An Oracle Exadata machine would cost Joel Gilbert at least $200,000, and that's just the hardware. Instead, Gilbert built his own version for one-fifth the cost.
A SaaS model suffers from network latency anyway. Team that with I/O latency and it's just a losing model.
CTO, Pipkins Inc.
True, it's not as powerful as an Exadata. Gilbert, chief technology officer at workforce management software company Pipkins Inc., acknowledges that. But according to him, he doesn't need something as powerful as Exadata. All he needed to do was find a cheaper way to solve the same problem that Exadata aims to solve: I/O latency.
Any IT shop doing a lot of database writes will run into the problem of I/O latency. The database needs to pass information to the storage layer and vice versa, and the pipe is only so thick. Oracle is aware of the problem, and Exadata attacks it at the storage layer. In particular, the storage cells that come with Exadata can do much of the processing previously left up to the CPUs. It's called Exadata Smart Scan.
Pipkins, based in Chesterfield, Mo., is a Software as a Service (SaaS)-based workforce management company founded almost 30 years ago. According to Gilbert, a big part of his job is dealing with and mitigating latency of all kinds.
"A SaaS model suffers from network latency anyway," he said. "Team that with I/O latency and it's just a losing model."
In 2009, Gilbert said, the situation reached a tipping point. The latency was brutal. So he brought all the storage area network (SAN) vendors in and told them they needed faster I/O. EMC and the rest of them handed him options that would stress the bottom line.
"They wanted to sell us racks," Gilbert said. "Big expensive things."
But he wasn't buying it. Gilbert is of the opinion that hard disks are going the way of the dinosaur, what he calls a "monolith of spinning spindles, waiting to die." Not only are they costly, but they also hog up a lot of room in the data center, and square footage there is at a premium. Gilbert was faced with a tough choice: Buy a bunch of SANs and then turn around and raise prices on Pipkins' customers -- which they would hate -- or find some kind of alternative.
Gilbert found the alternative.
It came not in the form of hordes of new hardware he had to rack and stack in his data center. Instead he bought a bunch of PCI-E flash memory cards from Fusion-io that he could plug into his existing commodity server hardware. The effects were immediately apparent. Engineers at the company thought the reports were wrong at first because the I/O latency "just disappeared." Gilbert said they can get more use out of each of their servers now as well. Before, CPU use was low because the system was waiting on I/O; with the cards installed, I/O isn't a problem and so the CPU can be saturated.
Don't get Gilbert wrong -- he still loves Oracle. He was a major proponent of moving Pipkins to Oracle Database from another proprietary database platform. But when it came to Exadata, Gilbert felt like Pipkins just didn't need it.
Gilbert said that Exadata does some of the same things Pipkins' architecture does. For example, he said, Exadata has architected the database to recognize when some of the data is on flash storage so that database algorithms can work more efficiently. But it also has some features that Gilbert finds unnecessary.
"We don't need all the things Exadata provides," he said, "like automatic updates. We're engineers. We know how to manually update. In the end, Oracle gets paid what they get paid [with Pipkins running Oracle Database]. We don't break the bank and we don't have to charge our customers prices they can't afford."