In this expert video, storage consultant and Storage Decisions presenter Randy Kerns discusses the differences between scale-up and scale-out storage. Kerns, senior strategist at the Boulder, Colo.-based analyst firm Evaluator Group, offers diagrams showing classic scale-out architectures and explains how newer scale-out network-attached storage and SAN systems can provide much-needed capacity without slowing performance. View his expert video and read his remarks below for more on the scale-up vs. scale-out discussion.
For the most part, data storage managers historically have been scaling up and not out. In its simplest terms, scaling up means adding more disks while scaling out requires a more modular approach in which users add individual components as needed without having to add an entire storage node.
In the old days, "You bought a system and then you needed more space, so you plugged in more disk drives," Kerns said. "The problem was that it wasn't practical for most people in a financial sense; if you added capacity, you had a problem with synchronizing depreciation schedules."
But scale-out storage takes a different approach, he said. "Let's grow rather than just adding more capacity; let's grow by adding the control function and the capacity in some equal measure. Now I can scale performance and capacity at the same time, and I've got a separate asset, essentially, to start a separate depreciation schedule."
With scale-up, Kerns said, disks are added for capacity like "logs on a fire." As capacity is added, the number of IOPS shrinks. "Overall, I've reduced the performance because I put more capacity in there. That's the last thing you want to do. So you've done exactly the wrong thing for the customer.
"[With] scale-up, I just added devices," he explained, "[while with] scale-out, I've added more functions."
The key is how to link these functions, or federate them, so they're a single image to the outside world. This happens when distributed file systems allow for files to be spread across multiple nodes. "I want to be able to capacity-balance and load-balance between these nodes," Kerns noted. "That's the extra embedded software intelligence."
When it comes to scale-out block storage, there are two chief ways to go about implementing it, he said. One is a "monolithic storage system that you saw in the mainframe world. There's essentially a backplane with a bunch of cards plugged into it [EMC's VMAX architecture, for example]. The other is a federation of separate controller nodes that are connected, and they have to deal with cache coherency. An example of this is EqualLogic."