The mfr. date on the HDDs in NOV 2007 (you can see it in the high-res. CNET pic).

 Reply

And they have over 1000 servers in each shipping container.

So I think they'd be pretty set with 2 PB of storage.

Even if they only have .5 PB of storage per shipping container, they've got a lot of containers...

 Reply

I'm thinking it's real, but what I don't get is where the 5V is coming from for the HDDs. And if it's coming from the mobo, why is that more efficient than putting a 5V rail in the PSU? Either way you'd have to go from 12V -> 5V..

The image shows what appears to be a standard-looking 110-240v PSU PC power lead.

Are you talking about the input lead, or the leads to the mobo? In case of the latter, it's in no way standard-looking. All the wires are yellow or black (indicating 12V/gnd if they're following that tradition), and there are fewer wires than you'd see on a standard ATX (20- or 24-pin) connector.

Also, the battery leads go into the PSU, which is totally non-standard.

 Reply

I work in a HUGE data warehouse. I'm making a big guess here, because the systems I work on don't even represent 1% of the storage this place has, but I'm pretty positive we've got at least 5 to 10 petabytes in this facility. And I could be underestimating. I have no idea how much storage space Google's got sitting around in their data centers, but I'm sure it's much more than that.

What I'm trying to say is that in my relatively "small" data center, the types of servers we're running typically run at least 12 to 32 GB of ram, with usually more like 64 to 250GB of ram, with at least 4 quad core cpu's on the mother board. And many of them have more than one of those motherboards. And most of them don't carry their databases on internal storage, they are attached via high speed fiber to a SAN, where they have access to large banks of terabytes of storage space. I'm thinking google would be closer to those routes rather than some dinky little Gigabyte mobo with 8 slots for ram.

 Reply

Google is [in]famous for not doing things in the usual way. And the hardware pictured is a logical future iteration of what they started with:

[i.gizmodo.com]

If you read the CNET article, you'll see that this commodity (or slightly-modified commodity) hardware approach leads to a lower cost-per-query.

I tend to believe it; in your datacenter I would wager that your servers themselves are not wholly redundant to each other. IIRC, many arbitrary nodes in Google's architecture can fail without disrupting their entire system; hence they can get away with using cheaper, lower-end hardware than other DCs.

 Reply

* 6 x Registered DDRII 400 MHz Module

* Max. memory capacity up to 12 GB

* Supports ECC

@Marty200: @Allwheeldriveturbosportwagon is not an XUV/SUV/CUV or MPV: Damn my dyslexia!

 Reply

Back in college I had to work on machines that booted to a default date of Jan 1st, 1976, were 7 bit based, and programmed via toggle switches.

Mmmm, hard wired backplanes.

 Reply