Distributed Caching with Memcached

Memcached is a high-performance, distributed caching system. Although application-neutral, it's most commonly used to speed up dynamic Web applications by alleviating database load. Memcached is used on LiveJournal, Slashdot, Wikipedia and other high-traffic sites.

For the past eight years I've been creating large, interactive, database-backed Web sites spanning multiple servers. Approximately 70 machines currently run LiveJournal.com, a blogging and social networking system with 2.5 million accounts. In addition to the typical blogging and friend/interest/profile declaration features, LiveJournal also sports forums, polls, a per-user news aggregator, audio posts by phone and other features useful for bringing people together.

Optimizing the speed of dynamic Web sites is always a challenge, and LiveJournal is no different. The task is made all the more challenging, because nearly any content item in the system can have an associated security level and be aggregated into many different views. From prior projects with dynamic, context-aware content, I knew from the beginning of LiveJournal's development that pregenerating static pages wasn't a viable optimization technique. It's impossible due to the constituent objects' cacheability and lifetimes being so different, so you make a bunch of sacrifices and waste a lot of time precomputing pages more often than they're requested.

This isn't to say caching is a bad thing. On the contrary, one of the core factors of a computer's performance is the speed, size and depth of its memory hierarchy. Caching definitely is necessary, but only if you do it on the right medium and at the right granularity. I find it best to cache each object on a page separately, rather than caching the entire page as a whole. That way you don't end up wasting space by redundantly caching objects and template elements that appear on more than one page.

In the end, though, it's all a series of trade-offs. Because processors keep getting faster, I find it preferable to burn CPU cycles rather than wait for disks. Modern disks keeping growing larger and cheaper, but they aren't getting much faster. Considering how slow and crash-prone they are, I try to avoid disks as much as possible. LiveJournal's Web nodes are all diskless, Netbooting off a common yet redundant NFS root image. Not only is this cheaper, but it requires significantly less maintenance.

Of course, disks are necessary for our database servers, but there we stick to fast disks with fast RAID setups. We actually have ten different database clusters, each with two or more machines. Nine of the clusters are user clusters, containing data specific to the users partitioned among them. One is our global cluster with non-user data and the table that maps users to their user clusters. The rationale for independent clusters is to spread writes. The alternative is having one big cluster with hundreds of slaves. The difficulty with such a monolithic cluster is it only spreads reads. The problem of diminishing returns appears as each new slave is added and increasingly is consumed by the writes necessary to stay up to date.

At this point you can see LiveJournal's back-end philosophy:

My definition of a little machine is more about re-usability than cost. I want a machine I can keep using as long as it's worth its space and heat output. I don't want to scale by throwing out machines every six months, replacing them with bigger machines.

Prior to Memcached, our Web nodes unconditionally hit our databases. This worked, but it wasn't as optimal as it could've been. I realized that even with 4G or 8G of memory, our database server caches were limited, both in raw memory size and by the address space available to our database server processes running on 32-bit machines. Yes, I could've replaced all our databases with 64-bit machines with much more memory, but recall that I'm stubborn and frugal.

I wanted to cache more on our Web nodes. Unfortunately, because we're using mod_perl 1.x, caching is a pain. Each process and thus, each Web request, is in its own address space and can't share data with the others. Each of the 30–50 processes could cache on its own, but doing so would be wasteful.

System V shared memory has too many weird limitations and isn't portable. It also works only on a single machine, not across 40+ Web nodes. These issues reflect what I saw as the main problem with most caching solutions. Even if our application platform was multithreaded with data easily shared between processes, we still could cache on only a single machine. I didn't want all 40+ machines caching independently and duplicating information.