July 9th, 2012
By Charles Babcock
Joyent's unique infrastructure design doesn't follow the path of Google or Amazon, includes built-in security and large data pipes for big business use.
While Amazon and Google steal the limelight, Joyent has been hard at work off in the wings building out a competitive cloud infrastructure to compete with both of them. So far it's raised $120 million to produce an infrastructure that departs significantly from both Amazon and Google.
At the same time, it competes on price and has the option of offering a 100% uptime guaranteed service level agreement. Amazon guarantees 99.9% or extends credits equal to the downtime experienced by the customer, an approach that doesn't recognize the cost to the business of being down. Google guarantees 99.95% uptime and also extends credits for downtime. Joyent returns 5% of the monthly fee for each 30 minutes of downtime encountered, even though 30 minutes is less than 1% of the month.
Joyent servers don't run Linux, as Google and Amazon servers do. It runs SmartOS, a derivative of Illumos, an open source operating system started in 2010 to make the former Sun OpenSolaris operating system more open than Oracle was inclined to allow.
Joyent's use of SmartOS enables the use of the Dtrace utility to track the performance of a running application. "We do tons of instrumentation with Dtrace across the system," said Jason Hoffman, founder and CTO of Joyent.
SmartOS invokes the use of Solaris-style zones in virtualization, where one operating system supports the operation of many virtual machines, instead of each VM needing to fire up its own. That means one host has one operating system, not a whole set of them, with more efficient multi-tenant operations. And it includes the ZFS file system, originated by Sun and allowing for the verified movement of files, something like the two-phase commit pattern of relational databases, said Hoffman. That provides for assured data movement in the cloud, supervised by the file system itself.
The typical x86 malware writer is not familiar with these features. If an intruder gets beyond the perimeter into the Joyent cloud, he lands in an operating system he doesn't recognize, using a file system with its own protections built in and running virtual machines in a way that doesn't look like anything he's seen before. Hoffman compares the protective layers to the double-hulled oil tanker that can suffer an outer hull breach but not lose any oil.
Instead of being a front-end CPU- and server-based design, Joyent's cloud infrastructure considers the backend more important. Its design is built to provide CPU and memory at the service of a storage system that resembles the Sun Unified Storage platform, a system that included a large, high-speed cache memory based on flash drives. It also uses a ZFS file system and lots of storage management smarts built in.
That design springs from Hoffman's experience as a cancer pathologist, building in an earlier life a supercomputer to help with cancer research. "We were building an I/O constrained supercomputer," he said, self-mockingly, of that effort. It was loaded with CPU and memory and could tackle complex algorithms, but it was restricted by the small pipes leading to and from the data. He knew from experience a successful cloud design would require heavy I/O throughput to storage.
The result, Hoffman said, is "purposeful engineering to deliver a highly scalable, flexible, and resilient cloud for business." Amazon has taken what worked for its retail operation and generalized it for public use. Google has taken what worked for its search engine and generalized it as public infrastructure. Joyent has designed cloud infrastructure from the ground up for business. The CPU to storage connection "is like the back end of a SAN with I/O on it," he said. High I/O rates are built-in.
Hoffman leaves the impression that a lot of Sun expertise has migrated into Joyent. Bryan Cantrill, author of Dtrace and designer of Sun's Unified Storage platform, works there, along with much of the Unified Storage platform team. Joyent also maintains Seattle offices, where it employs a group of former Amazon engineers.
Joyent has received a total $120 million in backing, with Intel one of its early backers, to build out its infrastructure as a service. It's a large figure, but Amazon, Google, or Microsoft probably spend $200 million to $400 million per large data center.
Joyent competes with other vendors at prices "similar to Amazon's or 20% lower," Hoffman claimed. That also translates into a 4-cent to 8-cent charge per hour for a one GB RAM virtual machine. Joyent is running an estimated 200,000 VMs for 20,000 to 30,000 customers on infrastructure in multiple data centers. It will sell its infrastructure design direct to customers through partners, such as Dell, for installation on premises to be managed as a private cloud. Or it will supply infrastructure through partners, such as Telefonica in Spain.
Hoffman said San Francisco-based Joyent is not focused on competing with big established cloud vendors so much as it seeks to convert more of its natural business clientele to its style of cloud computing. Joyent can run workloads running any hypervisor, whether VMware ESX Server, Microsoft Hyper-V, Citrix Systems XenServer, or Red Hat’s KVM. The SmartOS operating system knows to use the hardware extensions built into Intel (and AMD) chips that allow some calls for hardware operations to occur directly between the application and the hardware instead of being handled by the hypervisor, another aid to efficient operation.
"It wouldn't be possible for us to be in business without SmartOS. We need an always consistent file system. It's impossible for data to be lost or corrupted in our file system," he asserted.
With its unique cloud infrastructure and growing customer base, Joyent will become a large infrastructure vendor by continuing its unique approach, not copying the competition, Hoffman said.