in Rails (not wanting to serve that with Mongrel) I decided to look for alternatives.

Lighty and FastCGI tests did not even complete, so I gave up on Lighty. The FastCGI processes were becoming Zombie for some unknown reason.

Then I tried nginx. I have 5 workers, each consuming around 3Mb. Wow … I call that Spring savings.

nginx is sending all blog traffic to a couple of php-cgi handlers. I tried both TCP and Unix sockets, but went for TCP since I was getting dropped requests and stalling processes for Unix sockets. php-cgi is not very fast, but it’s stable and it has a reduced memory footprint.

Then nginx is acting as a reverse proxy and balancing across 3 Mongrels. I tried also FastCGI, but again it’s not stable.

And man is this thing flying or what!? I have enough memory spare to make MySQL happy. Plus I am caching a lot of Rails generated content that nginx serves happily.

Now, some numbers from Apache Benchmark (running on Linux 2.6.16.29-xen #3 SMP, x86_64 Dual Core AMD Opteron(tm) Processor 265 AuthenticAMD GNU/Linux, Xen host with 256Mb):
nginx + pure HTML content:


Concurrency Level: 50
Time taken for tests: 0.96657 seconds
Complete requests: 1000
Failed requests: 0
Write errors: 0
Requests per second: 10345.86 [#/sec] (mean)
Time per request: 4.833 [ms] (mean)
Time per request: 0.097 [ms] (mean, across all concurrent requests)


nginx + php Wordpress home page (not cached, no APC):

Concurrency Level: 50
Time taken for tests: 192.387414 seconds
Complete requests: 1000
Failed requests: 0
Write errors: 0
Requests per second: 5.20 [#/sec] (mean)
Time per request: 9619.371 [ms] (mean)
Time per request: 192.387 [ms] (mean, across all concurrent requests)


nginx + mongrel (not cached):

Concurrency Level: 50
Time taken for tests: 8.290116 seconds
Complete requests: 1000
Failed requests: 0
Write errors: 0
Requests per second: 120.63 [#/sec] (mean)
Time per request: 414.506 [ms] (mean)
Time per request: 8.290 [ms] (mean, across all concurrent requests)


Sure, it’s not a scientific test by any means, but it tells you something. And so far, no complains. If I ever outgrow the server, I can always get Pound on front of a couple of nginx web servers.

One of the most useful recent additions to my Gnome desktop has been Beagle, Nat’s personal desktop search daemon. I also discovered that in Ubuntu Feisty there is a deskbar that traces all the actions you do from your desktop, including your web activity and beagle searches. The more you use it, the more relevant it becomes since humans are repetitive.

Then, it happened: Tim pointed me this morning to Beagle++, a semantic desktop search engine based on Beagle, and that triggered the thought.

There ain’t a semantic web. No, you read me well, there won’t be one. It won’t be the Web 3.0. No. The next revolution will be on the OS, not on the web.

Web 2.0, user generated content, and data mashups present a taxonomy of data available from an almost infinite variety of sources. And that will continue to grow. But understanding the meaning of data is something that belongs to our brains. The ability of a server to interact with a human being is very limited and it would require a tremendous flow of data between us and many servers out there. Whereas our desktops currently track much more information which can be used to understand the meaning of for example, a search query.

Eventually, the desktop and the OS will capture our facial expressions, our mood, our feelings. Really, it’s not that hard.

What this will mean is that the applications that deliver the semantic web will actually reside on our desktops, not on the server. There won’t be a need for processing on the server. The semantic analysis will happen on the desktop.

When you really think about it, it sort of makes sense. People are reluctant to leave a behavioral trail on the web, let alone have things such as feelings captured on somebody’s database. You want to keep that privacy, and ensure that sort of data lives somewhere where we have much better control of it: our desktops.

Web 3.0 will be about client-side applications, powered on Javascript, Flash, Java Web Start, .NET Smart clients, etc. The current browsers will change to integrate better with the desktop. The OS will change to feed a lot of behavioural data to the desktop. The desktop will control our semantic experience. The OS will become a browser, and the browser will become an OS. The OS will become a search engine, and the search engine will be our OS.

Don’t believe me? Well, I am not asking you to do so … my good friends at MIT have already proven it:

• Re:Search, a search engine on past search results: http://people.csail.mit.edu/teevan/rse/
• Exhibit, a CSS+Javascript mashup technology with no server-side processing: http://simile.mit.edu/exhibit/
• The Universal Information Client, a semantic office/browser/…: http://haystack.csail.mit.edu/home.html

Who said Gnome wasn’t cool?