The Secret Source of Google's Power

Much is being written about Gmail, Google's free webmail system. There's something deeper to learn about Google from this product than the initial reaction to the product features, however. Ignore for a moment the observations about Google leapfrogging their competitors with more user value and a new feature or two. Or Google diversifying away from search into other applications; they've been doing that for a while. Or the privacy red herring.

No, the story is about seemingly incremental features that are actually massively expensive for others to match, and the platform that Google is building which makes it cheaper and easier for them to develop and run web-scale applications than anyone else.

I've written before about Google's snippet service, which required that they store the entire web in RAM. All so they could generate a slightly better page excerpt than other search engines.

Google has taken the last 10 years of systems software research out of university labs, and built their own proprietary, production quality system. What is this platform that Google is building? It's a distributed computing platform that can manage web-scale datasets on 100,000 node server clusters. It includes a petabyte, distributed, fault tolerant filesystem, distributed RPC code, probably network shared memory and process migration. And a datacenter management system which lets a handful of ops engineers effectively run 100,000 servers. Any of these projects could be the sole focus of a startup.

Speculation: Gmail's Architecture and Economics

Let's make some guesses about how one might build a Gmail.

Hotmail has 60 million users. Gmail's design should be comparable, and should scale to 100 million users. It will only have to support a couple of million in the first year though.

  The most obvious challenge is the storage. You can't lose people's email, and you don't want to ever be down, so data has to be replicated. RAID is no good; when a disk fails, a human needs to replace the bad disk, or there is risk of data loss if more disks fail. One imagines the old ENIAC technician running up and down the isles of Google's data center with a shopping cart full of spare disk drives instead of vacuum tubes. RAID also requires more expensive hardware -- at least the hot swap drive trays. And RAID doesn't handle high availability at the server level anyway.

No. Google has 100,000 servers. If a server/disk dies, they leave it dead in the rack, to be reclaimed/replaced later. Hardware failures need to be instantly routed around by software.

Google has built their own distributed, fault-tolerant, petabyte filesystem, the Google Filesystem. This is ideal for the job. Say GFS replicates user email in three places; if a disk or a server dies, GFS can automatically make a new copy from one of the remaining two. Compress the email for a 3:1 storage win, then store user's email in three locations, and their raw storage need is approximately equivalent to the user's mail size.

The Gmail servers wouldn't be top-heavy with lots of disk. They need the CPU for indexing and page view serving anyway. No fancy RAID card or hot-swap trays, just 1-2 disks per 1U server.

It's straightforward to spreadsheet out the economics of the service, taking into account average storage per user, cost of the servers, and monetization per user per year. Google apparently puts the operational cost of storage at $2 per gigabyte. My napkin math comes up with numbers in the same ballpark. I would assume the yearly monetized value of a webmail user to be in the $1-10 range.

Cheap Hardware

Here's an anecdote to illustrate how far Google's cultural approach to hardware cost is different from the norm, and what it means as a component of their competitive advantage.

In a previous job I specified 40 moderately-priced servers to run a new internet search site we were developing. The ops team overrode me; they wanted 6 more expensive servers, since they said it would be easier to manage 6 machines than 40.

What this does is raise the cost of a CPU second. We had engineers that could imagine algorithms that would give marginally better search results, but if the algorithm was 10 times slower than the current code, ops would have to add 10X the number of machines to the datacenter. If you've already got $20 million invested in a modest collection of Suns, going 10X to run some fancier code is not an option.

Google has 100,000 servers.

Any sane ops person would rather go with a fancy $5000 server than a bare $500 motherboard plus disks sitting exposed on a tray. But that's a 10X difference to the cost of a CPU cycle. And this frees up the algorithm designers to invent better stuff.

Without cheap CPU cycles, the coders won't even consider algorithms that the Google guys are deploying. They're just too expensive to run.

Google doesn't deploy bare motherboards on exposed trays anymore; they're on at least the fourth iteration of their cheap hardware platform. Google now has an institutional competence building and maintaining servers that cost a lot less than the servers everyone else is using. And they do it with fewer people.

Think of the little internal factory they must have to deploy servers, and the level of automation needed to run that many boxes. Either network boot or a production line to pre-install disk images. Servers that self-configure on boot to determine their network config and load the latest rev of the software they'll be running. Normal datacenter ops practices don't scale to what Google has.

What are all those OS Researchers doing at Google?

Rob Pike

has gone to Google. Yes, that Rob Pike -- the OS researcher, the member of the original Unix team from Bell Labs. This guy isn't just some labs hood ornament; he writes code, lots of it. Big chunks of whole new operating systems like Plan 9.

Look at the depth of the research background of the Google employees in OS, networking, and distributed systems. Compiler Optimization. Thread migration. Distributed shared memory.

I'm a sucker for cool OS research. Browsing papers from Google employees about distributed systems, thread migration, network shared memory, GFS, makes me feel like a kid in Tomorrowland wondering when we're going to Mars. Wouldn't it be great, as an engineer, to have production versions of all this great research.

Google engineers do!

Competitive Advantage

Google is a company that has built a single very large, custom computer. It's running their own cluster operating system. They make their big computer even bigger and faster each month, while lowering the cost of CPU cycles. It's looking more like a general purpose platform than a cluster optimized for a single application.

While competitors are targeting the individual applications Google has deployed, Google is building a massive, general purpose computing platform for web-scale programming.

This computer is running the world's top search engine, a social networking service, a shopping price comparison engine, a new email service, and a local search/yellow pages engine. What will they do next with the world's biggest computer and most advanced operating system?