How Facebook Works
Facebook is a wonderful example of the network effect, in which the value of a network to a user is exponentially proportional to the number of other users that network has.
Facebook’s power derives from what Jeff Rothschild, its vice president of technology, calls the “social graph”–the sum of the wildly various connections between the site’s users and their friends; between people and events; between events and photos; between photos and people; and between a huge number of discrete objects linked by metadata describing them and their connections.
Facebook maintains data centers in Santa Clara, CA; San Francisco; and Northern Virginia. The centers are built on the backs of three tiers of x86 servers loaded up with open-source software, some that Facebook has created itself.
Let’s look at the main facility, in Santa Clara, and then show how it interacts with its siblings.
The top tier of the Facebook network is made up of the Web servers that create the Web pages that users see, most with eight cores running 64-bit Linux and Apache. Many of the social network’s pages and features are created using PHP, a computer scripting language specialized for simple, automated functions. But Facebook also develops complex core applications using a variety of full-featured computer languages, including C++, Java, Python, and Ruby. To manage the complexity of this approach, the company created Thrift, an application framework that lets programs compiled from different languages work together.
The bottom tier consists of eight-core Linux servers running MySQL, an open-source database server application. Rothschild estimates that Facebook has about 800 such servers distributing about 40 terabytes of user data. This tier stores all the metadata about every object in the database, such as a person, photo, or event.
The middle tier consists of caching servers. Even 800 database servers can’t serve up all the needed data: Facebook receives 15 million requests per second for both data and connections. Bulked-up cache servers, running Linux and the open-source Memcache software, fill the gap. About 95 percent of data queries can be filled from the cache servers’ 15 terabytes of RAM, so that only 500,000 queries per second have to be passed to the MySQL databases and their relatively slow hard drives.
Photos, videos, and other objects that populate the Web tier are stored in separate filers within the data center.
The San Francisco facility replicates the Web and cache tiers, as well as the filers with the database objects, but it uses the Santa Clara MySQL database tier.
The Virginia data center is too far away to share MySQL databases: with 70 milliseconds of Internet delay, give or take, it just won’t work. Thus, it completely duplicates the Santa Clara facility, using MySQL replication to keep the database tiers in sync.
What’s next for Facebook’s technology? For one thing, says Rothschild, the company has discovered that interrupts on the servers’ Ethernet controllers–which let the servers process myriad requests arriving at the same time–are a bottleneck, since they’re generally handled by only one core. So Facebook rewrote the controllers’ drivers to scale on multicore systems. Facebook is also experimenting with solid-state drives, which could speed the performance of the MySQL database tier by a factor of 100.
Given that Facebook is growing–and that connections grow exponentially–the site is going to need that performance soon.
The inside story of how ChatGPT was built from the people who made it
Exclusive conversations that take us behind the scenes of a cultural phenomenon.
ChatGPT is about to revolutionize the economy. We need to decide what that looks like.
New large language models will transform many jobs. Whether they will lead to widespread prosperity or not is up to us.
Sam Altman invested $180 million into a company trying to delay death
Can anti-aging breakthroughs add 10 healthy years to the human life span? The CEO of OpenAI is paying to find out.
GPT-4 is bigger and better than ChatGPT—but OpenAI won’t say why
We got a first look at the much-anticipated big new language model from OpenAI. But this time how it works is even more deeply under wraps.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.